Object recommendation methods, user behavior prediction methods, devices and electronic devices

Abstract

This invention discloses an object recommendation method and apparatus. The method includes: obtaining target user features of a target user, candidate target object features of candidate target objects, and spatiotemporal environment features corresponding to the target user; inputting the target user features, candidate target object features, and spatiotemporal environment features into a user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target objects; and obtaining a judgment result on whether to recommend the candidate target objects as the target user's recommended objects based on the prediction data. Using this processing method, the obtained recommended objects for the target user have high precision and accuracy, and good adaptability to the target user.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to an object recommendation method. This invention also relates to a user behavior prediction method, a method for obtaining model output results, and a model training method; this application also relates to devices, electronic devices, and storage media related to the above-mentioned methods. Background Technology

[0002] With the development of internet technology, users in real-world scenarios live in an era of information overload. User-oriented recommendation services are commonplace in daily life. When users are faced with massive amounts of data, such as the complex product information on various e-commerce platforms, it becomes particularly important for recommendation services to obtain highly relevant and accurate recommendations.

[0003] In current object recommendation processes, static or adaptive model parameters are typically used to obtain recommended objects for user display. However, this approach often ignores the diversity and variability of user interests across different spatiotemporal environments. Consequently, its recommendation performance is poor, its match with user needs is low, and it fails to meet users' actual requirements. Therefore, improving the accuracy and precision of recommended objects and enhancing their suitability for target users becomes a pressing issue. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, one objective of this invention is to provide an object recommendation method that accurately obtains prediction data on the target user's execution of specified operations, thereby providing a foundation for more accurate object recommendations. Another objective of this invention is to provide a user behavior prediction method, a model output result acquisition method, and a model training method, making it possible to obtain more accurate recommended objects.

[0005] This application provides an object recommendation method, comprising: obtaining target user features of a target user, candidate target object features of candidate target objects, and spatiotemporal environment features corresponding to the target user; inputting the target user features, candidate target object features, and spatiotemporal environment features into a user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object; and obtaining a judgment result on whether to recommend the candidate target object as a recommended object for the target user based on the prediction data; wherein, in the user operation prediction model, at least one of the following processes is performed: assigning weights to the target user features and candidate target object features based on the spatiotemporal environment features, and obtaining the prediction data based on the target user feature weights and candidate target object feature weights; obtaining spatiotemporal semantic features based on the spatiotemporal environment features, target user features, and candidate target object features, and obtaining the prediction data based on the spatiotemporal semantic features; obtaining a fusion output feature with spatiotemporal deviation features based on the spatiotemporal environment features, target user features, and candidate target object features, and obtaining the prediction data based on the fusion output feature.

[0006] Optionally, assigning weights to the target user features based on the spatiotemporal environment features includes: performing concatenation processing on the spatiotemporal environment features and the target user features to obtain spatiotemporal user concatenation features; performing fully connected layer processing on the spatiotemporal user concatenation features to obtain spatiotemporal user fully connected processing results; and using the spatiotemporal user fully connected processing results as independent variables of the activation function to obtain the target user feature weights.

[0007] Optionally, the spatiotemporal environment features are spatiotemporal environment feature vectors, and the target user features are target user feature vectors; concatenating the spatiotemporal environment features and the target user features to obtain spatiotemporal user connection features includes: concatenating the spatiotemporal environment feature vector and the target user feature vector to obtain a spatiotemporal user connection feature vector; performing fully connected layer processing on the spatiotemporal user connection features to obtain a spatiotemporal user fully connected processing result includes: obtaining a spatiotemporal user fully connected processing result vector based on the spatiotemporal user connection feature vector, weight allocation of fully connected layer weight parameters, and weight allocation of fully connected layer bias parameters.

[0008] Optionally, assigning weights to the candidate target object features based on the spatiotemporal environment features includes: performing concatenation processing on the spatiotemporal environment features and the candidate target object features to obtain spatiotemporal object concatenation features; performing fully connected layer processing on the spatiotemporal object concatenation features to obtain a spatiotemporal object fully connected processing result; and using the spatiotemporal object fully connected processing result as the independent variable of the activation function to obtain the target candidate target object feature weights.

[0009] Optionally, the spatiotemporal environment features are spatiotemporal environment feature vectors, and the candidate target object features are candidate target object feature vectors; the step of concatenating the spatiotemporal environment features and the candidate target object features to obtain spatiotemporal object connection features includes: concatenating the spatiotemporal environment feature vectors and the candidate target object feature vectors to obtain spatiotemporal object connection feature vectors; the step of performing fully connected layer processing on the spatiotemporal object connection features to obtain spatiotemporal object fully connected processing results includes: obtaining a spatiotemporal object fully connected processing result vector based on the spatiotemporal object connection feature vectors, weight allocation of fully connected layer weight parameters, and weight allocation of fully connected layer bias parameters.

[0010] Optionally, obtaining the predicted data based on the target user feature weights and candidate target object feature weights includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; obtaining fused output features based on the spatiotemporally weighted target user features and the spatiotemporally weighted candidate target object features; and obtaining the predicted data based on the fused output features.

[0011] Optionally, the target user feature is a target user feature vector, and the candidate target object feature is a candidate target object feature vector; obtaining a spatiotemporally weighted target user feature with spatiotemporal weights based on the target user feature weights and the target user feature includes: performing a multiplication operation between the target user feature vector and the target user feature weights to obtain a spatiotemporally weighted target user feature vector; obtaining a spatiotemporally weighted candidate target object feature with spatiotemporal weights based on the candidate target object feature weights and the candidate target object feature includes: performing a multiplication operation between the candidate target object feature vector and the candidate target object feature weights to obtain a spatiotemporally weighted candidate target object feature vector.

[0012] Optionally, obtaining the fused output features based on the spatiotemporal weighted target user features and the spatiotemporal weighted candidate target object features includes: obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporal weighted target user features, and the spatiotemporal weighted candidate target object features; and obtaining the fused output features based on the spatiotemporal semantic features.

[0013] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporal weight target user features, and the spatiotemporal weight candidate target object features includes: obtaining original semantic features based on the spatiotemporal weight target user features and the spatiotemporal weight candidate target object features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0014] Optionally, obtaining the original semantic features based on the spatiotemporal weighted target user features and the spatiotemporal weighted candidate target object features includes: performing a concatenation process on the spatiotemporal weighted target user features and the spatiotemporal weighted candidate target object features to obtain the original semantic features.

[0015] Optionally, a target user's historical behavior sequence matching the spatiotemporal environment features is obtained; obtaining the spatiotemporal semantic fully connected layer weight parameters and the spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features includes: inputting the spatiotemporal environment features and the target user's historical behavior sequence matching the spatiotemporal environment features into the meta-network to obtain the spatiotemporal semantic fully connected layer weight parameters and the spatiotemporal semantic fully connected layer bias parameters.

[0016] Optionally, the step of inputting the spatiotemporal environmental features and the target user's historical behavior sequence matching the spatiotemporal environmental features into the meta-network to obtain the weight parameters and bias parameters of the spatiotemporal semantic fully connected layer includes: performing concatenation processing on the spatiotemporal environmental features and the target user's historical behavior sequence matching the spatiotemporal environmental features to obtain spatiotemporal historical behavior features; and obtaining the weight parameters and bias parameters of the spatiotemporal semantic fully connected layer based on the spatiotemporal historical behavior features, the meta-network weight parameters, and the meta-network bias parameters.

[0017] Optionally, the original historical behavior sequence of the target user is obtained; obtaining the target user's historical behavior sequence that matches the spatiotemporal environmental features includes: filtering out the target user's historical behavior sequence that matches the spatiotemporal environmental features from the target user's original historical behavior sequence.

[0018] Optionally, obtaining the fused output features based on the spatiotemporal semantic features includes: performing processing on the spatiotemporal semantic features using multiple fused fully connected batch normalization layers to obtain the fused output features.

[0019] Optionally, the step of performing multiple fusion fully connected batch normalization layers on the spatiotemporal semantic features to obtain the fusion output features includes: inputting the spatiotemporal semantic features into a first fusion fully connected batch normalization layer to obtain a first fusion intermediate feature; inputting the first fusion intermediate feature into a second fusion fully connected batch normalization layer to obtain a second fusion intermediate feature; inputting the second fusion intermediate feature into a third fusion fully connected batch normalization layer, and so on, until the last fusion intermediate feature output by the penultimate fusion fully connected batch normalization layer is obtained; and inputting the last fusion intermediate feature into the last fusion fully connected batch normalization layer to obtain the fusion output features.

[0020] Optionally, the step of performing processing on multiple fused fully connected batch normalization layers for the spatiotemporal semantic features to obtain the fused output features includes: for any fused fully connected batch normalization layer, obtaining the output result of the arbitrary fused fully connected batch normalization layer based on the output result of the previous fused fully connected batch normalization layer, the spatiotemporal adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer, and the batch normalization adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer.

[0021] Optionally, it also includes: obtaining the spatiotemporal adjustment parameters corresponding to any one of the fused fully connected batch normalization layers and the batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics.

[0022] Optionally, obtaining the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics includes: performing fully connected layer processing on the spatiotemporal environment characteristics for any one of the fused fully connected batch normalization layers to obtain the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers; and using the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers as the independent variable of the activation function to obtain the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers.

[0023] Optionally, obtaining the predicted data based on the fused output features includes: performing fully connected layer processing on the fused output features to obtain a fused output fully connected processing result; and using the fused output fully connected processing result as the independent variable of the activation function to obtain the predicted data.

[0024] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: obtaining original semantic features based on the target user features and the candidate target object features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0025] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; and obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted candidate target object features.

[0026] Optionally, obtaining the predicted data based on the spatiotemporal semantic features includes: obtaining fused output features based on the spatiotemporal semantic features; and obtaining the predicted data based on the fused output features.

[0027] Optionally, obtaining the fusion output feature with spatiotemporal deviation features based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted candidate target object features; and obtaining the fusion output feature based on the spatiotemporal semantic features.

[0028] Optionally, it further includes: detecting the actual result data of the target user performing a specified operation on the candidate target object; if the deviation data between the predicted data and the actual result data does not exceed a predetermined deviation threshold, then the target user features, the candidate target object features, the spatiotemporal environment features, and the predicted data are used as positive samples to update the user operation prediction model; otherwise, the target user features, the candidate target object features, the spatiotemporal environment features, and the predicted data are used as negative samples to update the user operation prediction model.

[0029] Optionally, it further includes: obtaining combined features and / or the original historical behavior sequence of the target user, wherein the combined features are features obtained by combining at least two features from the target user, at least one feature of the candidate target object, and the spatiotemporal environment features; the step of inputting the target user features, the candidate target object features, and the spatiotemporal environment features into the user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object includes: inputting the combined features and / or the original historical behavior sequence of the target user, the target user features, the candidate target object features, and the spatiotemporal environment features into the user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object.

[0030] Optionally, if the original historical behavior sequence of the target user is obtained, then the step of inputting the combined features and / or the original historical behavior sequence of the target user, the target user features, the candidate target object features, and the spatiotemporal environment features into the user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object includes: obtaining target user interest features of the target user on the candidate target object features based on the original historical behavior sequence of the target user and the candidate target object features; and obtaining prediction data of the target user performing a specified operation on the candidate target object based on the combined features, the original historical behavior sequence of the target user, the target user features, the candidate target object features, the target user interest features, and the spatiotemporal environment features.

[0031] Optionally, obtaining the target user interest features of the target user for the candidate target object features based on the target user's original historical behavior sequence and the candidate target object features includes: processing the target user's original historical behavior sequence and the candidate target object features using a multi-head target attention processing layer to obtain the target user interest features.

[0032] Optionally, obtaining the target user interest features of the target user for the candidate target object features based on the target user's original historical behavior sequence and the candidate target object features includes: determining the historical object in which the target user performed a specified historical operation based on the target user's original historical behavior sequence; and obtaining the target user interest features based on the similarity between the candidate target object and the historical object.

[0033] Optionally, the candidate target objects include at least one of candidate products, candidate merchants, candidate benefits, and candidate promotional activities; the user operation prediction model is used to predict at least one of the following data for the target user regarding the candidate target objects: click-through rate, add-to-cart rate, purchase rate, and repurchase rate.

[0034] Optionally, the method is applied to the server side, and the method further includes: obtaining a page content request message sent by the client to request page content, and returning page content to the client in response to the page content request message, wherein if the judgment result indicates that the candidate target object is used as the recommended object for the target user, then the page content includes the candidate target object; or, obtaining an object query message sent by the client, obtaining candidate target objects in response to the object query message, and returning object query results to the client, wherein if the judgment result indicates that the candidate target object is used as the recommended object for the target user, then the object query results include the candidate target object.

[0035] Optionally, the method is applied to a client, and the method further includes: displaying page content in response to detecting a trigger operation for displaying page content, wherein if the determination result indicates that the candidate target object is a recommended object for the target user, then the page content includes the candidate target object; or, displaying object query results in response to detecting a trigger operation for querying an object, wherein if the determination result indicates that the candidate target object is a recommended object for the target user, then the object query results include the candidate target object.

[0036] This application provides a user behavior prediction method, comprising: obtaining target user features of a target user, target operation features of a specified operation, and spatiotemporal environment features corresponding to the target user; inputting the target user features, the target operation features, and the spatiotemporal environment features into a user operation prediction model to obtain prediction data of the target user performing the specified operation; wherein, in the user operation prediction model, at least one of the following processes is performed: assigning weights to the target user features and the target operation features according to the spatiotemporal environment features, and obtaining the prediction data according to the target user feature weights and the target operation feature weights; obtaining spatiotemporal semantic features according to the spatiotemporal environment features, the target user features, and the target operation features, and obtaining the prediction data according to the spatiotemporal semantic features; obtaining a fused output feature with spatiotemporal deviation features according to the spatiotemporal environment features, the target user features, and the target operation features, and obtaining the prediction data according to the fused output feature.

[0037] Optionally, assigning weights to the target operation features based on the spatiotemporal environment features includes: performing concatenation processing on the spatiotemporal environment features and the target operation features to obtain spatiotemporal target operation concatenation features; performing fully connected layer processing on the spatiotemporal target operation to obtain a fully connected processing result for the spatiotemporal target operation; and using the fully connected processing result for the spatiotemporal target operation as the independent variable of the activation function to obtain the target operation feature weights.

[0038] Optionally, the spatiotemporal environment features are spatiotemporal environment feature vectors, and the target operation features are target operation feature vectors; the step of concatenating the spatiotemporal environment features and the target operation features to obtain spatiotemporal target operation connection features includes: concatenating the spatiotemporal environment feature vector and the target operation feature vector to obtain a spatiotemporal target operation connection feature vector; the step of performing fully connected layer processing on the spatiotemporal target operation to obtain a spatiotemporal target operation fully connected processing result includes: obtaining a spatiotemporal target operation fully connected processing result vector based on the spatiotemporal target operation connection feature vector, weight allocation of fully connected layer weight parameters, and weight allocation of fully connected layer bias parameters.

[0039] Optionally, obtaining the predicted data based on the target user feature weights and target operation feature weights includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted target operation features with spatiotemporal weights based on the candidate target object feature weights and the target operation features; obtaining fused output features based on the spatiotemporally weighted target user features and the spatiotemporally weighted target operation features; and obtaining the predicted data based on the fused output features.

[0040] Optionally, the target user feature is a target user feature vector, and the target operation feature is a target operation feature vector; obtaining a spatiotemporally weighted target user feature with spatiotemporal weights based on the target user feature weights and the target user feature includes: performing a multiplication operation on the target user feature vector and the target user feature weights to obtain a spatiotemporally weighted target user feature vector; obtaining a spatiotemporally weighted target operation feature with spatiotemporal weights based on the target operation feature weights and the target operation feature includes: performing a multiplication operation on the target operation feature vector and the target operation feature weights to obtain a spatiotemporally weighted target operation feature vector.

[0041] Optionally, obtaining the fused output features based on the spatiotemporal weighted target user features and the spatiotemporal weighted target operation features includes: obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporal weighted target user features, and the spatiotemporal weighted target operation features; and obtaining the fused output features based on the spatiotemporal semantic features.

[0042] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporal weight target user features, and the spatiotemporal weight target operation features includes: obtaining original semantic features based on the spatiotemporal weight target user features and the spatiotemporal weight target operation features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0043] Optionally, obtaining the original semantic features based on the spatiotemporal weight target user features and the spatiotemporal weight target operation features includes: performing concatenation processing on the spatiotemporal weight target user features and the spatiotemporal weight target operation features to obtain the original semantic features.

[0044] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the target operation features includes: obtaining original semantic features based on the target user features and the target operation features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0045] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the target operation features includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted target operation features with spatiotemporal weights based on the target operation feature weights and the target operation features; and obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted target operation features.

[0046] Optionally, obtaining the fusion output feature with spatiotemporal deviation features based on the spatiotemporal environment features, the target user features, and the target operation features includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted target operation features with spatiotemporal weights based on the target operation feature weights and the target operation features; obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted target operation features; and obtaining the fusion output feature based on the spatiotemporal semantic features.

[0047] Optionally, it further includes: detecting the actual result data of the target user performing the specified operation; if the deviation data between the predicted data and the actual result data does not exceed a predetermined deviation threshold, then the target user features, the target operation features, the spatiotemporal environment features, and the predicted data are used as positive samples to update the user operation prediction model; otherwise, the target user features, the target operation features, the spatiotemporal environment features, and the predicted data are used as negative samples to update the user operation prediction model.

[0048] Optionally, it further includes: obtaining combined features and / or the original historical behavior sequence of the target user, wherein the combined features are features obtained by combining at least two features from at least one feature of the target user, at least one target operation feature, and the spatiotemporal environment feature; the step of inputting the target user feature, the target operation feature, and the spatiotemporal environment feature into a user operation prediction model to obtain prediction data of the target user performing a specified operation includes: inputting the combined features and / or the original historical behavior sequence of the target user, the target user feature, the target operation, and the spatiotemporal environment feature into a user operation prediction model to obtain prediction data of the target user performing a specified operation.

[0049] Optionally, if the original historical behavior sequence of the target user is obtained, then the step of inputting the combined features and / or the original historical behavior sequence of the target user, the target user features, the target operation features, and the spatiotemporal environment features into the user operation prediction model to obtain prediction data of the target user performing a specified operation includes: obtaining the target user interest features of the target user for the target operation features based on the original historical behavior sequence of the target user and the target operation features; and obtaining the prediction data of the target user performing a specified operation based on the combined features, the original historical behavior sequence of the target user, the target user features, the target operation features, the target user interest features, and the spatiotemporal environment features.

[0050] Optionally, obtaining the target user interest features of the target user in relation to the target operation features based on the target user's original historical behavior sequence and the target operation features includes: processing the target user's original historical behavior sequence and the target operation features using a multi-head target attention processing layer to obtain the target user interest features.

[0051] Optionally, obtaining the target user interest features of the target user for the target operation features based on the target user's original historical behavior sequence and the target operation features includes: determining the historical operations performed by the target user based on the target user's original historical behavior sequence; and obtaining the target user interest features based on the similarity between the target operation features and the historical operation features of the historical operations performed by the target user.

[0052] This application provides a method for obtaining model output results, comprising: obtaining spatiotemporal environmental features corresponding to a target user and multiple basic input features of the user operation prediction model; assigning weights to the multiple basic input features according to the spatiotemporal environmental features to obtain spatiotemporal weights corresponding to the multiple basic input features respectively; obtaining multiple spatiotemporal weighted basic input features according to the multiple basic input features and the spatiotemporal weights corresponding to the multiple basic input features respectively; and obtaining the output result of the user operation prediction model according to the multiple spatiotemporal weighted basic input features.

[0053] Optionally, the step of assigning weights to the plurality of basic input features based on the spatiotemporal environment features to obtain the spatiotemporal weights corresponding to the plurality of basic input features includes: concatenating the spatiotemporal environment features with any one of the basic input features to obtain spatiotemporal basic input concatenation features; performing fully connected layer processing on the spatiotemporal basic input concatenation features to obtain a spatiotemporal basic input fully connected processing result; and using the spatiotemporal basic input fully connected processing result as the independent variable of the activation function to obtain the basic input feature weights of any one of the basic input features.

[0054] Optionally, the spatiotemporal environment feature is a spatiotemporal environment feature vector, and the basic input feature is a basic input feature vector. Connecting the spatiotemporal environment feature with any basic input feature to obtain a spatiotemporal basic input connection feature includes: connecting the spatiotemporal environment feature vector with the basic input feature vector of any basic input feature to obtain a spatiotemporal basic input connection feature vector. Performing fully connected layer processing on the spatiotemporal basic input connection feature to obtain a spatiotemporal basic input fully connected processing result includes: obtaining a spatiotemporal basic input fully connected processing result vector based on the spatiotemporal basic input connection feature vector, weight allocation of the fully connected layer weight parameters, and weight allocation of the fully connected layer bias parameters.

[0055] Optionally, the basic input features are basic input feature vectors; obtaining multiple spatiotemporal weighted basic input features based on the multiple basic input features and the spatiotemporal weights corresponding to the multiple basic input features includes: performing a multiplication operation between the basic input feature vector of any basic input feature and the spatiotemporal weight corresponding to the any basic input feature to obtain the spatiotemporal weighted basic input feature vector of the any basic input feature.

[0056] Optionally, obtaining the output result of the user operation prediction model based on the plurality of spatiotemporal weighted basic input features includes: obtaining fused output features based on the plurality of spatiotemporal weighted basic input features; and obtaining the output result of the user operation prediction model based on the fused output features.

[0057] Optionally, obtaining the fused output features based on the plurality of spatiotemporal weighted basic input features includes: obtaining spatiotemporal semantic features based on the spatiotemporal environment features and the plurality of spatiotemporal weighted basic input features; and obtaining the fused output features based on the spatiotemporal semantic features.

[0058] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features and the plurality of spatiotemporal weighted basic input features includes: obtaining original semantic features based on the plurality of spatiotemporal weighted basic input features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0059] Optionally, obtaining the fused output features based on the spatiotemporal semantic features includes: performing processing on the spatiotemporal semantic features using multiple fused fully connected batch normalization layers to obtain the fused output features.

[0060] Optionally, obtaining the output result of the user operation prediction model based on the fused output features includes: performing fully connected layer processing on the fused output features to obtain a fused output fully connected processing result; and using the fused output fully connected processing result as the independent variable of the activation function to obtain the prediction data.

[0061] This application provides a method for obtaining model output results, comprising: obtaining spatiotemporal environmental features corresponding to a target user and multiple basic input features of the user operation prediction model; obtaining spatiotemporal semantic features based on the spatiotemporal environmental features and the multiple basic input features; and obtaining the output results of the user operation prediction model based on the spatiotemporal semantic features.

[0062] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features and the plurality of basic input features includes: obtaining original semantic features based on the spatiotemporal weighted basic input features corresponding to the plurality of basic input features respectively; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0063] Optionally, obtaining the original semantic features based on the spatiotemporal weighted basic input features corresponding to the plurality of basic input features includes: concatenating the spatiotemporal weighted basic input features corresponding to the plurality of basic input features to obtain the original semantic features.

[0064] Optionally, it further includes: obtaining the target user's historical behavior sequence matching the spatiotemporal environment features; obtaining the spatiotemporal semantic fully connected layer weight parameters and the spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features includes: inputting the spatiotemporal environment features and the target user's historical behavior sequence matching the spatiotemporal environment features into the meta-network to obtain the spatiotemporal semantic fully connected layer weight parameters and the spatiotemporal semantic fully connected layer bias parameters.

[0065] Optionally, the step of inputting the spatiotemporal environmental features and the target user's historical behavior sequence matching the spatiotemporal environmental features into the meta-network to obtain the weight parameters and bias parameters of the spatiotemporal semantic fully connected layer includes: performing concatenation processing on the spatiotemporal environmental features and the target user's historical behavior sequence matching the spatiotemporal environmental features to obtain spatiotemporal historical behavior features; and obtaining the weight parameters and bias parameters of the spatiotemporal semantic fully connected layer based on the spatiotemporal historical behavior features, the meta-network weight parameters, and the meta-network bias parameters.

[0066] Optionally, it further includes: obtaining the original historical behavior sequence of the target user; obtaining the target user's historical behavior sequence that matches the spatiotemporal environmental features includes: filtering out the target user's historical behavior sequence that matches the spatiotemporal environmental features from the target user's original historical behavior sequence.

[0067] Optionally, obtaining the output result of the user operation prediction model based on the spatiotemporal semantic features includes: obtaining fused output features based on the spatiotemporal semantic features; and obtaining the output result of the user operation prediction model based on the fused output features.

[0068] This application provides a method for obtaining model output results, including: obtaining spatiotemporal environmental features corresponding to a target user and multiple basic input features of the user operation prediction model; obtaining fused output features with spatiotemporal deviation features based on the spatiotemporal environmental features and the multiple basic input features; and obtaining the output result of the user operation prediction model based on the fused output features.

[0069] Optionally, obtaining the fused output feature with spatiotemporal deviation features based on the spatiotemporal environmental features and the plurality of basic input features includes: obtaining a spatiotemporal weighted basic input feature corresponding to any one basic input feature and the spatiotemporal weight corresponding to the one basic input feature; obtaining spatiotemporal semantic features based on the spatiotemporal environmental features and the spatiotemporal weighted basic input features corresponding to the plurality of basic input features; and obtaining the fused output feature based on the spatiotemporal semantic features.

[0070] Optionally, obtaining the fused output features based on the spatiotemporal semantic features includes: performing processing on the spatiotemporal semantic features using multiple fused fully connected batch normalization layers to obtain the fused output features.

[0071] Optionally, the step of performing multiple fusion fully connected batch normalization layers on the spatiotemporal semantic features to obtain the fusion output features includes: inputting the spatiotemporal semantic features into a first fusion fully connected batch normalization layer to obtain a first fusion intermediate feature; inputting the first fusion intermediate feature into a second fusion fully connected batch normalization layer to obtain a second fusion intermediate feature; inputting the second fusion intermediate feature into a third fusion fully connected batch normalization layer, and so on, until the last fusion intermediate feature output by the penultimate fusion fully connected batch normalization layer is obtained; and inputting the last fusion intermediate feature into the last fusion fully connected batch normalization layer to obtain the fusion output features.

[0072] Optionally, the step of performing processing on multiple fused fully connected batch normalization layers for the spatiotemporal semantic features to obtain the fused output features includes: for any fused fully connected batch normalization layer, obtaining the output result of the arbitrary fused fully connected batch normalization layer based on the output result of the previous fused fully connected batch normalization layer, the spatiotemporal adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer, and the batch normalization adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer.

[0073] Optionally, it also includes: obtaining the spatiotemporal adjustment parameters corresponding to any one of the fused fully connected batch normalization layers and the batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics.

[0074] Optionally, obtaining the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics includes: performing fully connected layer processing on the spatiotemporal environment characteristics for any one of the fused fully connected batch normalization layers to obtain the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers; and using the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers as the independent variable of the activation function to obtain the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers.

[0075] Optionally, obtaining the output result of the user operation prediction model based on the fused output features includes: performing fully connected layer processing on the fused output features to obtain a fused output fully connected processing result; and using the fused output fully connected processing result as the independent variable of the activation function to obtain the output result of the user operation prediction model.

[0076] This application provides a model training method for a user operation prediction model, comprising: acquiring prediction data of a target user performing a specified operation on a candidate target object and actual result data of the target user performing the specified operation on the candidate target object; calculating the deviation data between the prediction data and the actual result data; if the deviation data between the prediction data and the actual result data does not exceed a predetermined deviation threshold, then using the current user operation prediction model as the user operation prediction model recommended for the object; if the deviation data between the prediction data and the actual result data is greater than the predetermined deviation threshold, then training the current user operation prediction model until the deviation data meets the predetermined condition of not exceeding the predetermined deviation threshold.

[0077] This application provides an object recommendation method, comprising: obtaining recommended objects to a target user in response to a trigger condition for displaying information; and displaying the recommended objects on a display interface for displaying information; wherein the recommended objects are determined based on prediction data of the target user performing a specified operation on a candidate target object; the prediction data is obtained by assigning target user feature weights to the target user features of the target user based on the spatiotemporal environmental features corresponding to the target user and assigning candidate target object feature weights to the candidate target object features; or, the prediction data is obtained based on spatiotemporal semantic features obtained based on the spatiotemporal environmental features, the target user features, and the candidate target object features; or, the prediction data is obtained based on fused output features with spatiotemporal deviation features obtained based on the spatiotemporal environmental features, the target user features, and the candidate target object features.

[0078] Optionally, the response to meeting the triggering condition for displaying information includes: responding to detecting a triggering operation for displaying page content; or, responding to detecting a triggering operation for querying an object.

[0079] Optionally, target user feature weights are assigned to the target user features of the target user based on the spatiotemporal environmental features corresponding to the target user in the following manner: the spatiotemporal environmental features are concatenated with the target user features to obtain spatiotemporal user concatenation features; the spatiotemporal user concatenation features are processed by a fully connected layer to obtain a spatiotemporal user fully connected processing result; the spatiotemporal user fully connected processing result is used as the independent variable of the activation function to obtain the target user feature weights.

[0080] Optionally, candidate target object feature weights are assigned to the candidate target object features of the candidate target object based on the spatiotemporal environment features corresponding to the target user in the following manner: the spatiotemporal environment features are concatenated with the candidate target object features to obtain spatiotemporal object concatenation features; the spatiotemporal object concatenation features are processed by a fully connected layer to obtain a spatiotemporal object fully connected processing result; the spatiotemporal object fully connected processing result is used as the independent variable of the activation function to obtain the candidate target object feature weights.

[0081] Optionally, the predicted data is obtained according to the target user feature weights and the candidate target object feature weights in the following manner: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; obtaining fused output features based on the spatiotemporally weighted target user features and the spatiotemporally weighted candidate target object features; and obtaining the predicted data based on the fused output features.

[0082] Optionally, the target user feature is a target user feature vector, and the candidate target object feature is a candidate target object feature vector; obtaining a spatiotemporally weighted target user feature with spatiotemporal weights based on the target user feature weights and the target user feature includes: performing a multiplication operation between the target user feature vector and the target user feature weights to obtain a spatiotemporally weighted target user feature vector; obtaining a spatiotemporally weighted candidate target object feature with spatiotemporal weights based on the candidate target object feature weights and the candidate target object feature includes: performing a multiplication operation between the candidate target object feature vector and the candidate target object feature weights to obtain a spatiotemporally weighted candidate target object feature vector.

[0083] Optionally, obtaining the fused output features based on the spatiotemporal weighted target user features and the spatiotemporal weighted candidate target object features includes: obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporal weighted target user features, and the spatiotemporal weighted candidate target object features; and obtaining the fused output features based on the spatiotemporal semantic features.

[0084] Optionally, spatiotemporal semantic features are obtained based on the spatiotemporal environment features, the target user features, and the candidate target object features in the following manner: obtaining original semantic features based on the target user features and the candidate target object features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0085] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; and obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted candidate target object features.

[0086] Optionally, the predicted data can be obtained based on the spatiotemporal semantic features in the following manner: obtain fused output features based on the spatiotemporal semantic features; obtain the predicted data based on the fused output features.

[0087] Optionally, the fusion output feature with spatiotemporal deviation features is obtained according to the spatiotemporal environment features, the target user features, and the candidate target object features in the following manner: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted candidate target object features; and obtaining the fusion output feature based on the spatiotemporal semantic features.

[0088] Optionally, the fused output features are obtained based on the spatiotemporal semantic features in the following manner: for the spatiotemporal semantic features, multiple fused fully connected batch normalization layers are processed to obtain the fused output features.

[0089] Optionally, the step of performing multiple fusion fully connected batch normalization layers on the spatiotemporal semantic features to obtain the fusion output features includes: inputting the spatiotemporal semantic features into a first fusion fully connected batch normalization layer to obtain a first fusion intermediate feature; inputting the first fusion intermediate feature into a second fusion fully connected batch normalization layer to obtain a second fusion intermediate feature; inputting the second fusion intermediate feature into a third fusion fully connected batch normalization layer, and so on, until the last fusion intermediate feature output by the penultimate fusion fully connected batch normalization layer is obtained; and inputting the last fusion intermediate feature into the last fusion fully connected batch normalization layer to obtain the fusion output features.

[0090] Optionally, the step of performing processing on multiple fused fully connected batch normalization layers for the spatiotemporal semantic features to obtain the fused output features includes: for any fused fully connected batch normalization layer, obtaining the output result of the arbitrary fused fully connected batch normalization layer based on the output result of the previous fused fully connected batch normalization layer, the spatiotemporal adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer, and the batch normalization adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer.

[0091] Optionally, it also includes: obtaining the spatiotemporal adjustment parameters corresponding to any one of the fused fully connected batch normalization layers and the batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics.

[0092] Optionally, obtaining the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics includes: performing fully connected layer processing on the spatiotemporal environment characteristics for any one of the fused fully connected batch normalization layers to obtain the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers; and using the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers as the independent variable of the activation function to obtain the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers.

[0093] Optionally, the predicted data can be obtained based on the fused output features in the following manner: the fused output features are processed by a fully connected layer to obtain a fused output fully connected processing result; the fused output fully connected processing result is used as the independent variable of the activation function to obtain the predicted data.

[0094] This application provides an apparatus for object recommendation, comprising: an acquisition unit for acquiring target user features of a target user, candidate target object features of candidate target objects, and spatiotemporal environment features corresponding to the target user; a first processing unit for inputting the target user features, candidate target object features, and spatiotemporal environment features into a user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object; and a second processing unit for obtaining a judgment result on whether to recommend the candidate target object as a recommended object for the target user based on the prediction data; wherein, in the user operation prediction model, at least one of the following processes is performed: assigning weights to the target user features and candidate target object features based on the spatiotemporal environment features, and obtaining the prediction data based on the target user feature weights and candidate target object feature weights; obtaining spatiotemporal semantic features based on the spatiotemporal environment features, target user features, and candidate target object features, and obtaining the prediction data based on the spatiotemporal semantic features; obtaining a fusion output feature with spatiotemporal deviation features based on the spatiotemporal environment features, target user features, and candidate target object features, and obtaining the prediction data based on the fusion output feature.

[0095] This application provides a user behavior prediction apparatus, comprising: an acquisition unit for acquiring target user features of a target user, target operation features of a specified operation, and spatiotemporal environment features corresponding to the target user; and a processing unit for inputting the target user features, the target operation features, and the spatiotemporal environment features into a user operation prediction model to obtain prediction data of the target user performing the specified operation; wherein, in the user operation prediction model, at least one of the following processes is performed: assigning weights to the target user features and the target operation features according to the spatiotemporal environment features, and obtaining the prediction data according to the target user feature weights and the target operation feature weights; obtaining spatiotemporal semantic features according to the spatiotemporal environment features, the target user features, and the target operation features, and obtaining the prediction data according to the spatiotemporal semantic features; obtaining a fused output feature with spatiotemporal deviation features according to the spatiotemporal environment features, the target user features, and the target operation features, and obtaining the prediction data according to the fused output feature.

[0096] This application provides a model output result acquisition device, comprising: an acquisition unit for acquiring spatiotemporal environmental features corresponding to a target user and multiple basic input features of the user operation prediction model; a first processing unit for assigning weights to the multiple basic input features according to the spatiotemporal environmental features, thereby acquiring spatiotemporal weights corresponding to the multiple basic input features respectively; a second processing unit for acquiring multiple spatiotemporal weighted basic input features according to the multiple basic input features and the spatiotemporal weights corresponding to the multiple basic input features respectively; and a result acquisition unit for acquiring the output result of the user operation prediction model according to the multiple spatiotemporal weighted basic input features.

[0097] This application provides a model output result acquisition device, comprising: an acquisition unit for acquiring spatiotemporal environmental features corresponding to a target user and multiple basic input features of the user operation prediction model; a first processing unit for acquiring spatiotemporal semantic features based on the spatiotemporal environmental features and the multiple basic input features; and a result acquisition unit for acquiring the output result of the user operation prediction model based on the spatiotemporal semantic features.

[0098] This application provides a model output result acquisition device, comprising: an acquisition unit for acquiring spatiotemporal environmental features corresponding to a target user and multiple basic input features of the user operation prediction model; a first processing unit for acquiring fused output features with spatiotemporal deviation features based on the spatiotemporal environmental features and the multiple basic input features; and a result acquisition unit for acquiring the output result of the user operation prediction model based on the fused output features.

[0099] This application provides a model training apparatus for a user operation prediction model, comprising: an acquisition unit for acquiring prediction data of a target user performing a specified operation on a candidate target object and actual result data of the target user performing the specified operation on the candidate target object; and a processing unit for calculating deviation data between the prediction data and the actual result data; if the deviation data between the prediction data and the actual result data does not exceed a predetermined deviation threshold, then the current user operation prediction model is used as the recommended user operation prediction model; if the deviation data between the prediction data and the actual result data is greater than the predetermined deviation threshold, then the current user operation prediction model is trained until the deviation data meets the predetermined condition of not exceeding the predetermined deviation threshold.

[0100] This application provides an object recommendation device, characterized in that it includes: an acquisition unit configured to acquire recommended objects to a target user in response to a trigger condition for displaying information; and a display unit configured to display the recommended objects on a display interface for displaying information; wherein the recommended objects are determined based on prediction data of the target user performing a specified operation on a candidate target object; the prediction data is obtained by assigning target user feature weights to the target user features of the target user based on the spatiotemporal environmental features corresponding to the target user and assigning candidate target object feature weights to the candidate target object features; or, the prediction data is obtained based on spatiotemporal semantic features obtained based on the spatiotemporal environmental features, the target user features, and the candidate target object features; or, the prediction data is obtained based on fused output features with spatiotemporal deviation features obtained based on the spatiotemporal environmental features, the target user features, and the candidate target object features.

[0101] This application provides an electronic device including a processor and a memory; wherein the memory is used to store one or more computer instructions, which are executed by the processor to implement any method embodiment. This application also provides a computer-readable storage medium storing one or more computer instructions that are executed by a processor to implement any method embodiment. Attached Figure Description

[0102] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0103] Figure 1 This is a schematic diagram illustrating an application scenario of an embodiment of this application;

[0104] Figure 2 This is a flowchart illustrating the first embodiment of this application;

[0105] Figure 3 This is a model architecture diagram corresponding to the user operation prediction model in the first embodiment of this application;

[0106] Figure 4 This is a schematic diagram corresponding to the first embodiment of this application, which assigns weights to target user features based on spatiotemporal environmental characteristics;

[0107] Figure 5 This is a schematic diagram illustrating the transformation of original semantic features into spatiotemporal semantic features corresponding to the first embodiment of this application;

[0108] Figure 6 This is a schematic diagram illustrating the acquisition of fused output features based on spatiotemporal semantic features, corresponding to the first embodiment of this application.

[0109] Figure 7 This is a flowchart of a user behavior prediction method according to Embodiment 2 of this application;

[0110] Figure 8 This is a flowchart of a method for obtaining model output results according to Embodiment 3 of this application;

[0111] Figure 9 This is a flowchart of a method for obtaining model output results corresponding to the fourth embodiment of this application;

[0112] Figure 10 This is a flowchart of a method for obtaining model output results according to the fifth embodiment of this application;

[0113] Figure 11 This is a flowchart of a model training method according to the sixth embodiment of this application;

[0114] Figure 11A This is a flowchart of an object recommendation method according to the seventh embodiment of this application;

[0115] Figure 12 This is a schematic diagram of the structure of an object recommendation device according to the eighth embodiment of this application;

[0116] Figure 13 This is a schematic diagram of the structure of a user behavior prediction device according to the ninth embodiment of this application;

[0117] Figure 14 This is a schematic diagram of a model output result obtaining device according to the tenth embodiment of this application;

[0118] Figure 15 This is a schematic diagram of a model output result obtaining device according to the eleventh embodiment of this application;

[0119] Figure 16 This is a schematic diagram of a model output result obtaining device according to the twelfth embodiment of this application;

[0120] Figure 17 This is a schematic diagram of the structure of a model training device according to the thirteenth embodiment of this application;

[0121] Figure 17A This is a schematic diagram of the structure of an object recommendation device according to the fourteenth embodiment of this application;

[0122] Figure 18 This is a schematic diagram of the logic structure of an electronic device provided in the fifteenth embodiment of this application. Detailed Implementation

[0123] To make the objectives, advantages, and features of this application clearer, the following detailed description, in conjunction with the accompanying drawings and specific embodiments, provides a method for object recommendation, user behavior prediction, model output result acquisition, and model training proposed in this application. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the invention can be implemented in many ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0124] It should be noted that in the description of this application, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance, or a specific order or sequence. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances. Furthermore, in the description of this application, unless otherwise stated, the term "multiple" refers to two or more. The term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices.

[0125] First, the terms and concepts involved in one or more embodiments of this application will be explained.

[0126] The user action prediction model, also known as the adaptive bottom-up spatiotemporal model (BASM model), adaptively adapts to spatiotemporal data distribution, thereby improving the model's fitting ability. This adaptive bottom-up spatiotemporal model includes a spatiotemporal-aware embedding layer, a spatiotemporal semantic transformation layer, and a spatiotemporal adaptive bias tower. Each of these layers can be processed independently, unaffected by other modules; however, they can also be used in combination to achieve superior data processing results. As a data prediction model, this user action prediction model can predict user actions on recommended objects. For example, it can predict at least one of the following: click-through rate, add-to-cart rate, purchase rate, and repurchase rate for target users on candidate objects. Furthermore, the model's own training optimization can be achieved based on the predicted user actions on recommended objects.

[0127] Click-through rate (CTR) is the probability that a user clicks to view an object when it is shown to them. It is calculated as: CTR = Clicks / Impressions, where impressions refer to the number of times the object is shown to the user, and clicks refer to the number of times the object is viewed by the user.

[0128] Next, in order to facilitate understanding of the various embodiments of this application, the application background of the embodiments will be explained.

[0129] User interests change with time and space, exhibiting diverse transformations. For example, a user might be interested in coffee at their workplace for breakfast, but at home, they might prefer soy milk. Similarly, over time, this same user might be interested in fast food at their workplace for lunch, but at home, they might prefer noodles. This demonstrates that user interests are influenced by temporal and spatial factors, meaning they diversify across different time and space contexts. Therefore, the role of temporal and spatial factors in recommending target users cannot be ignored. If recommendations can be based on temporal and spatial factors, more accurate capture of user interests can be achieved, improving the precision of recommended items.

[0130] Because user interests vary across different spatiotemporal contexts, complex spatiotemporal data distributions exist. Traditional recommendation models are ill-suited for situations with diverse user interests, and these complex data distributions further complicate their learning process. Current technologies typically involve processing all user sample data to train a set of static model parameters for use in user recommendations. This approach makes capturing diverse user interests across different spatiotemporal contexts challenging, and the modeling process often ignores the influence of spatiotemporal factors. Therefore, recommendation models using static parameters generally produce mediocre results. While dynamic parameter-based recommendation models exist, these methods typically ignore spatiotemporal factors, rely on predefined parameters, struggle to adaptively fit spatiotemporal data distributions, and are unsuitable for complex data distributions. Recommendation models, as tools for acquiring user recommendations, need to obtain highly accurate, well-matched, and compatible recommendations, as well as improve their data fitting capabilities to enhance information delivery efficiency. To facilitate understanding of the embodiments, the embodiments of this application are all illustrated using the scenario of a target user obtaining recommendations for takeout products.

[0131] In view of this, this application provides a method for object recommendation, a method for user behavior prediction, a method for obtaining model output results, a method for obtaining model output results, a method for model training, and related devices, electronic devices, and storage media, aiming to predict user behavior and obtain recommended objects. Naturally, the method of this application can also be applied to other real-world scenarios that require object recommendation. This application does not limit the specific application field; for the sake of convenience, the following uses the example of an e-commerce platform recommending suitable takeaway products to target users (users ordering food online) as a typical application scenario, and the above-mentioned method provided by this application is described around this scenario. Of course, the various embodiments of this application can also be applied to shopping scenarios, video playback scenarios, advertising interaction scenarios, etc.

[0132] To facilitate understanding of the method embodiments of this application, their application scenarios are described. Please refer to... Figure 1 , Figure 1 This is a schematic diagram illustrating an application scenario of the method embodiment of this application. It can be applied to cloud servers; however, this application scenario is merely an illustrative example and not intended to limit the specific application scenario. Figure 1 As shown, in this application scenario, a client 101 and a cloud server 102 are configured. In this embodiment, the client 101 and the cloud server 102 establish a connection directly through network communication. The client program runs on the client 101, and the server program runs on the server of the cloud server 102.

[0133] The client 101, also known as a user terminal or workstation, corresponds to the server and is controlled and managed by the server. In this embodiment, the client can be understood as a client device running a specific application, such as a mobile phone or tablet, or more specifically, as an application (APP) used to provide local services to the user. The client device can also be an in-vehicle device, wearable device, terminal device, logical entity, or a communication device such as a gateway, base station, or controller, or an Internet of Things (IoT) device such as a sensor, electricity meter, or water meter. A specific communication connection needs to be established between the client 101 and the cloud server 102 to provide various services to the user. In various embodiments of this application, the client 101 is used to obtain target user-related data, which can be converted into a vector in the server of the server. Furthermore, the client 101 can send page content request messages to request page content.

[0134] The cloud server 102 possesses high computing power. The server of the cloud server 102 has high-speed central processing unit (CPU) computing power, long-term reliable operation, powerful input / output (I / O) external data throughput capability, and better scalability. The server 102 can be a single server or a server cluster. The cloud server 102 is one end where the user operation prediction model is deployed. Of course, the user operation prediction model of this application can also be directly deployed on the client 101; this embodiment is for illustration only and not a practical limitation. In specific implementation, the server of the cloud server 102 is equipped with a user operation prediction model (BASM model). The server is used to receive page content request messages from the client 101, as well as the relevant data of the target user transmitted, and converts the data into vectors as input to the user operation prediction model. The input of the BASM model includes the target user characteristics, the candidate target object characteristics, and the spatiotemporal environment characteristics corresponding to the target user. Through the processing of this model, the judgment result of the candidate target object as the recommended object of the target user can be obtained, and the matched recommended object is transmitted to the client 101 in the form of page content through network communication connection, so that the client 101 can expose and display it to the target user.

[0135] Client 101 can communicate with cloud server 102. Client 101 and cloud server 102 can communicate using various communication systems, such as wired or wireless communication systems. Wireless communication systems can include, for example, Global System for Mobile Communications (GSM), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), General Packet Radio Service (GPRS), Long Term Evolution (LTE), LTE Frequency Division Duplex (FDD), LTE Time Division Duplex (TDD), Universal Mobile Telecommunication System (UMTS), Worldwide Interoperability for Microwave Access (WiMAX), future 5th Generation (5G) systems or New Radio (NR), and satellite communication systems.

[0136] In this embodiment, client 101 is mainly used to transmit relevant data of the target user, such as click behavior data, browsing behavior data, and add-to-cart behavior data, to the server of cloud server 102 via network communication connection. Client 101 also needs to send a page content request message to request page content. The server of cloud server 102 receives the page content request message sent by client 101. Since the server is equipped with a user action prediction model (BASM model), the server performs vectorization processing on the acquired data to obtain the target user features, candidate target object features, and spatiotemporal environment features corresponding to the target user. These are used as inputs to the BASM model to obtain the prediction data of candidate target objects and the judgment result of the candidate target objects as recommended objects for the target user. The server obtains the matching recommended objects based on the judgment result of the recommended objects and transmits the recommended objects to client 101 in the form of page content for exposure to the target user.

[0137] First Embodiment

[0138] Based on this, the following describes the method embodiments provided in this application. Please refer to... Figure 2 This is a flowchart illustrating the first embodiment of this application. The following is a summary of the process. Figure 2 The method will be described in detail below.

[0139] S201, obtain the target user characteristics of the target user, the candidate target object characteristics of the candidate target object, and the spatiotemporal environment characteristics corresponding to the target user.

[0140] This step is used to obtain the input data for the user action prediction model (adaptive bottom-up spatiotemporal model, BASM model), which includes: target user features, candidate target object features, and spatiotemporal environment features.

[0141] The target user is a user with a need for object recommendations. The target user type includes users of any form with a need for object recommendations, such as natural persons, animals, and machines. For example, a food delivery robot may need information on recommended delivery locations, a user in a shopping application may need recommendations for browsing products, and a pet in a real-world scenario may need recommendations for pet toys. The object can be any form of matter existing in a real-world scenario, such as virtual information or physical items. For ease of understanding, this application's embodiments use natural persons as the target user and takeout products in a shopping application as examples for illustration.

[0142] The target user's user feature refers to the user characteristics of the target user, used to represent the user's attribute features. The target user's user feature can unify the user's various attribute features into a single feature for holistic expression. The user attribute features originate from the user's basic information. For example, basic user information includes: education level, role, marital status, etc. Correspondingly, user attribute features can include: female, 28 years old, student, from Shaanxi, unmarried, etc. Of course, user attribute features can also include the user's behavioral habits, interests, preferences, and needs. By integrating multiple user attribute features representing the user into one user feature, the resulting user feature can represent the user's attributes. In specific implementation, the aforementioned obtained target user feature can be represented as "X". j＝1 "; where "j" is a subscript identifier, used to distinguish the target user features from other features in this embodiment.

[0143] The candidate target object refers to the pre-selected object in the object recommendation pre-selection pool. This pre-selected object serves as the source of the recommended object for the user operation prediction model; that is, whether the candidate target object can be used as the recommended object for the target user requires further judgment in subsequent steps. The pre-selected objects in the object recommendation pre-selection pool can be selected based on object information data pre-stored in the database. Furthermore, the process of selecting candidate target objects in the object recommendation pre-selection pool can be random selection by the model or pre-set manually; this embodiment does not specifically limit this. Illustratively, the candidate target object includes at least one of the following: candidate product, candidate merchant, candidate benefit, and candidate promotional activity.

[0144] The candidate target object feature refers to the attribute feature representing the pre-selected object itself. The candidate target object feature can be used to characterize one or more attributes of the candidate target object. The candidate target object feature is a feature vector formed by uniformly organizing the attribute values ​​corresponding to one or more attributes of the candidate target object.

[0145] In specific implementation, the candidate target object features include candidate item features. Combining multiple item attribute features of the candidate item yields a single candidate item feature representing the target item. In other words, the aforementioned candidate item feature is a holistic expression of the product's numerous item attribute features unified into one feature. These item attribute features include at least: static item attribute features and dynamic item attribute features. Static item attribute features represent the inherent attributes of the product, such as color, shape, material, and brand identity; dynamic item attribute features represent attributes that change with the product's flow, such as price, inventory, monthly sales volume, distributor, and manufacturer. By unifying multiple item attribute features of the target item into a holistic candidate item feature, the characteristic expression of the candidate item can be achieved. Of course, the process of forming a candidate item feature can also involve unifying all item attribute features of the product or unifying only one item attribute feature. This embodiment is merely an illustrative example and not a practical limitation. In specific implementation, the aforementioned obtained candidate object feature can be represented as "X". j＝2 ".

[0146] The spatiotemporal environmental features corresponding to the target user refer to spatiotemporal features (Context Features), also known as contextual features, which represent the characteristics of the time and space scene in which the target user is located. In spatiotemporal features, both the temporal features representing the target user's time and the spatial features representing their location coexist. Illustratively, these spatiotemporal features include today's day of the week, the current time period (e.g., breakfast time, afternoon tea time), the current time point, and the current geographical location (e.g., city, city region, geographical location code). For example, when a target user searches for and places an order for food delivery product A in a food delivery application, the target user's current time period is afternoon tea time, and their current location is a high-tech zone in the city. By obtaining the temporal features corresponding to the afternoon tea time period and the spatial features corresponding to the high-tech zone in the city, spatiotemporal features representing both of these characteristics can be obtained. It should be understood that the target user in the spatiotemporal environmental features corresponding to the target user and the target user features of the target user mentioned above refer to the same user; furthermore, the spatiotemporal features of the target user mentioned above can be the target user's historical spatiotemporal features or the target user's current spatiotemporal features. In practical implementation, the spatiotemporal environmental characteristics can be represented as "X". c ".

[0147] In this embodiment, the target user features, candidate target object features, and spatiotemporal environment features corresponding to the target user obtained above are all vectorized features. In other words, each of the features obtained above is a vector, and this vectorized feature can be used as input data for the user operation prediction model (adaptive bottom-up spatiotemporal model, BASM model) for data recognition processing of the BASM model.

[0148] S202, the target user features, the candidate target object features, and the spatiotemporal environment features are input into the user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object.

[0149] This step is used to obtain prediction data on the target user's performance of a specified operation on a candidate target object.

[0150] The user action prediction model refers to an adaptive bottom-up spatiotemporal model (BASM model). This model can obtain prediction data for a specified action based on the characteristics of the target user, the characteristics of the candidate target object, and the spatiotemporal environment. For a better understanding of the user action prediction model (BASM model), please refer to [link to relevant documentation]. Figure 3 The indication, Figure 3 A model architecture diagram for the user-operated prediction model (BASM model).

[0151] The specified operation refers to the action performed by the target user on the candidate target object. Examples include clicking, adding to favorites, adding to cart, sharing, and placing an order. In other words, the specified operation is the operation that the target user operation prediction model needs to estimate data for. By estimating the data of the target user and the candidate target object during the specified operation using the target user operation prediction model, the predicted data can be obtained. The data type of the predicted data corresponds to the operation type of the specified operation. For example, the predicted data type could be click-through rate, order rate, repurchase rate, or share rate. If the operation type of the specified operation is a favorites action, then the data type of the predicted data is a favorites data type.

[0152] In the user operation prediction model, at least one of the following processes is performed:

[0153] The first method involves assigning weights to the target user features and the candidate target object features based on the spatiotemporal environment characteristics, and obtaining the prediction data based on the target user feature weights and the candidate target object feature weights.

[0154] The second method involves obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the candidate target object features, and then obtaining the prediction data based on the spatiotemporal semantic features.

[0155] The third method involves obtaining fused output features with spatiotemporal deviation characteristics based on the spatiotemporal environment features, the target user features, and the candidate target object features, and then obtaining the predicted data based on the fused output features.

[0156] Next, each processing method of the user operation prediction model (adaptive bottom-up spatiotemporal model, BASM model) will be explained in detail.

[0157] The first method involves assigning weights to the target user features and the candidate target object features based on the spatiotemporal environment characteristics, and obtaining the prediction data based on the target user feature weights and the candidate target object feature weights.

[0158] To better understand the process of assigning weights to target user features based on spatiotemporal environmental characteristics, please refer to [reference needed]. Figure 4 The indication, Figure 4 This diagram illustrates the process of assigning weights to target user features based on spatiotemporal environmental characteristics. It's important to note that this process of assigning weights to target user features based on spatiotemporal environmental characteristics is implemented within the Spatiotemporal-Aware Embedding Layer of the User Action Prediction Model (BASM model).

[0159] Here, we will first explain the Spatiotemporal-Aware Embedding Layer. The Spatiotemporal-Aware Embedding Layer is used to dynamically perceive spatiotemporal context. As an embedding layer in a neural network, it better models spatiotemporal structural information by fusing spatiotemporal information into node embeddings. In this embedding layer, traditional convolutional neural networks or attention mechanisms are first used to extract features from nodes. Then, spatiotemporal information is introduced by encoding time-series data and spatial location data, fusing this information into the node embeddings to obtain an embedding representation containing spatiotemporal information. The Spatiotemporal-Aware Embedding Layer is widely used in object recommendation, such as product recommendation, traffic flow prediction, video classification, and behavior recognition, effectively capturing the spatiotemporal evolution patterns in data.

[0160] The step of assigning weights to the target user features based on the spatiotemporal environmental features includes:

[0161] S202-1-1. Connect the spatiotemporal environment features with the target user features to obtain spatiotemporal user connection features;

[0162] The spatiotemporal environment features are spatiotemporal environment feature vectors, and the target user features are target user feature vectors; in specific implementation, refer to... Figure 4 To illustrate, the spatiotemporal environment feature vector is represented by the symbol "X". c The target user feature vector is represented by the symbol "X". j＝1 The symbol "j" indicates the subscript order, for easy distinction from other vectors mentioned later. The spatiotemporal environmental feature "X" is used... c "and the target user feature vector "X j＝1 "Performing concatenation enables the target user's feature vector to possess the characteristics of the target user's spatiotemporal information. It's important to understand that due to the spatiotemporal environmental feature 'X'..." c "X" represents the spatiotemporal environmental characteristics corresponding to the target user. Therefore, when the spatiotemporal environmental characteristics of the target user change, the spatiotemporal user connection characteristics obtained through connection processing will also change accordingly. For example, if the spatiotemporal environmental characteristics "X" are input into the BASM model... c "If the spatiotemporal environment features of the target user's current spatiotemporal environment are given, then the spatiotemporal user connection features obtained through connection processing are the current spatiotemporal user connection features; if the spatiotemporal environment features "X" of the input BASM model are given..." c "If the spatiotemporal environment characteristics are the spatiotemporal environment characteristics of the target user's historical spatiotemporal environment, then the spatiotemporal user connection characteristics obtained by the connection processing are the historical spatiotemporal user connection characteristics."

[0163] The concatenation process, also known as combined concatenation, is a common method for integrating multi-channel features. It involves concatenating two or more feature vectors to obtain a new feature vector. In this new feature vector, the feature dimension is increased, while the information of the original feature vectors is preserved. In this embodiment, by concatenating the spatiotemporal environment feature vector with the target user feature vector, the resulting spatiotemporal user concatenation feature retains both the information of the target user's spatiotemporal environment and the information of the target user's features. For specific implementation, refer to... Figure 4 Illustrative symbols used for connection processing express.

[0164] The spatiotemporal environment features are concatenated with the target user features to obtain spatiotemporal user connection features. This includes concatenating the spatiotemporal environment feature vector with the target user feature vector to obtain a spatiotemporal user connection feature vector. For specific implementation, refer to... Figure 4 As illustrated, the above refers to the spatiotemporal environmental feature vector "X". c "and target user feature vector "X j＝1 The spatiotemporal user connection feature vector obtained by splicing is represented by the symbol "X". j＝1 ;X c This is indicated by "". It is important to note that the spatiotemporal environmental feature vector "X" used for connection processing mentioned above... c " is a vector that has undergone zero-value initialization.

[0165] S202-1-2. Perform fully connected layer processing on the spatiotemporal user connection characteristics to obtain the spatiotemporal user fully connected processing result;

[0166] The fully connected layer (Multilayer Perceptron, MLP layer), also known as a multilayer perceptron, is an artificial neural network structure with a directional architecture. In this layer, a set of input vectors is mapped to a set of output vectors. Besides the input and output layers, it can have multiple hidden layers, which can be viewed as a directed graph composed of multiple node layers, each fully connected to the next. In addition to the input nodes, each node is a neuron with a non-linear activation function. In this embodiment, by using the spatiotemporal user connectivity features as the input vector of the fully connected layer, the corresponding output vector spatiotemporal user fully connected processing result can be obtained.

[0167] The step of performing fully connected layer processing on the spatiotemporal user connection features to obtain the spatiotemporal user fully connected processing result includes: obtaining the spatiotemporal user fully connected processing result vector based on the spatiotemporal user connection feature vector, weight parameters of the fully connected layer, and bias parameters of the fully connected layer. In specific implementation, the spatiotemporal user fully connected processing result vector is represented by the symbol "X". j＝1 ;X c "" indicates that the weights of the MLP layer are assigned to the weights of the fully connected layer using the symbol "W". p "b" indicates that the weights of the MLP layer are assigned to the bias parameters of the fully connected layer. p "Indicates". (Reference) Figure 4 This illustrates that, based on the spatiotemporal user connection feature vector "X", j＝1 ;X c Weight allocation of fully connected layer weight parameters "W" p "Weight allocation of fully connected layer bias parameters "b" p The spatiotemporal user fully connected processing result vector output by the MLP layer is represented as "W". p [X j＝1 ;X c ]+b p ".

[0168] It is important to understand that in the user action prediction model of this embodiment, the weights of the fully connected layer are assigned to the weight parameter "W". p "and weight allocation of fully connected layer bias parameters "b p These are all parameters pre-set by the user in the predictive model (adaptive bottom-up spatiotemporal model, BASM model). The specific values ​​of these parameters can be adjusted during the BASM model training process.

[0169] S202-1-3. The spatiotemporal user fully connected processing result is used as the independent variable of the activation function to obtain the target user feature weights.

[0170] The activation function (sigmoid), also known as the sigmoid function or sigmoid growth curve, is a commonly used non-linear activation function that normalizes data using a non-linear method; it is typically used in regression prediction and binary classification models. In this embodiment, the spatiotemporal user fully connected processing result can be used as the independent variable of the activation function to calculate the corresponding target user feature weights. For specific implementation, refer to... Figure 4 To illustrate, the target user feature weights are represented by the symbol "α". j＝1 The term "" indicates that the target user feature weight "α" is calculated in this embodiment. j＝1 The formula for " is:

[0171] α j =2*δ(Wp [x j ;x c ]+b p ) Formula 1-1

[0172] In Formula 1-1, the symbol "j" is the subscript order symbol, used to distinguish it from other vectors mentioned later; in this embodiment, "j = 1"; "α" j＝1 " represents the target user feature weights; "δ" represents the sigmoid activation function; "W" represents the target user feature weights. p "" indicates the weight allocation of the MLP layer and the weight parameters of the fully connected layer; "b p "X" indicates the weight allocation of the MLP layer and the bias parameters of the fully connected layer; j＝1 "X" represents the target user feature vector. c " represents the spatiotemporal environment feature vector.

[0173] Reference Figure 4 The illustration shows the obtained spatiotemporal environmental feature vector "X". c "and the target user feature vector "X" initialized with zero values j＝1 Perform combined connections between vectors This leads to the acquisition of the spatiotemporal user connection feature vector "X". j＝1 ;X c The spatiotemporal user connection feature vector "X" j＝1 ;X c "As the input vector of the MLP layer, after vector mapping, the fully connected processing result vector "W" is output. p [X j＝1 ;X c ]+b p ";Then process the fully connected result vector "W p [X j＝1 ;X c ]+b p "As the independent variable of the activation function (sigmoid), the target user feature weight "α" is calculated using the above formula 1-1." j＝1 ".

[0174] The step of assigning weights to the candidate target object features based on the spatiotemporal environment features includes: concatenating the spatiotemporal environment features with the candidate target object features to obtain spatiotemporal object connection features; in this embodiment, the spatiotemporal environment features are spatiotemporal environment feature vectors, and the candidate target object features are candidate target object feature vectors; it should be understood that the process of obtaining the candidate target object features and assigning weights is similar to the process of obtaining the target user features and assigning weights mentioned above, therefore, the spatiotemporal environment feature vector obtained in this step is still represented by the symbol "X". cIn contrast, the feature vector of the candidate target object is represented by the symbol "X". j＝2 The step of connecting the spatiotemporal environment features with the candidate target object features to obtain spatiotemporal object connection features includes: connecting the spatiotemporal environment feature vector with the candidate target object feature vector to obtain a spatiotemporal object connection feature vector;

[0175] The connection processing (Concatenate) is explained in step S202-1-1 and will not be repeated here. Unlike the spatiotemporal user connection feature, the spatiotemporal object connection feature in this embodiment is represented by the symbol "X". j＝2 ;X c "Indicates. Refer to Figure 4 The diagram illustrates the field input, X. j At the specified location, the input vector is the feature vector of the candidate target object, "X". j＝2 ".

[0176] S202-2-2, Perform fully connected layer processing on the spatiotemporal object connection features to obtain the spatiotemporal object fully connected processing result;

[0177] In specific implementation, the spatiotemporal object connection feature "X" obtained through connection processing will be used... j＝2 ;X c "As the input vector of the MLP layer, after vector mapping processing, the spatiotemporal object fully connected processing result is obtained. The process of performing fully connected layer processing on the spatiotemporal object connection features to obtain the spatiotemporal object fully connected processing result includes: obtaining the spatiotemporal object fully connected processing result vector based on the spatiotemporal object connection feature vector, weight allocation of fully connected layer weight parameters, and weight allocation of fully connected layer bias parameters. It should be understood that the process of obtaining the spatiotemporal object fully connected processing result vector is similar to the process of obtaining the spatiotemporal user fully connected processing result vector, both based on the pre-set weight allocation of fully connected layer weight parameters 'W' in the MLP layer." p "Weight allocation of fully connected layer bias parameters "b" p This is the process of obtaining the corresponding fully connected processing result vector. Unlike the previous step, the input data variable in this step is the spatiotemporal object connectivity feature "X". j＝2 ;X c The result of the fully connected spatiotemporal object obtained in this step is represented by the symbol "W". p [X j＝2 ;X c ]+b p "express.

[0178] S202-2-3. The result of the fully connected processing of the spatiotemporal object is used as the independent variable of the activation function to obtain the feature weights of the candidate target object.

[0179] For the specific meaning of the activation function, please refer to the description in step S202-1-3; it will not be repeated here. In specific implementation, the fully connected result of the spatiotemporal object, "W", will be processed... p [X j＝2 ;X c ]+b p "As the independent variable of the sigmoid activation function, it is calculated again using the above formula 1-1 to obtain the feature weights "α" of the candidate target object. j＝2 ".

[0180] Here, the field input involved is explained. Field embedding, also known as field coding, is a technique that maps discrete features to a continuous vector space. It is commonly used to process various types of sparse data. In this embodiment, during the model learning process, the target user feature vector "X" can be used as the target user feature vector. j＝1 ", candidate target object feature vector "X" j＝2 "The discrete features of the target user and the candidate target object are respectively used as field inputs; the discrete features of the target user and the candidate target object are converted into continuous vector representations. Specifically, this is achieved by multiplying each of the discrete features by an embedding matrix of a fixed size, which maps each different discrete feature to a continuous vector representation. Furthermore, the field inputs can also involve other types of vectors, such as combined features, which will be described in detail in the subsequent steps of the embodiments of this application. By using different types of vectors as field inputs, the spatiotemporal attention embedding layer can assign the input feature vectors to the spatiotemporal characteristics of the spatiotemporal environment of the target user. Furthermore, the candidate target object feature weights obtained in the spatiotemporal attention embedding layer are adaptive weights, which can be adjusted according to the spatiotemporal environment feature vector "X". c The weights of the target user features, α, change adaptively with the changes in the target user feature weights. j＝1 It can also be based on the spatiotemporal environmental feature vector "X" c The spatiotemporal environment feature vector "X" adapts and changes adaptively. It is important to note that the spatiotemporal environment feature vector "X" changes accordingly. c "No need to obtain weight parameters."

[0181] The step of obtaining the prediction data based on the target user feature weights and the candidate target object feature weights includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; wherein, the target user features are target user feature vectors. The step of obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features includes: performing a multiplication operation between the target user feature vector and the target user feature weights to obtain a spatiotemporally weighted target user feature vector.

[0182] In this embodiment, based on the target user feature weight "α" j＝1 "Target user feature vector "X" j＝1 The process of obtaining spatiotemporally weighted target user features is a multiplication operation. For specific implementation, refer to... Figure 4 The illustration shows that the spatiotemporally weighted target user features obtained by the above multiplication operation are represented by the symbol "h". j＝1 " indicates; multiplication operations are handled using symbols express.

[0183] Multiplication, also known as element-wise product, refers to multiplying two vectors element by element to form a new vector with the same size as the original vector. In this embodiment, the obtained target user feature weights "α" are... j＝1 "and target user feature vector "X j＝1 "Perform element-wise multiplication to obtain the spatiotemporal weighted target user features"h j＝1 "and the target user feature vector "X j＝1 "Same size. Due to the target user feature weight "α" j＝1 "Based on the spatiotemporal environmental feature vector "X c "The obtained adaptive weights, therefore, the spatiotemporal weights of the target user features obtained by the multiplication operation" j＝1 "It possesses spatiotemporal adaptability. That is, as the target user feature weights "α" change..." j＝1 The weights of "h" adapt to changes, and the spatiotemporal weights target user features "h" j＝1 "This will change accordingly. Specifically, it depends on the target user feature weight α." j＝1 "and target user feature vector "X j＝1 "Calculate and obtain the spatiotemporal weighted target user features "h" j＝1 The formula for " is as follows:

[0184] h j =α j x j Formula 1-2

[0185] In formulas 1-2, the symbol "j" is the subscript order symbol; in this implementation, "j = 1" is used to distinguish it from other vectors mentioned later. j＝1 " indicates the target user feature weight; "X j＝1 This represents the target user feature vector. (See also: [link to relevant documentation]). Figure 4 The illustration shows the target user feature weights "α" obtained through adaptive weighting. j＝1 "After being compared with the target user feature vector "X" j＝1Perform element-wise multiplication Able to obtain spatiotemporal weighted target user features "h" j＝1 Through the above process, the spatiotemporal weighted target user feature "h" is... j＝1 It gives it adaptive characteristics in the time and space environment of the target user.

[0186] Based on the candidate target object feature weights and the candidate target object features, spatiotemporal weighted candidate target object features are obtained; wherein, the candidate target object features are candidate target object feature vectors. The step of obtaining spatiotemporal weighted candidate target object features based on the candidate target object feature weights and the candidate target object features includes: performing a multiplication operation between the candidate target object feature vector and the candidate target object feature weights to obtain a spatiotemporal weighted candidate target object feature vector. The multiplication operation, as described above, is the process of multiplying vectors element-wise to form a new vector. In this embodiment, based on the candidate target object feature weights "α", j＝2 ", candidate target object feature vector "X" j＝2 ", to obtain the features of the candidate target object as the feature vector "h" j＝2 The process of obtaining the spatiotemporally weighted target user features "h" described above is similar to the process of obtaining the spatiotemporally weighted target user features "h" with spatiotemporal weights. j＝1 The process is similar. Both involve element-wise multiplication, where the weights of the candidate target object features are multiplied by "α". j＝2 "and the feature vector of the candidate target object "X j＝2 "By performing the product operation again using formulas 1-2 above, we can obtain the spatiotemporal weighted candidate target object features with spatiotemporal weights." j＝2 ".

[0187] Based on the spatiotemporal weighted target user features and the spatiotemporal weighted candidate target object features, fused output features are obtained; based on the fused output features, the predicted data is obtained. In this embodiment, the spatiotemporal weighted target user features "h" are obtained through a spatiotemporal attention embedding layer. j＝1 ", Spatiotemporal weighted candidate target object features "h j＝2 The two spatiotemporal weighted features mentioned above can be used to obtain the fused output features. In specific implementation, the spatiotemporal weighted target user feature "h" is... j＝1 ", Spatiotemporal weighted candidate target object features "h j＝2 "After processing by the Spatiotemporal Semantic Transform Layer and the Spatiotemporal Adaptive BiasTower, the fused output features are obtained."

[0188] The step of obtaining the fused output features based on the spatiotemporal weighted target user features and the spatiotemporal weighted candidate target object features includes: obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporal weighted target user features, and the spatiotemporal weighted candidate target object features; and obtaining the fused output features based on the spatiotemporal semantic features. In specific implementation, the spatiotemporal environment feature vector "X" is used. c ", Spatiotemporal weighted target user features" h j＝1 ", Spatiotemporal weighted candidate target object features "h j＝2 The obtained spatiotemporal semantic features are represented by symbols. This is indicated. For easier understanding, please refer to... Figure 3 The illustration.

[0189] The step of obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporal weight target user features, and the spatiotemporal weight candidate target object features includes: obtaining original semantic features based on the spatiotemporal weight target user features and the spatiotemporal weight candidate target object features; specifically, concatenating the spatiotemporal weight target user features and the spatiotemporal weight candidate target object features to obtain the original semantic features; in this embodiment, the original semantic features are represented by symbols. This indicates that the above is based on the spatiotemporal weighted target user feature "h". j＝1 ", Spatiotemporal weighted candidate target object features "h j＝2 "By performing concatenation, we can obtain the original semantic features." The connection process described above, also known as combinatorial connection, is a common method for integrating multi-channel features. It involves concatenating two or more feature vectors to obtain a new feature vector. For ease of understanding, refer to the following formula to illustrate combinatorial connection:

[0190]

[0191] In formulas 1-3, the symbol "n" represents the number of types of input features. In specific implementation, the above input features include: spatiotemporal weighted target user feature "h". j＝1 ", Spatiotemporal weighted candidate target object features "h j＝2 In this formula, "h0; ...; h n-1 " and "h j＝1 “h” j＝2 The correspondence between "" and "" can be set according to actual needs. For example, "h" can be set as "h". j＝1 "corresponding to "h0" in the formula, "h" j＝2 "This corresponds to "h1" in the formula; of course, subscript misalignment can be used for correspondence, but this example is for illustration only and is not a practical limitation. Through the original semantic features..." Obtaining spatiotemporal semantic features The process is implemented in the spatiotemporal semantic transformation layer. For a better understanding of the transformation process from original semantic features to spatiotemporal semantic features, please refer to [link / reference needed]. Figure 5 The indication, Figure 5 The original semantic features of the embodiments of this application Spatiotemporal semantic features A diagram illustrating the transformation.

[0192] Here, we will first explain the Spatiotemporal Semantic Transform Layer. The Spatiotemporal Semantic Transform Layer explicitly transforms the original semantic features into spatiotemporal semantic features, thereby increasing the spatiotemporal semantic expression in different spatiotemporal contexts. As a deep learning layer, it can convert the original semantic information in the original semantic feature vector into spatiotemporal semantic information. In this layer, a convolutional neural network is first used to extract features from the original semantic information. Then, spatiotemporal information is introduced, and feature fusion is performed in the temporal and spatial domains respectively. Next, further processing is performed through self-attention mechanisms or convolutional operations to transform it into spatiotemporal semantic information, thus obtaining spatiotemporal semantic features.

[0193] In this embodiment, the original semantic features refer to the features obtained by concatenating the output vectors of the spatiotemporal attention embedding layer, and are used as one of the input features of the spatiotemporal semantic transformation layer. These original semantic features are the initial features used in the spatiotemporal semantic transformation layer, and these features typically reflect the characteristics and properties of the data itself. The spatiotemporal semantic features are the output features of the spatiotemporal semantic transformation layer.

[0194] Based on the spatiotemporal environmental features, the weight parameters and bias parameters of the spatiotemporal semantic fully connected layer are obtained; based on the original semantic features, the weight parameters and bias parameters of the spatiotemporal semantic fully connected layer, the spatiotemporal semantic fully connected layer are obtained, and the spatiotemporal semantic features are obtained. In specific implementation, the method further includes: obtaining the target user's historical behavior sequence matching the spatiotemporal environmental features. For ease of understanding, the target user's historical behavior sequence matching the spatiotemporal environmental features is represented by the symbol "h". ui " indicates; the obtained spatiotemporal semantic fully connected layer weight parameters are respectively represented by the symbol "W w “W” b " indicates; the obtained spatiotemporal semantic fully connected layer bias parameters are respectively represented by the symbol "b w “b” b This indicates that the weight parameter "W" in the aforementioned spatiotemporal semantic fully connected layer is used to represent this. w “W” b", Spatiotemporal semantic fully connected layer bias parameter "b w “b” b "These are also parameters preset in the BASM model. The specific values ​​of these parameters can be adjusted during the training of the BASM model."

[0195] The step of obtaining the spatiotemporal semantic fully connected layer weight parameters and bias parameters based on the spatiotemporal environment features includes: inputting the spatiotemporal environment features and the target user's historical behavior sequence matching the spatiotemporal environment features into the meta-network to obtain the spatiotemporal semantic fully connected layer weight parameters and bias parameters. In this embodiment, the obtained spatiotemporal semantic fully connected layer weight parameters are represented by the symbol "W". stl "b" indicates that the obtained spatiotemporal semantic fully connected layer bias parameters are represented by the symbol "b". stl The above indicates that the weight parameters "W" of the spatiotemporal semantic fully connected layer are obtained. stl ", Spatiotemporal semantic fully connected layer bias parameter "b stl "To be used in subsequent steps to calculate spatiotemporal semantic features" Specifically, the spatiotemporal semantic fully connected layer weight parameter "W" is obtained. stl ", Spatiotemporal semantic fully connected layer bias parameter "b stl For the process described below, see the description of the following section.

[0196] The meta-network refers to a special type of neural network used to generate parameters for other neural networks. Meta-networks are used to generate dynamic parameter weights and biases to adapt to input features such as spatiotemporal environmental features "X". c The change in "W". In this embodiment, the meta-network is used to generate the spatiotemporal semantic fully connected layer weight parameters "W". stl "and the spatiotemporal semantic fully connected layer bias parameter "b stl The advantage of using metanetworks is that they allow the model to adaptively adjust parameters without requiring manual parameter adjustments or model retraining. Furthermore, metanetworks can accelerate model training and inference by sharing parameters across different tasks. Specifically, the step of inputting the spatiotemporal environmental features and the target user's historical behavior sequence matching the spatiotemporal environmental features into the metanetwork to obtain the weight parameters and bias parameters of the spatiotemporal semantic fully connected layer includes: concatenating the spatiotemporal environmental features and the target user's historical behavior sequence matching the spatiotemporal environmental features to obtain spatiotemporal historical behavior features; and obtaining the weight parameters and bias parameters of the spatiotemporal semantic fully connected layer based on the spatiotemporal historical behavior features, the metanetwork weight parameters, and the metanetwork bias parameters.

[0197] The source of the aforementioned target user's historical behavior sequence is the target user's original historical behavior sequence. Therefore, the method further includes: obtaining the target user's original historical behavior sequence; in this embodiment, the target user's original historical behavior sequence is represented by the symbol "x". j＝3 "" indicates; through the target user's original historical behavior sequence "x j＝3 "Able to obtain the target user's historical behavior sequence" ui "X" and its spatiotemporal environmental characteristics c "Perform combined connections" Able to obtain the concatenated vector "X" c h ui ", i.e., spatiotemporal historical behavioral characteristics; the spatiotemporal historical behavioral characteristics "X" are spliced ​​together as described above. c h ui "As input data for the meta-network."

[0198] The target user's original historical behavior sequence refers to a sequence of behavioral objects that the target user has previously performed, arranged chronologically according to the order in which the actions occurred. Illustratively, these behaviors may include: clicking, adding to favorites, adding to cart, sharing, placing an order, repeat purchases, and deleting records; correspondingly, behavioral objects may include products, videos, documents, and data tables. It is important to understand that while all of these behavioral objects are exposed within the user's client application, not all objects exposed within the application are behavioral objects. The behavioral objects mentioned above are those objects that the user has interacted with within the application. In other words, the target user's behavioral objects refer to the objects within the application that the target user has interacted with. In the target user's original historical behavior sequence, the user's behavioral objects and actions exist in key-value pairs.

[0199] The target user's historical behavior sequence (spatiotemporal filtering behavior) is a feature vector obtained by filtering the target user's original historical behavior sequence based on spatiotemporal environmental feature vectors. Obtaining the target user's historical behavior sequence matching the spatiotemporal environmental features includes: filtering out target user historical behavior sequences that match the spatiotemporal environmental features from the target user's original historical behavior sequence. For specific implementation, refer to... Figure 5 As illustrated above, the target user's historical behavior sequence "h" is selected. ui The process is as follows: based on the spatiotemporal environmental feature vector "X" c "The original historical behavior sequence of the target user" x j＝3 "Perform spatiotemporal filtering to obtain the filtered target user's historical behavior sequence."ui The filtered target user's historical behavior sequence "h" ui "X" and its spatiotemporal environmental characteristics c "Combined connections are made for data processing of the source network. It is important to understand that due to the spatiotemporal environmental characteristics of 'X'..." c "This refers to the characteristics of the spatiotemporal environment in which the target user resides. Therefore, by defining the spatiotemporal environment characteristics 'X'..." c "The original historical behavior sequence of the target user" x j＝3 "By performing filtering, we can obtain X that conforms to the characteristics of the spatiotemporal environment." c The spatiotemporal characteristics of the target user's historical behavior sequence "h" ui In this embodiment, the spatiotemporal semantic fully connected layer weight parameters calculated in the meta-network are as follows:

[0200] W stl =Reshape(W w [h c h ui ]+b w ) Formula 1-4

[0201] In formulas 1-4, "W stl " represents the weight parameters of the spatiotemporal semantic fully connected layer; "Reshape" represents the vector reshaping operation; "W" represents the vector reshaping operation. w "W" represents the weight parameter "W" used to calculate the spatiotemporal semantic fully connected layer. stl The meta-network weight parameters; "b w "W" represents the weight parameter "W" used to calculate the spatiotemporal semantic fully connected layer. stl The meta-network bias parameters; "h c "In the formula, it represents the spatiotemporal environmental characteristics, namely, the "X" mentioned above." c ";"h ui "" represents the target user's historical behavior sequence. It should be understood that, in this embodiment, the meta-network weight parameter "W" in the above formula... w "、Meta-network bias parameter "b w "These are all parameters pre-set by the user operation prediction model (adaptive bottom-up spatiotemporal model, BASM model). The specific values ​​of these parameters can be adjusted during the BASM model training process. In this embodiment, the formula for calculating the spatiotemporal semantic fully connected layer bias parameters in the meta-network is as follows:"

[0202] b stl =Reshape(W b [h c h ui ]+b b ) Formula 1-5

[0203] In formulas 1-5, “b stl " represents the bias parameters of the spatiotemporal semantic fully connected layer; "Reshape" represents the vector reshaping operation; "W" represents the vector reshaping operation. b "b" represents the bias parameter used to calculate the spatiotemporal semantic fully connected layer. stl The meta-network weight parameters; "b b "b" represents the bias parameter used to calculate the spatiotemporal semantic fully connected layer. stl The meta-network bias parameters; "h c "In the formula, it represents the spatiotemporal environmental characteristics, namely, the "X" mentioned above." c ";"h ui "" represents the target user's historical behavior sequence. It should be understood that, in this embodiment, the meta-network weight parameter "W" in the above formula... b "、Meta-network bias parameter "b b "Similarly, as a preset parameter of the BASM model, it can be adjusted."

[0204] Here, we will explain "Reshape" in detail. "Reshape is a reshaping operation, typically used to transform input data into a shape suitable for a specific neural network layer. Specifically, it can change the shape of the tensor corresponding to the input data vector, transforming a tensor into a new shape containing the same number of elements, but with a different arrangement. Reshape converts the output embedding of the meta-network into a dynamic weight matrix. The obtained spatiotemporal semantic fully connected layer weight parameters 'W'..." stl ", Spatiotemporal semantic fully connected layer bias parameter "b stl "The spatiotemporal semantic features are obtained through calculation using the following formula:"

[0205]

[0206] In formulas 1-6, Represents spatiotemporal semantic features; "W stl " represents the weight parameters of the spatiotemporal semantic fully connected layer; Indicates the original semantic features; "b stl "" represents the bias parameters of the spatiotemporal semantic fully connected layer.

[0207] For ease of understanding, combined with Figure 5 The illustration in the diagram is explained in detail. In the spatiotemporal semantic transformation layer, the spatiotemporal environment feature vector "X" is transformed... c "x" represents the original historical behavior sequence of the target user. j＝3 " and original semantic features As input data for this module; the original historical behavior sequence of the target user "x" j＝3 "Based on the spatiotemporal environmental feature vector "X" c"Perform filtering to obtain the target user's historical behavior sequence under specific spatiotemporal characteristics." ui "; to transfer the spatiotemporal environmental feature vector "X" c "and the target user's historical behavior sequence" h ui "By combining and connecting the elements, we can obtain the spatiotemporal historical behavioral characteristics of 'X'." c h ui "; to define the spatiotemporal historical behavioral characteristics "X" c h ui "As input to the meta-network, the spatiotemporal semantic fully connected layer weight parameters "W" are obtained through calculations using formulas 1-4 and 1-5. stl ", Spatiotemporal semantic fully connected layer bias parameter "b stl "; to transform the above original semantic features The weight parameter "W" of the spatiotemporal semantic fully connected layer stl ", Spatiotemporal semantic fully connected layer bias parameter "b stl "As the input to the MLP layer of the spatiotemporal semantic transformation layer, the spatiotemporal semantic features are obtained through the operations of formulas 1-6."

[0208] Through the above steps, the spatiotemporal semantic transformation layer transforms the original semantic features into spatiotemporal semantic features. Based on the obtained spatiotemporal semantic features, the process of obtaining the fused output features is implemented in the spatiotemporal adaptive bias tower of the BASM model.

[0209] Here, we will first explain the Spatiotemporal Adaptive Bias Tower. The Spatiotemporal Adaptive Bias Tower is used to obtain spatiotemporal biases, thereby reducing the difficulty of spatiotemporal differential modeling. Spatiotemporal bias refers to the characteristic differences of different spatiotemporal environmental features in the time and spatial dimensions. That is, the differences between multiple spatiotemporal environmental features used to represent the spatiotemporal environment in which the target user is located in the time and spatial dimensions are represented by spatiotemporal bias.

[0210] To better understand the process of obtaining fused output features based on spatiotemporal semantic features, please refer to... Figure 6 The illustration; Figure 6 Based on spatiotemporal semantic features and the spatiotemporal environmental characteristics "X" c "to obtain fused output features" The diagram illustrates this. Obtaining the fused output feature based on the spatiotemporal semantic features includes: performing processing on the spatiotemporal semantic features using multiple fused fully connected batch normalization layers to obtain the fused output feature. For specific implementation, refer to... Figure 6As illustrated, the number of the above-mentioned fully connected batch normalization layers is 3, including fused FC / BN (1), fused FC / BN (2), and fused FC / BN (3). Of course, the number of the above-mentioned fully connected batch normalization layers can also be set to one or more. This embodiment is for illustration only and is not a practical limitation.

[0211] Specifically, the process of performing multiple fusion fully connected batch normalization layers on the spatiotemporal semantic features to obtain the fusion output features includes: inputting the spatiotemporal semantic features into a first fusion fully connected batch normalization layer to obtain a first fusion intermediate feature; in specific implementation, refer to... Figure 6 The illustration shows the spatiotemporal semantic features. As The input is fed into the fused FC / BN(1) layer to obtain the first fused intermediate features. Specifically, the first fusion intermediate feature is obtained. In addition to the above process In addition, other parameters are required; please refer to the detailed discussion in the following steps. The first fused intermediate features are input into the second fused fully connected batch normalization layer to obtain the second fused intermediate features; in specific implementation, refer to... Figure 6 The illustration shows the first fusion intermediate feature. The input is fed into the second fused FC / BN(2) layer to obtain the second fused intermediate features. With obtaining The process is similar and also requires other parameters. The second fused intermediate feature is input into the third fused fully connected batch normalization layer, and so on, until the last fused intermediate feature output by the penultimate fused fully connected batch normalization layer is obtained; the last fused intermediate feature is then input into the last fused fully connected batch normalization layer to obtain the fused output feature. For specific implementation, refer to... Figure 6 As illustrated, in this embodiment of the application, the fused FC / BN(3) is used as the last fused fully connected batch normalization layer, by fusing the second fused intermediate features. The input is fed into the third fused FC / BN(3) layer to obtain the fused output features. With obtaining The process is similar and also requires other parameters.

[0212] The process of performing multiple fused fully connected batch normalization layers on the spatiotemporal semantic features to obtain the fused output features includes: for any fused fully connected batch normalization layer, obtaining the output result of the arbitrary fused fully connected batch normalization layer based on the output result of the previous fused fully connected batch normalization layer, the spatiotemporal adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer, and the batch normalization adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer. In specific implementation, in the first fused fully connected batch normalization layer "fused FC / BN(1) layer", the spatiotemporal semantic features are processed by multiple fused fully connected batch normalization layers to obtain the fused output features. As The input is fed into the fused FC / BN(1) layer to obtain the output of the "fused FC / BN(1) layer". Will The input data of the second fused fully connected batch normalization layer, "fused FC / BN(2) layer", is used to obtain the output results of the "fused FC / BN(2) layer". The process continues based on the number of fused FC / BN layers. Furthermore, the process of obtaining the output of any fused fully connected batch normalization layer also involves the spatiotemporal adjustment parameters corresponding to any fused fully connected batch normalization layer and the batch normalization adjustment parameters corresponding to any fused fully connected batch normalization layer. In specific implementation, the spatiotemporal adjustment parameters corresponding to any fused fully connected batch normalization layer are represented by the symbol... Indicates; the batch normalization adjustment parameter corresponding to any fully connected batch normalization layer is represented by the symbol. express.

[0213] In this embodiment, based on the spatiotemporal environmental characteristics, the spatiotemporal adjustment parameters corresponding to any one of the fused fully connected batch normalization layers and the batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers are obtained. Specifically, refer to... Figure 6 The above-mentioned fused fully connected batch normalization layer is represented by the symbol "FCNbias"; the spatiotemporal adjustment parameters are respectively represented by the symbols... The batch normalization adjustment parameters are represented by symbols. The superscript "m" is used to distinguish the specific number of layers in the fused fully connected batch normalization layer. For example, "m=1" represents "FCNbias(1)" and "m=2" represents "FCNbias(2)". In this embodiment, the fused fully connected batch normalization layer "FCNbias" has 3 layers, including "FCNbias(1)", "FCNbias(2)" and "FCNbias(3)". It should be understood that in the spatiotemporal adaptive bias tower of this application embodiment, the number of layers of the fused fully connected batch normalization layer "fused FC / BN" and the fused fully connected batch normalization layer "FCNbias" maintains a consistent correspondence. For example, "FCNbias(1)" corresponds to "fused FC / BN(1) layer" and "FCNbias(2)" corresponds to "fused FC / BN(2) layer".

[0214] The step of obtaining the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics includes: performing fully connected layer processing on the spatiotemporal environment characteristics corresponding to any one of the fused fully connected batch normalization layers to obtain the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers; and using the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers as the independent variable of the activation function to obtain the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers.

[0215] In practical implementation, the spatiotemporal adjustment parameters include: spatiotemporal adjustment parameter weights. and spatiotemporal adjustment parameter weight bias Different calculation formulas are used for the operation. Among them, the weights of the spatiotemporal adjustment parameters are obtained. The calculation formula is as follows:

[0216]

[0217] In formulas 1-7, "δ" represents the weights of the spatiotemporal adjustment parameters; "δ" represents the activation function (sigmoid); "h" represents the weights of the spatiotemporal adjustment parameters. c "In the formula, it represents the spatiotemporal environmental characteristics, namely, the "X" mentioned above." c ”; and This represents the adjustment parameters of the FCNbias layer; "m" in the formula indicates that the fused fully connected batch normalized layer FCNbias is at the m-th layer. For ease of understanding, an example is provided. The spatiotemporal environmental feature "X" is... c"After inputting into FCNbias(1) and performing calculations using formulas 1-7, we can obtain..." Furthermore, the adjustment parameters used in this formula These are pre-set parameters; similarly, the spatiotemporal environmental characteristics "X" are... c "After inputting into FCNbias(2) and performing calculations using formulas 1-7, we can obtain..." And the calling parameters used in this formula These are also pre-set parameters.

[0218] It is important to understand that in formulas 1-7, and As the preset parameters of the FCNbias(m) layer (which are also the preset parameters of the BASM model), the parameters can be adjusted during the model training process.

[0219] Among them, the weight bias of the spatiotemporal adjustment parameter is obtained. The calculation formula is as follows:

[0220]

[0221] In formulas 1-8, Indicates the weight bias of the spatiotemporal adjustment parameter; "δ" represents the activation function (sigmoid); "h" c "In the formula, it represents the spatiotemporal environmental characteristics, namely, the "X" mentioned above." c ”; This also represents the adjustment parameters of the FCNbias layer. For ease of understanding, an example will be provided. The spatiotemporal environmental feature "X" will be used... c "After inputting into FCNbias(1) and performing calculations using formulas 1-8, we can obtain..." And the calculation parameters used in this formula are Similarly, the spatiotemporal environmental characteristics "X" c "After inputting into FCNbias(2) and performing calculations using formulas 1-8, we can obtain..." And the calculation parameters used in this formula are Similarly, in formulas 1-8, and Similarly, the default parameters for the FCNbias(m) layer can also be adjusted.

[0222] The weights of the spatiotemporal adjustment parameters are calculated using the steps described above. and spatiotemporal adjustment parameter weight bias The following describes how to obtain batch normalization adjustment parameters. The process is as follows: The specific formula used is:

[0223]

[0224] In formulas 1-9, the corresponding Figure 6 The illustration in the image shows the spatiotemporal semantic features. As The following formula is based on the weights of the spatiotemporal adjustment parameters. and spatiotemporal adjustment parameter weight bias Calculate and obtain fusion intermediate features The formula:

[0225]

[0226] In formula 1-10, The fusion intermediate features of the FC / BN layer are represented; the superscript "m" is used to indicate the specific layer corresponding to the FC / BN layer, such as "m=1" for the FC / BN(1) layer and "m=2" for the FC / BN(2) layer; "δ" represents the activation function (sigmoid); Indicates the weights of the spatiotemporal adjustment parameters; "⊙" represents the Hadamard product; This represents the output of the previous fully connected batch normalization layer. This indicates the weight bias of the spatiotemporal adjustment parameters; This represents the adjustment parameters of the FCNbias layer, which are preset parameters; in the formula, Where L is the number of tower layers. Further, obtain the batch normalized adjustment parameters. The following calculation formula is used:

[0227]

[0228] In formula 1-11, "δ" represents the first-type batch normalization adjustment parameter of the m-th layer of FCNbias; "δ" represents the activation function (sigmoid); "h" represents the first-type batch normalization adjustment parameter of the m-th layer of FCNbias. c "In the formula, it represents the spatiotemporal environmental characteristics, namely, the "X" mentioned above." c ”; This represents the adjustment parameters of the FCNbias layer.

[0229]

[0230] In formula 1-12, "δ" represents the second-type batch normalization adjustment parameter of the m-th layer of FCNbias; "δ" represents the activation function (sigmoid); "h" represents the second-type batch normalization adjustment parameter of the m-th layer of FCNbias. c "In the formula, it represents the spatiotemporal environmental characteristics, namely, the "X" mentioned above." c ”; This represents the adjustment parameters of the FCNbias layer. The specific calculations are as follows: The following formula is used:

[0231]

[0232] In formula 1-13, Indicates the fused output features; This represents the first type of batch normalization adjustment parameter of the m-th layer of FCNbias; This represents the second type of batch normalization adjustment parameter of the m-th layer of FCNbia; "σ" represents the fusion of intermediate features; "σ" represents the variance; "∈" represents the activity parameter; "μ" represents the mean of the batch vector samples; "γ" represents the mean of the batch vector samples. (m) “β” (m) "All of these represent the original learnable BN parameters."

[0233] In specific implementation, refer to Figure 6 The illustration will be based on the spatiotemporal environmental characteristics of "X". c "The result calculated in the input FCNbias(1) layer" Spatiotemporal semantic features As Input into the fused FC / BN(1), and simultaneously, obtain Input into the fused FC / BN(1) and calculate to obtain The spatiotemporal environmental characteristics "X" c "The result calculated in the input FCNbias(2) layer" Will get and Input into the fused FC / BN(2), calculate to obtain This process continues until the fused output features are obtained. In this embodiment, specifically...

[0234] In this embodiment, obtaining the predicted data based on the fused output features includes: performing fully connected layer processing on the fused output features to obtain a fused output fully connected processing result; and using the fused output fully connected processing result as the independent variable of the activation function to obtain the predicted data.

[0235] In practice, the output data of the spatiotemporal adaptive deviation tower will be fused to output features. The data is input into the sigmoid activation function and calculated using a formula to obtain predicted data on the target user's execution of a specified operation on a candidate target object. The calculation formula for obtaining the predicted data is as follows:

[0236]

[0237] In formula 1-14, This represents the predicted data for the target user to perform a specified operation on the candidate target object; "δ" represents the activation function. This represents the fused output characteristics, i.e., the output of the spatiotemporal adaptive bias tower. W o " indicates preset weight parameters; "b o " indicates the preset bias parameter. It is important to understand that the above W..." o “b” o "Specific parameter values ​​can be adjusted during the BASM model training process. Through the above steps, predictive data on the target user's performance of a specified operation on the target candidate object can be obtained. For example, predictive order data on user A's ordering operation for target candidate product 1 in a food delivery shopping application; or, at least one of the following data for the target user's click-through rate, add-to-cart rate, purchase rate, and repurchase rate for the candidate target object."

[0238] The second method involves obtaining spatiotemporal semantic features based on the spatiotemporal environment features, target user features, and candidate target object features, and then obtaining the predicted data based on these features. It should be understood that this method refers to the process by which the Spatiotemporal Semantic Transform Layer of the User Operation Prediction Model (BASM model) obtains spatiotemporal semantic features based on the spatiotemporal environment features, target user features, and candidate target object features, and then obtains the predicted data. The difference from the first method is that in the user operation prediction model of this application, the spatiotemporal semantic transformation layer can also perform data processing independently to obtain the predicted data. The step of obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; and obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted candidate target object features. The step of obtaining the prediction data based on the spatiotemporal semantic features includes: obtaining fused output features based on the spatiotemporal semantic features; and obtaining the prediction data based on the fused output features. The processing procedure of the second method described above can be referred to the description in the first method, and will not be repeated here.

[0239] The third method involves obtaining fused output features with spatiotemporal bias characteristics based on the spatiotemporal environment features, the target user features, and the candidate target object features, and then obtaining the predicted data based on these fused output features. It should be understood that this method refers to the process by which the Spatiotemporal Adaptive Bias Tower of the User Operation Prediction Model (BASM model) obtains fused output features based on spatiotemporal environment features, target user features, and candidate target object features, and then obtains the predicted data. The difference between this method and the first and second methods is that in the user operation prediction model of this application, the spatiotemporal adaptive bias tower can also perform data processing independently to obtain the predicted data. The step of obtaining the fusion output feature with spatiotemporal deviation characteristics based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted candidate target object features; and obtaining the fusion output feature based on the spatiotemporal semantic features. The processing procedure of the third method described above can also refer to the description in the first method, and will not be repeated here.

[0240] Through the above step S202, prediction data of the target user performing a specified operation on the candidate target object is obtained; next, the candidate target object needs to be judged based on the prediction data.

[0241] S203, based on the predicted data, obtain a judgment result on whether to use the candidate target object as the recommended object for the target user.

[0242] This step is used to obtain the judgment result of the candidate target object as the recommended object for the target user.

[0243] The determination of whether to recommend the candidate target object to the target user includes: comparing the predicted data with a preset condition threshold to obtain a comparison result of the predicted data; and obtaining the determination result based on the comparison result of the predicted data. For example, by comparing the predicted click-through rate (CTR) data of the candidate target object estimated by the BASM model with a preset CTR condition threshold. If the predicted CTR data of the candidate target object is greater than or equal to the CTR condition threshold, a determination result is obtained that the candidate target object is recommended to the target user; if the predicted CTR data of the candidate target object is less than the CTR condition threshold, a determination result is obtained that the candidate target object is not recommended to the target user.

[0244] Furthermore, in this embodiment, the User Action Prediction Model (BASM model) can also be used to detect the actual result data of the target user performing a specified operation on the candidate target object; illustratively, the BASM model detects that the actual data result of target user A performing an order operation on candidate target takeaway product 1 in a shopping application is 0, and the actual data result of performing a click operation on candidate target takeaway product 2 is 100%. Of course, the actual result data can be presented in various ways in the model, and can also be in a binarized form; this is only for illustration and not a practical limitation.

[0245] If the deviation between the predicted data and the actual result data does not exceed a predetermined deviation threshold, then the target user features, the candidate target object features, the spatiotemporal environment features, and the predicted data are used as positive samples to update the user operation prediction model. Otherwise, the target user features, the candidate target object features, the spatiotemporal environment features, and the predicted data are used as negative samples to update the user operation prediction model. It should be understood that the deviation between the predicted data and the actual result data not exceeding the predetermined deviation threshold indicates that the complete set of data formed by the target user features, candidate target object features, spatiotemporal environment features, and predicted data has a high confidence level in the BASM model, and therefore, the entire complete set of data can be used as a positive sample. Conversely, if the deviation between the predicted data and the actual result data exceeds the predetermined deviation threshold, then the input data, intermediate data processed by the model, and model output data (i.e., the predicted data) all have low confidence levels, and the entire complete set of data can be used as a negative sample for the BASM model. These positive and negative samples can be used for training the user operation prediction model.

[0246] In practice, the formula for calculating the deviation between the predicted data and the actual result data (binary cross-entropy loss) is as follows:

[0247]

[0248] In formula 1-15, "y i "Indicates actual result data; Represents predicted data; by using the actual result data "y" i "and forecast data" Logarithmic function operations are performed to obtain the deviation data between the two.

[0249] In this embodiment, the method further includes: obtaining combined features and / or the original historical behavior sequence of the target user, wherein the combined features are features obtained by combining at least two features from the target user, at least one feature from the candidate target object, and the spatiotemporal environment features; in this embodiment, the combined features and the original historical behavior sequence of the target user can also be directly used as input data for the BASM model. Here, the combined features are explained in detail. A combined feature is a feature that combines target user features, candidate target object features, and spatiotemporal environment features to describe a single feature. For example, the click-through rate of a user for a product within their geographical activity range over the past three days can be combined to form a combined feature. For ease of understanding, refer to... Figure 3 The illustration shows that the combined features can be represented as "X". j＝4 It is important to understand here that in step S202, the target user's original historical behavior sequence "x" is input into the Spatiotemporal Semantic Transform Layer. j＝3 "Directly applied to this conversion layer. Therefore, in step S201 of this application embodiment, the original historical behavior sequence of the target user "x" can also be obtained." j＝3 ".

[0250] Further, the step of inputting the target user features, the candidate target object features, and the spatiotemporal environment features into the user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object includes: inputting the combined features and / or the target user's original historical behavior sequence, the target user features, the candidate target object features, and the spatiotemporal environment features into the user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object. It should be understood that the aforementioned target user feature "X" j＝1 "Candidate target object feature "X" j＝2 "and spatiotemporal environmental characteristics" X c "In user operation prediction models, as the basic features for obtaining prediction data, after obtaining the above basic features, the combined feature "X" can also be used." j＝4 "and / or the target user's original historical behavior sequence" X j＝3 "Directly used as input data for the model. For example, simultaneously using the target user feature 'X'..." j＝1 "Candidate target object feature "X" j＝2 "Spatiotemporal environmental characteristics" X c "X", the original historical behavior sequence of the target user j＝3 ", Combination feature "X" j＝4"Input into the model. For example, the target user feature "X" j＝1 "Candidate target object feature "X" j＝2 "Spatiotemporal environmental characteristics" X c "X", the original historical behavior sequence of the target user j＝3 "Simultaneously input into the model; for example, the target user feature "X" j＝1 "Candidate target object feature "X" j＝2 "Spatiotemporal environmental characteristics" X c ", Combination feature "X" j＝4 "The input is fed into the BASM model. Of course, the above description is only an illustrative example, and other feature vectors related to object recommendation can also be used as input to the model."

[0251] If the original historical behavior sequence of the target user is obtained, then inputting the combined features and / or the original historical behavior sequence of the target user, the target user features, the candidate target object features, and the spatiotemporal environment features into the user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object includes: obtaining target user interest features of the target user for the candidate target object features based on the original historical behavior sequence of the target user and the candidate target object features; wherein, obtaining target user interest features of the target user for the candidate target object features based on the original historical behavior sequence of the target user and the candidate target object features includes: determining the historical object on which the target user performed the specified historical operation based on the original historical behavior sequence of the target user; and obtaining the target user interest features based on the similarity between the candidate target object and the historical object.

[0252] Here, the original historical behavior sequence of the target user is explained. The original historical behavior sequence of the target user (User Behavior Sequence) refers to the sequence of behavioral objects that the target user has previously performed, arranged chronologically according to the order in which the actions occurred. Illustratively, these behaviors may include: clicking, adding to favorites, adding to cart, sharing, placing an order, repeat purchase, deleting records, etc.; correspondingly, behavioral objects may include products, videos, documents, data tables, etc. It is important to understand that all of the above behavioral objects are exposed in the application on the user's client side, but objects exposed in the application are not necessarily behavioral objects. The behavioral objects mentioned above are objects that the user has interacted with within the application. That is, the behavioral objects of the target user refer to objects in the application that the target user has interacted with. For example, for some objects exposed to the target user, although they are shown to the target user, if the target user does not click on them, purchase them, or browse them, but instead directly flips through the page, this type of object cannot be considered a behavioral object. In this embodiment, the above application can be various types of applications such as shopping, office, games, and video playback applications. For ease of understanding, this application embodiment uses food delivery applications as examples. In the original historical behavior sequence of the target user, the user's behavior objects and behaviors exist in the form of key-value pairs; for example, the "takeout product A" ordered by target user A and the "purchase order" behavior are formed into a key-value pair. By sorting the original historical behavior sequence in timeline, the historical objects in which the target user performed the specified historical operation can be determined; that is, the historical behaviors and historical objects existing in the form of key-value pairs can be obtained.

[0253] The target user interest features are used to represent the target user's interest in candidate target objects. For example, by analyzing the target user's original historical behavior sequence, we can know which products the user has purchased in the past. By calculating the similarity between the features of candidate target products and historical objects (i.e., the features of products the user has purchased), we can determine the user's interest in the candidate target product. The specific process of obtaining the target user interest features is as follows: by calculating the similarity or correlation between the features of the target user's historical objects and the features of candidate target objects, we obtain the target user's interest features for the features of candidate target objects based on the similarity or correlation. For example, we can use the Euclidean distance formula to calculate the similarity or correlation, measuring the degree of difference or correlation between two feature vectors by measuring the straight-line distance between two points in Euclidean space. Alternatively, we can calculate the correlation between the two using formulas such as the Pearson correlation coefficient or the maximum information coefficient. Or, we can calculate the similarity between the two using a self-attention mechanism. The self-attention mechanism (Scaled Dot-Product Attention) can calculate a similarity matrix based on two input data, thereby obtaining relevant information between the inputs. (Refer to...) Figure 3 The illustration shows that the candidate target object feature "X" will be used. j＝2 "X" is the original historical behavior sequence of the target user. j＝3 The obtained target user interest features are represented by the symbol "X". j＝5 "express.

[0254] Based on the combined features, the target user's original historical behavior sequence, the target user features, the candidate target object features, the target user's interest features, and the spatiotemporal environment features, prediction data is obtained regarding the target user performing a specified operation on the candidate target object. In specific implementation, refer to... Figure 3 The illustration shows the combination of features "X". j＝4 "X", the original historical behavior sequence of the target user j＝3 "Target user characteristics" X j＝1 "Candidate target object feature "X" j＝2 "Target user interest characteristics" X j＝5 "Spatiotemporal environmental characteristics" X c All of these are used as input data for the BASM model to obtain corresponding prediction data. It is important to understand that the more diverse the types of input data and the larger the amount of data, the better the model's fitting performance.

[0255] The step of obtaining the target user's interest features based on the target user's original historical behavior sequence and the candidate target object features includes: processing the target user's original historical behavior sequence and the candidate target object features using a multi-head target attention processing layer to obtain the target user's interest features. It should be understood that the above-mentioned acquisition of target user interest features "X"... j＝5 The process of "" is a separate computational process of the multi-head target attention mechanism and does not affect the combined feature "X". j＝4 "X", the original historical behavior sequence of the target user j＝3 "Target user characteristics" X j＝1 "Candidate target object feature "X" j＝2 "Spatiotemporal environmental characteristics" X c "Direct use in the BASM model."

[0256] In this embodiment, the method is applied to the server side. The method further includes: obtaining a page content request message from a client requesting page content; returning page content to the client in response to the page content request message; wherein, if the judgment result indicates that the candidate target object is a recommended object for the target user, then the page content includes the candidate target object; or, obtaining an object query message from a client; obtaining a candidate target object in response to the object query message; and returning an object query result to the client; wherein, if the judgment result indicates that the candidate target object is a recommended object for the target user, then the object query result includes the candidate target object. It should be understood that when a target user on the client side is browsing a page, they can perform a page content search operation by opening the application's display page. For example, if a target user opens a food delivery application and searches for the keyword "hamburger," the client can send a "hamburger" object query message based on this keyword. After receiving the "hamburger" object query message, the server on the server side processes it using the BASM model to obtain a recommended object and sends the recommended object to the client, exposing it to the target user.

[0257] The method is applied to a client and further includes: displaying page content in response to detecting a triggering operation for displaying page content, wherein if the determination result indicates that the candidate target object is a recommended object for the target user, then the page content includes the candidate target object; or, displaying object query results in response to detecting a triggering operation for querying an object, wherein if the determination result indicates that the candidate target object is a recommended object for the target user, then the object query results include the candidate target object. It should be understood that the above triggering operation is not necessarily an operation related to the target user's intention to search; it can be a triggering operation by the target user under random and occasional circumstances; the above-mentioned triggering operation by the target user can also obtain recommended objects on the client side.

[0258] The object recommendation method provided in the first embodiment above can adaptively adapt to the spatiotemporal data distribution through user operation prediction models, thereby improving the model's data fitting ability. Specifically, the spatiotemporal attention embedding layer performs weight adaptation in feature embedding to achieve dynamic perception of spatiotemporal background. The spatiotemporal semantic transformation layer explicitly converts the original semantic connection input into spatiotemporal semantics, thereby enhancing semantic representation under different spatiotemporal backgrounds. The spatiotemporal adaptive bias tower captures different spatiotemporal biases, reducing the difficulty of modeling spatiotemporal differences. The candidate target objects obtained through this process have good adaptability to the target user, and the recommended objects have high accuracy and precision.

[0259] Second Embodiment

[0260] The following combination Figure 7 This document introduces an embodiment of the user behavior prediction method provided in this application, applied to a user operation prediction model. Please refer to... Figure 7 This is a flowchart illustrating the second embodiment of this application. The specific implementation process of this embodiment can be found in the description of the first embodiment, and the method includes the following steps:

[0261] Step S701: Obtain the target user characteristics, the target operation characteristics of the specified operation, and the spatiotemporal environment characteristics corresponding to the target user. The target operation characteristics of the specified operation can be understood as the object characteristics of the target user's operation. For example, if the target user clicks on a product, the target operation characteristics are the target product characteristics, and the specified operation is clicking that product. Alternatively, the target operation characteristics can also refer to the characteristics of the action itself. For example, the action of kicking a soccer ball involves one foot first moving backward, then forcefully extending forward, and the soccer ball moving after touching it.

[0262] Step S702: Input the target user features, the target operation features, and the spatiotemporal environment features into the user operation prediction model to obtain prediction data of the target user performing a specified operation. Specifically, the user operation prediction model performs at least one of the following processes: assigning weights to the target user features and the target operation features based on the spatiotemporal environment features, and obtaining the prediction data based on the target user feature weights and the target operation feature weights; obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the target operation features, and obtaining the prediction data based on the spatiotemporal semantic features; obtaining a fused output feature with spatiotemporal deviation features based on the spatiotemporal environment features, the target user features, and the target operation features, and obtaining the prediction data based on the fused output feature.

[0263] Optionally, assigning weights to the target operation features based on the spatiotemporal environment features includes: performing concatenation processing on the spatiotemporal environment features and the target operation features to obtain spatiotemporal target operation concatenation features; performing fully connected layer processing on the spatiotemporal target operation to obtain a fully connected processing result for the spatiotemporal target operation; and using the fully connected processing result for the spatiotemporal target operation as the independent variable of the activation function to obtain the target operation feature weights.

[0264] Optionally, the spatiotemporal environment features are spatiotemporal environment feature vectors, and the target operation features are target operation feature vectors; the step of concatenating the spatiotemporal environment features and the target operation features to obtain spatiotemporal target operation connection features includes: concatenating the spatiotemporal environment feature vector and the target operation feature vector to obtain a spatiotemporal target operation connection feature vector; the step of performing fully connected layer processing on the spatiotemporal target operation to obtain a spatiotemporal target operation fully connected processing result includes: obtaining a spatiotemporal target operation fully connected processing result vector based on the spatiotemporal target operation connection feature vector, weight allocation of fully connected layer weight parameters, and weight allocation of fully connected layer bias parameters.

[0265] Optionally, obtaining the predicted data based on the target user feature weights and target operation feature weights includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted target operation features with spatiotemporal weights based on the candidate target object feature weights and the target operation features; obtaining fused output features based on the spatiotemporally weighted target user features and the spatiotemporally weighted target operation features; and obtaining the predicted data based on the fused output features.

[0266] Optionally, the target user feature is a target user feature vector, and the target operation feature is a target operation feature vector; obtaining a spatiotemporally weighted target user feature with spatiotemporal weights based on the target user feature weights and the target user feature includes: performing a multiplication operation on the target user feature vector and the target user feature weights to obtain a spatiotemporally weighted target user feature vector; obtaining a spatiotemporally weighted target operation feature with spatiotemporal weights based on the target operation feature weights and the target operation feature includes: performing a multiplication operation on the target operation feature vector and the target operation feature weights to obtain a spatiotemporally weighted target operation feature vector.

[0267] Optionally, obtaining the fused output features based on the spatiotemporal weighted target user features and the spatiotemporal weighted target operation features includes: obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporal weighted target user features, and the spatiotemporal weighted target operation features; and obtaining the fused output features based on the spatiotemporal semantic features.

[0268] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporal weight target user features, and the spatiotemporal weight target operation features includes: obtaining original semantic features based on the spatiotemporal weight target user features and the spatiotemporal weight target operation features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0269] Optionally, obtaining the original semantic features based on the spatiotemporal weight target user features and the spatiotemporal weight target operation features includes: performing concatenation processing on the spatiotemporal weight target user features and the spatiotemporal weight target operation features to obtain the original semantic features.

[0270] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the target operation features includes: obtaining original semantic features based on the target user features and the target operation features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0271] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the target operation features includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted target operation features with spatiotemporal weights based on the target operation feature weights and the target operation features; and obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted target operation features.

[0272] Optionally, obtaining the fusion output feature with spatiotemporal deviation features based on the spatiotemporal environment features, the target user features, and the target operation features includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted target operation features with spatiotemporal weights based on the target operation feature weights and the target operation features; obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted target operation features; and obtaining the fusion output feature based on the spatiotemporal semantic features.

[0273] Optionally, it further includes: detecting the actual result data of the target user performing the specified operation; if the deviation data between the predicted data and the actual result data does not exceed a predetermined deviation threshold, then the target user features, the target operation features, the spatiotemporal environment features, and the predicted data are used as positive samples to update the user operation prediction model; otherwise, the target user features, the target operation features, the spatiotemporal environment features, and the predicted data are used as negative samples to update the user operation prediction model.

[0274] Optionally, it further includes: obtaining combined features and / or the original historical behavior sequence of the target user, wherein the combined features are features obtained by combining at least two features from at least one feature of the target user, at least one target operation feature, and the spatiotemporal environment feature; the step of inputting the target user feature, the target operation feature, and the spatiotemporal environment feature into a user operation prediction model to obtain prediction data of the target user performing a specified operation includes: inputting the combined features and / or the original historical behavior sequence of the target user, the target user feature, the target operation, and the spatiotemporal environment feature into a user operation prediction model to obtain prediction data of the target user performing a specified operation.

[0275] Optionally, if the original historical behavior sequence of the target user is obtained, then the step of inputting the combined features and / or the original historical behavior sequence of the target user, the target user features, the target operation features, and the spatiotemporal environment features into the user operation prediction model to obtain prediction data of the target user performing a specified operation includes: obtaining the target user interest features of the target user for the target operation features based on the original historical behavior sequence of the target user and the target operation features; and obtaining the prediction data of the target user performing a specified operation based on the combined features, the original historical behavior sequence of the target user, the target user features, the target operation features, the target user interest features, and the spatiotemporal environment features.

[0276] Optionally, obtaining the target user interest features of the target user in relation to the target operation features based on the target user's original historical behavior sequence and the target operation features includes: processing the target user's original historical behavior sequence and the target operation features using a multi-head target attention processing layer to obtain the target user interest features.

[0277] Optionally, obtaining the target user interest features of the target user for the target operation features based on the target user's original historical behavior sequence and the target operation features includes: determining the historical operations performed by the target user based on the target user's original historical behavior sequence; and obtaining the target user interest features based on the similarity between the target operation features and the historical operation features of the historical operations performed by the target user.

[0278] Third Embodiment

[0279] The following combination Figure 8 This paper introduces an embodiment of a model output result acquisition method provided in this application, applied to a spatiotemporal attention embedding layer. Please refer to... Figure 8 This figure is a flowchart illustrating the third embodiment of this application. For a detailed description of the implementation process of this embodiment, please refer to the description in the first embodiment.

[0280] Step S801: Obtain the spatiotemporal environmental features corresponding to the target user and multiple basic input features of the user operation prediction model. These multiple basic input features include target user features and candidate target object features.

[0281] Step S802: Assign weights to the multiple basic input features according to the spatiotemporal environmental features, and obtain the spatiotemporal weights corresponding to the multiple basic input features respectively.

[0282] Step S803: Based on the multiple basic input features and the spatiotemporal weights corresponding to the multiple basic input features, obtain multiple spatiotemporal weighted basic input features.

[0283] Step S804: Based on the multiple spatiotemporal weighted basic input features, obtain the output result of the user operation prediction model.

[0284] Optionally, the step of assigning weights to the plurality of basic input features based on the spatiotemporal environment features to obtain the spatiotemporal weights corresponding to the plurality of basic input features includes: concatenating the spatiotemporal environment features with any one of the basic input features to obtain spatiotemporal basic input concatenation features; performing fully connected layer processing on the spatiotemporal basic input concatenation features to obtain a spatiotemporal basic input fully connected processing result; and using the spatiotemporal basic input fully connected processing result as the independent variable of the activation function to obtain the basic input feature weights of any one of the basic input features.

[0285] Optionally, the spatiotemporal environment feature is a spatiotemporal environment feature vector, and the basic input feature is a basic input feature vector. Connecting the spatiotemporal environment feature with any basic input feature to obtain a spatiotemporal basic input connection feature includes: connecting the spatiotemporal environment feature vector with the basic input feature vector of any basic input feature to obtain a spatiotemporal basic input connection feature vector. Performing fully connected layer processing on the spatiotemporal basic input connection feature to obtain a spatiotemporal basic input fully connected processing result includes: obtaining a spatiotemporal basic input fully connected processing result vector based on the spatiotemporal basic input connection feature vector, weight allocation of the fully connected layer weight parameters, and weight allocation of the fully connected layer bias parameters.

[0286] Optionally, the basic input features are basic input feature vectors; obtaining multiple spatiotemporal weighted basic input features based on the multiple basic input features and the spatiotemporal weights corresponding to the multiple basic input features includes: performing a multiplication operation between the basic input feature vector of any basic input feature and the spatiotemporal weight corresponding to the any basic input feature to obtain the spatiotemporal weighted basic input feature vector of the any basic input feature.

[0287] Optionally, obtaining the output result of the user operation prediction model based on the plurality of spatiotemporal weighted basic input features includes: obtaining fused output features based on the plurality of spatiotemporal weighted basic input features; and obtaining the output result of the user operation prediction model based on the fused output features.

[0288] Optionally, obtaining the fused output features based on the plurality of spatiotemporal weighted basic input features includes: obtaining spatiotemporal semantic features based on the spatiotemporal environment features and the plurality of spatiotemporal weighted basic input features; and obtaining the fused output features based on the spatiotemporal semantic features.

[0289] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features and the plurality of spatiotemporal weighted basic input features includes: obtaining original semantic features based on the plurality of spatiotemporal weighted basic input features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0290] Optionally, obtaining the fused output features based on the spatiotemporal semantic features includes: performing processing on the spatiotemporal semantic features using multiple fused fully connected batch normalization layers to obtain the fused output features.

[0291] Optionally, obtaining the output result of the user operation prediction model based on the fused output features includes: performing fully connected layer processing on the fused output features to obtain a fused output fully connected processing result; and using the fused output fully connected processing result as the independent variable of the activation function to obtain the output result of the user operation prediction model.

[0292] Fourth embodiment

[0293] The following combination Figure 9 This paper introduces an embodiment of a method for obtaining model output results, applied to a spatiotemporal semantic transformation layer. Please refer to... Figure 9 This figure is a flowchart illustrating the fourth embodiment of this application. For a detailed description of the implementation process of this embodiment, please refer to the description in the first embodiment.

[0294] Step S901: Obtain the spatiotemporal environmental features corresponding to the target user and multiple basic input features of the user operation prediction model. These multiple basic input features include target user features, candidate target object features, and the target user's historical behavior sequence.

[0295] Step S902: Obtain spatiotemporal semantic features based on the spatiotemporal environmental features and the multiple basic input features.

[0296] Step S903: Based on the spatiotemporal semantic features, obtain the output result of the user operation prediction model.

[0297] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features and the plurality of basic input features includes: obtaining original semantic features based on the spatiotemporal weighted basic input features corresponding to the plurality of basic input features respectively; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0298] Optionally, obtaining the original semantic features based on the spatiotemporal weighted basic input features corresponding to the plurality of basic input features includes: concatenating the spatiotemporal weighted basic input features corresponding to the plurality of basic input features to obtain the original semantic features.

[0299] Optionally, it further includes: obtaining the target user's historical behavior sequence matching the spatiotemporal environment features; obtaining the spatiotemporal semantic fully connected layer weight parameters and the spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features includes: inputting the spatiotemporal environment features and the target user's historical behavior sequence matching the spatiotemporal environment features into the meta-network to obtain the spatiotemporal semantic fully connected layer weight parameters and the spatiotemporal semantic fully connected layer bias parameters.

[0300] Optionally, the step of inputting the spatiotemporal environmental features and the target user's historical behavior sequence matching the spatiotemporal environmental features into the meta-network to obtain the weight parameters and bias parameters of the spatiotemporal semantic fully connected layer includes: performing concatenation processing on the spatiotemporal environmental features and the target user's historical behavior sequence matching the spatiotemporal environmental features to obtain spatiotemporal historical behavior features; and obtaining the weight parameters and bias parameters of the spatiotemporal semantic fully connected layer based on the spatiotemporal historical behavior features, the meta-network weight parameters, and the meta-network bias parameters.

[0301] Optionally, it further includes: obtaining the original historical behavior sequence of the target user; obtaining the target user's historical behavior sequence that matches the spatiotemporal environmental features includes: filtering out the target user's historical behavior sequence that matches the spatiotemporal environmental features from the target user's original historical behavior sequence.

[0302] Optionally, obtaining the output result of the user operation prediction model based on the spatiotemporal semantic features includes: obtaining fused output features based on the spatiotemporal semantic features; and obtaining the output result of the user operation prediction model based on the fused output features.

[0303] Fifth embodiment

[0304] The following combination Figure 10 This paper introduces an embodiment of a model output result acquisition method provided in this application, applied to a spatiotemporal adaptive bias tower. Please refer to... Figure 10 This figure is a flowchart illustrating the fifth embodiment of this application. For a detailed description of the implementation process of this embodiment, please refer to the description in the first embodiment.

[0305] Step S1001: Obtain the spatiotemporal environmental features corresponding to the target user and multiple basic input features of the user operation prediction model. These multiple basic input features include target user features and candidate target object features.

[0306] Step S1002: Based on the spatiotemporal environmental features and the multiple basic input features, obtain the fused output features with spatiotemporal deviation features.

[0307] Step S1003: Obtain the output result of the user operation prediction model based on the fused output features.

[0308] Optionally, obtaining the fused output feature with spatiotemporal deviation features based on the spatiotemporal environmental features and the plurality of basic input features includes: obtaining a spatiotemporal weighted basic input feature corresponding to any one basic input feature and the spatiotemporal weight corresponding to the one basic input feature; obtaining spatiotemporal semantic features based on the spatiotemporal environmental features and the spatiotemporal weighted basic input features corresponding to the plurality of basic input features; and obtaining the fused output feature based on the spatiotemporal semantic features.

[0309] Optionally, obtaining the fused output features based on the spatiotemporal semantic features includes: performing processing on the spatiotemporal semantic features using multiple fused fully connected batch normalization layers to obtain the fused output features.

[0310] Optionally, the step of performing multiple fusion fully connected batch normalization layers on the spatiotemporal semantic features to obtain the fusion output features includes: inputting the spatiotemporal semantic features into a first fusion fully connected batch normalization layer to obtain a first fusion intermediate feature; inputting the first fusion intermediate feature into a second fusion fully connected batch normalization layer to obtain a second fusion intermediate feature; inputting the second fusion intermediate feature into a third fusion fully connected batch normalization layer, and so on, until the last fusion intermediate feature output by the penultimate fusion fully connected batch normalization layer is obtained; and inputting the last fusion intermediate feature into the last fusion fully connected batch normalization layer to obtain the fusion output features.

[0311] Optionally, the step of performing processing on multiple fused fully connected batch normalization layers for the spatiotemporal semantic features to obtain the fused output features includes: for any fused fully connected batch normalization layer, obtaining the output result of the arbitrary fused fully connected batch normalization layer based on the output result of the previous fused fully connected batch normalization layer, the spatiotemporal adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer, and the batch normalization adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer.

[0312] Optionally, it also includes: obtaining the spatiotemporal adjustment parameters corresponding to any one of the fused fully connected batch normalization layers and the batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics.

[0313] Optionally, obtaining the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics includes: performing fully connected layer processing on the spatiotemporal environment characteristics for any one of the fused fully connected batch normalization layers to obtain the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers; and using the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers as the independent variable of the activation function to obtain the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers.

[0314] Optionally, obtaining the output result of the user operation prediction model based on the fused output features includes: performing fully connected layer processing on the fused output features to obtain a fused output fully connected processing result; and using the fused output fully connected processing result as the independent variable of the activation function to obtain the output result of the user operation prediction model.

[0315] Sixth Embodiment

[0316] The following combination Figure 11 This document introduces an embodiment of a model training method provided in this application, applied to a user operation prediction model. Please refer to... Figure 11 This figure is a flowchart illustrating the sixth embodiment of this application. For a detailed description of the implementation process of this embodiment, please refer to the description in the first embodiment.

[0317] Step S1101: Obtain the predicted data of the target user performing a specified operation on the candidate target object and the actual result data of the target user performing the specified operation on the candidate target object.

[0318] Step S1102: Calculate the deviation data between the predicted data and the actual result data.

[0319] If the deviation between the predicted data and the actual result data does not exceed a predetermined deviation threshold, the current user operation prediction model is used as the recommended user operation prediction model. If the deviation between the predicted data and the actual result data is greater than the predetermined deviation threshold, the current user operation prediction model is trained until the deviation data meets the predetermined condition of not exceeding the predetermined deviation threshold.

[0320] Seventh Embodiment

[0321] The following combination Figure 11A This paper introduces another embodiment of the object recommendation method provided in this application. Please refer to... Figure 11AThis figure is a flowchart illustrating the seventh embodiment of this application. For a detailed description of the implementation process of this embodiment, please refer to the description in the first embodiment.

[0322] Step S1111: In response to the fulfillment of the triggering condition for displaying information, obtain the recommended objects to be recommended to the target user;

[0323] Step S1112: Display the recommended object on the display interface for displaying information; wherein the recommended object is determined based on prediction data of the target user performing a specified operation on the candidate target object; the prediction data is obtained by assigning target user feature weights to the target user features of the target user based on the spatiotemporal environment features corresponding to the target user and assigning candidate target object feature weights to the candidate target object features of the candidate target object; or, the prediction data is obtained based on spatiotemporal semantic features obtained based on the spatiotemporal environment features, the target user features and the candidate target object features; or, the prediction data is obtained based on fused output features with spatiotemporal deviation features obtained based on the spatiotemporal environment features, the target user features and the candidate target object features.

[0324] Optionally, the response to meeting the triggering condition for displaying information includes: responding to detecting a triggering operation for displaying page content; or, responding to detecting a triggering operation for querying an object.

[0325] Optionally, target user feature weights are assigned to the target user features of the target user based on the spatiotemporal environmental features corresponding to the target user in the following manner: the spatiotemporal environmental features are concatenated with the target user features to obtain spatiotemporal user concatenation features; the spatiotemporal user concatenation features are processed by a fully connected layer to obtain a spatiotemporal user fully connected processing result; the spatiotemporal user fully connected processing result is used as the independent variable of the activation function to obtain the target user feature weights.

[0326] Optionally, candidate target object feature weights are assigned to the candidate target object features of the candidate target object based on the spatiotemporal environment features corresponding to the target user in the following manner: the spatiotemporal environment features are concatenated with the candidate target object features to obtain spatiotemporal object concatenation features; the spatiotemporal object concatenation features are processed by a fully connected layer to obtain a spatiotemporal object fully connected processing result; the spatiotemporal object fully connected processing result is used as the independent variable of the activation function to obtain the candidate target object feature weights.

[0327] Optionally, the predicted data is obtained according to the target user feature weights and the candidate target object feature weights in the following manner: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; obtaining fused output features based on the spatiotemporally weighted target user features and the spatiotemporally weighted candidate target object features; and obtaining the predicted data based on the fused output features.

[0328] Optionally, the target user feature is a target user feature vector, and the candidate target object feature is a candidate target object feature vector; obtaining a spatiotemporally weighted target user feature with spatiotemporal weights based on the target user feature weights and the target user feature includes: performing a multiplication operation between the target user feature vector and the target user feature weights to obtain a spatiotemporally weighted target user feature vector; obtaining a spatiotemporally weighted candidate target object feature with spatiotemporal weights based on the candidate target object feature weights and the candidate target object feature includes: performing a multiplication operation between the candidate target object feature vector and the candidate target object feature weights to obtain a spatiotemporally weighted candidate target object feature vector.

[0329] Optionally, obtaining the fused output features based on the spatiotemporal weighted target user features and the spatiotemporal weighted candidate target object features includes: obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporal weighted target user features, and the spatiotemporal weighted candidate target object features; and obtaining the fused output features based on the spatiotemporal semantic features.

[0330] Optionally, spatiotemporal semantic features are obtained based on the spatiotemporal environment features, the target user features, and the candidate target object features in the following manner: obtaining original semantic features based on the target user features and the candidate target object features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal environment features; and obtaining the spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters.

[0331] Optionally, obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; and obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted candidate target object features.

[0332] Optionally, the predicted data can be obtained based on the spatiotemporal semantic features in the following manner: obtain fused output features based on the spatiotemporal semantic features; obtain the predicted data based on the fused output features.

[0333] Optionally, the fusion output feature with spatiotemporal deviation features is obtained according to the spatiotemporal environment features, the target user features, and the candidate target object features in the following manner: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features with spatiotemporal weights based on the candidate target object feature weights and the candidate target object features; obtaining the spatiotemporal semantic features based on the spatiotemporal environment features, the spatiotemporally weighted target user features, and the spatiotemporally weighted candidate target object features; and obtaining the fusion output feature based on the spatiotemporal semantic features.

[0334] Optionally, the fused output features are obtained based on the spatiotemporal semantic features in the following manner: for the spatiotemporal semantic features, multiple fused fully connected batch normalization layers are processed to obtain the fused output features.

[0335] Optionally, the step of performing multiple fusion fully connected batch normalization layers on the spatiotemporal semantic features to obtain the fusion output features includes: inputting the spatiotemporal semantic features into a first fusion fully connected batch normalization layer to obtain a first fusion intermediate feature; inputting the first fusion intermediate feature into a second fusion fully connected batch normalization layer to obtain a second fusion intermediate feature; inputting the second fusion intermediate feature into a third fusion fully connected batch normalization layer, and so on, until the last fusion intermediate feature output by the penultimate fusion fully connected batch normalization layer is obtained; and inputting the last fusion intermediate feature into the last fusion fully connected batch normalization layer to obtain the fusion output features.

[0336] Optionally, the step of performing processing on multiple fused fully connected batch normalization layers for the spatiotemporal semantic features to obtain the fused output features includes: for any fused fully connected batch normalization layer, obtaining the output result of the arbitrary fused fully connected batch normalization layer based on the output result of the previous fused fully connected batch normalization layer, the spatiotemporal adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer, and the batch normalization adjustment parameter corresponding to the arbitrary fused fully connected batch normalization layer.

[0337] Optionally, it also includes: obtaining the spatiotemporal adjustment parameters corresponding to any one of the fused fully connected batch normalization layers and the batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics.

[0338] Optionally, obtaining the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers based on the spatiotemporal environment characteristics includes: performing fully connected layer processing on the spatiotemporal environment characteristics for any one of the fused fully connected batch normalization layers to obtain the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers; and using the spatiotemporal environment fully connected processing result corresponding to any one of the fused fully connected batch normalization layers as the independent variable of the activation function to obtain the spatiotemporal adjustment parameters and batch normalization adjustment parameters corresponding to any one of the fused fully connected batch normalization layers.

[0339] Optionally, the predicted data can be obtained based on the fused output features in the following manner: the fused output features are processed by a fully connected layer to obtain a fused output fully connected processing result; the fused output fully connected processing result is used as the independent variable of the activation function to obtain the predicted data.

[0340] Eighth embodiment

[0341] The eighth embodiment of this application provides an apparatus for recommending objects, corresponding to the first embodiment described above, with reference to... Figure 12 The indication, Figure 12For the structural schematic diagram of an object recommendation device corresponding to the eighth embodiment of this application, the device 1200 includes: an acquisition unit 1201, used to acquire target user features of a target user, candidate target object features of candidate target objects, and spatiotemporal environment features corresponding to the target user; a first processing unit 1202, used to input the target user features, the candidate target object features, and the spatiotemporal environment features into a user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object; and a second processing unit 1203, used to determine, based on the prediction data, whether to recommend the candidate target object as a recommended object for the target user. The judgment result; wherein, in the user operation prediction model, at least one of the following processes is performed: assigning weights to the target user features and the candidate target object features according to the spatiotemporal environment features, and obtaining the prediction data according to the target user feature weights and the candidate target object feature weights; obtaining spatiotemporal semantic features according to the spatiotemporal environment features, the target user features, and the candidate target object features, and obtaining the prediction data according to the spatiotemporal semantic features; obtaining fused output features with spatiotemporal deviation features according to the spatiotemporal environment features, the target user features, and the candidate target object features, and obtaining the prediction data according to the fused output features.

[0342] Ninth Embodiment

[0343] The ninth embodiment of this application provides a device for predicting user behavior, corresponding to the aforementioned second embodiment, with reference to... Figure 13 The indication, Figure 13 A schematic diagram of a user behavior prediction device corresponding to the ninth embodiment of this application is provided. The device 1300 includes: an acquisition unit 1301, used to acquire target user features of a target user, target operation features of a specified operation, and spatiotemporal environment features corresponding to the target user; and a processing unit 1302, used to input the target user features, the target operation features, and the spatiotemporal environment features into a user operation prediction model to obtain prediction data of the target user performing the specified operation; wherein, in the user operation prediction model, at least one of the following processes is performed: assigning weights to the target user features and the target operation features according to the spatiotemporal environment features, and obtaining the prediction data according to the target user feature weights and the target operation feature weights; obtaining spatiotemporal semantic features according to the spatiotemporal environment features, the target user features, and the target operation features, and obtaining the prediction data according to the spatiotemporal semantic features; obtaining a fusion output feature with spatiotemporal deviation features according to the spatiotemporal environment features, the target user features, and the target operation features, and obtaining the prediction data according to the fusion output feature.

[0344] Tenth Embodiment

[0345] The tenth embodiment of this application provides a model output result acquisition device, corresponding to the aforementioned third embodiment, and referring to... Figure 14 The indication, Figure 14 This is a schematic diagram of a model output result acquisition device according to the tenth embodiment of this application. The device 1400 includes: an acquisition unit 1401, used to acquire spatiotemporal environmental features corresponding to a target user and multiple basic input features of the user operation prediction model; a first processing unit 1402, used to assign weights to the multiple basic input features according to the spatiotemporal environmental features, and obtain spatiotemporal weights corresponding to the multiple basic input features respectively; a second processing unit 1403, used to acquire multiple spatiotemporal weighted basic input features according to the multiple basic input features and the spatiotemporal weights corresponding to the multiple basic input features respectively; and a result acquisition unit 1404, used to acquire the output result of the user operation prediction model according to the multiple spatiotemporal weighted basic input features.

[0346] Eleventh Embodiment

[0347] The eleventh embodiment of this application also provides a model output result acquisition device, corresponding to the aforementioned fourth embodiment, see below. Figure 15 The indication, Figure 15 This is a schematic diagram of a model output result acquisition device according to the eleventh embodiment of this application. The device 1500 includes: an acquisition unit 1501, used to acquire spatiotemporal environmental features corresponding to a target user and multiple basic input features of the user operation prediction model; a first processing unit 1502, used to acquire spatiotemporal semantic features based on the spatiotemporal environmental features and the multiple basic input features; and a result acquisition unit 1503, used to acquire the output result of the user operation prediction model based on the spatiotemporal semantic features.

[0348] Twelfth Embodiment

[0349] The twelfth embodiment of this application also provides a model output result acquisition device, corresponding to the aforementioned fifth embodiment, see below. Figure 16 The indication, Figure 16 This is a schematic diagram of a model output result acquisition device according to the twelfth embodiment of this application. The device 1600 includes: an acquisition unit 1601, used to acquire spatiotemporal environmental features corresponding to a target user and multiple basic input features of the user operation prediction model; a first processing unit 1602, used to acquire fused output features with spatiotemporal deviation features based on the spatiotemporal environmental features and the multiple basic input features; and a result acquisition unit 1603, used to acquire the output result of the user operation prediction model based on the fused output features.

[0350] Thirteenth Embodiment

[0351] The thirteenth embodiment of this application also provides a model training device applied to a user operation prediction model, corresponding to the aforementioned sixth embodiment, and referring to... Figure 17 The indication, Figure 17 This is a schematic diagram of a model training device according to the thirteenth embodiment of this application. The device 1700 includes: an acquisition unit 1701, used to acquire prediction data of a target user performing a specified operation on a candidate target object and actual result data of the target user performing the specified operation on the candidate target object; and a processing unit 1702, used to calculate and obtain deviation data between the prediction data and the actual result data; if the deviation data between the prediction data and the actual result data does not exceed a predetermined deviation threshold, then the current user operation prediction model is used as the user operation prediction model recommended for the object; if the deviation data between the prediction data and the actual result data is greater than the predetermined deviation threshold, then the current user operation prediction model is trained until the deviation data meets the predetermined condition of not exceeding the predetermined deviation threshold.

[0352] Fourteenth Embodiment

[0353] The fourteenth embodiment of this application also provides an object recommendation device, corresponding to the aforementioned seventh embodiment, see below. Figure 17A The indication, Figure 17A This is a schematic diagram of an object recommendation device according to the fourteenth embodiment of this application. The device 1710 includes: an acquisition unit 1711 configured to acquire recommended objects to a target user in response to a trigger condition for displaying information; and a display unit 1712 configured to display the recommended objects on a display interface for displaying information. The recommended objects are determined based on prediction data of a target user performing a specified operation on a candidate target object. The prediction data is obtained by assigning target user feature weights to the target user's target user features based on the spatiotemporal environment features corresponding to the target user, and assigning candidate target object feature weights to the candidate target object's candidate target object features. Alternatively, the prediction data is obtained based on spatiotemporal semantic features obtained from the spatiotemporal environment features, the target user features, and the candidate target object features. Or, the prediction data is obtained based on fused output features with spatiotemporal deviation features obtained from the spatiotemporal environment features, the target user features, and the candidate target object features.

[0354] Fifteenth Embodiment

[0355] The fifteenth embodiment of this application also provides an electronic device. Since the electronic device embodiment is basically similar to the method embodiment, it is described simply. For details of the relevant technical features, please refer to the corresponding descriptions of the method embodiments provided above. The following description of the electronic device embodiment is merely illustrative. The electronic device embodiment is as follows: Please refer to... Figure 18 To understand this embodiment, Figure 18 This is a schematic diagram of the electronic device provided in this embodiment. Figure 18 As shown, the electronic device provided in this embodiment includes: a processor 1801, a memory 1802, a communication bus 1803, and a communication interface 1804. The processor 1801 is used to execute one or more computer instructions to implement the steps of the above-described method embodiments. The memory 1802 is used to store one or more computer instructions for data processing. The communication bus 1803 is used to connect the processor 1801 and the memory 1802 mounted thereon. The communication interface 1804 is used to provide a connection interface for the processor 1801 and the memory 1802.

[0356] In the above embodiments, various methods and corresponding apparatuses are provided. Furthermore, this application also provides a computer-readable storage medium for implementing the above methods. The computer-readable storage medium embodiments provided in this application are described in a relatively simple manner; relevant parts can be found in the corresponding descriptions of the above method embodiments. The embodiments described below are merely illustrative.

[0357] The computer-readable storage medium provided in this embodiment stores computer instructions, which, when executed by a processor, implement the steps shown in the above method embodiment, and will not be repeated here.

[0358] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0359] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0360] 1. Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include non-transitory computer-readable media, such as modulated data signals and carrier waves.

[0361] 2. Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0362] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope defined in the claims of the present invention.

[0363] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

Claims

1. An object recommendation method, characterized in that, include: Obtain the target user characteristics of the target user, the candidate target object characteristics of the candidate target object, and the spatiotemporal environment characteristics corresponding to the target user; The target user features, the candidate target object features, and the spatiotemporal environment features are input into the user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object; Based on the predicted data, a judgment result is obtained on whether to use the candidate target object as the recommended object for the target user; In the user operation prediction model, weights are assigned to the target user features and the candidate target object features based on the spatiotemporal environment features, and the prediction data is obtained based on the target user feature weights and the candidate target object feature weights. The step of obtaining the prediction data based on the target user feature weights and candidate target object feature weights includes: obtaining spatiotemporal weighted target user features based on the target user feature weights and the target user features; obtaining spatiotemporal weighted candidate target object features based on the candidate target object feature weights and the candidate target object features; obtaining original semantic features based on the spatiotemporal weighted target user features and the spatiotemporal weighted candidate target object features; filtering out target user historical behavior sequences that match the spatiotemporal environment features from the target user's original historical behavior sequences; concatenating the spatiotemporal environment features and the target user historical behavior sequences that match the spatiotemporal environment features to obtain spatiotemporal historical behavior features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters; performing spatiotemporal semantic transformation on the original semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters to obtain spatiotemporal semantic features; obtaining fused output features based on the spatiotemporal semantic features; and obtaining the prediction data based on the fused output features.

2. The method according to claim 1, characterized in that, The step of assigning weights to the target user features and the candidate target object features based on the spatiotemporal environmental features includes: The spatiotemporal environment features are concatenated with the target user features to obtain spatiotemporal user connection features; Fully connected layer processing is performed on the spatiotemporal user connection characteristics to obtain the spatiotemporal user fully connected processing result; The spatiotemporal user fully connected processing result is used as the independent variable of the activation function to obtain the target user feature weights.

3. The method according to claim 1, characterized in that, The step of assigning weights to the target user features and the candidate target object features based on the spatiotemporal environmental features includes: The spatiotemporal environment features are connected with the candidate target object features to obtain spatiotemporal object connection features; Fully connected layer processing is performed on the spatiotemporal object connectivity features to obtain the spatiotemporal object fully connected processing result; The result of the fully connected processing of the spatiotemporal object is used as the independent variable of the activation function to obtain the feature weights of the candidate target object.

4. An object recommendation method, characterized in that, include: Obtain the target user characteristics of the target user, the candidate target object characteristics of the candidate target object, and the spatiotemporal environment characteristics corresponding to the target user; The target user features, the candidate target object features, and the spatiotemporal environment features are input into the user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object; Based on the predicted data, a judgment result is obtained on whether to use the candidate target object as the recommended object for the target user; In the user operation prediction model, spatiotemporal semantic features are obtained based on the spatiotemporal environment features, the target user features, and the candidate target object features, and the prediction data is obtained based on the spatiotemporal semantic features. The step of obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: obtaining original semantic features based on the target user features and the candidate target object features; filtering out target user historical behavior sequences that match the spatiotemporal environment features from the target user's original historical behavior sequences; concatenating the spatiotemporal environment features and the target user historical behavior sequences that match the spatiotemporal environment features to obtain spatiotemporal historical behavior features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters; and performing spatiotemporal semantic transformation on the original semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters to obtain the spatiotemporal semantic features.

5. An object recommendation method, characterized in that, include: Obtain the target user characteristics of the target user, the candidate target object characteristics of the candidate target object, and the spatiotemporal environment characteristics corresponding to the target user; The target user features, the candidate target object features, and the spatiotemporal environment features are input into the user operation prediction model to obtain prediction data of the target user performing a specified operation on the candidate target object; Based on the predicted data, a judgment result is obtained on whether to use the candidate target object as the recommended object for the target user; In the user operation prediction model, a fusion output feature with spatiotemporal deviation characteristics is obtained based on the spatiotemporal environment features, the target user features, and the candidate target object features, and the prediction data is obtained based on the fusion output feature. The process of obtaining fused output features with spatiotemporal deviation characteristics based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: assigning weights to the target user features and the candidate target object features based on the spatiotemporal environment features; obtaining spatiotemporally weighted target user features based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features based on the candidate target object feature weights and the candidate target object features; and obtaining original semantic features based on the spatiotemporally weighted target user features and the spatiotemporally weighted candidate target object features, and from the target user's original historical data... To filter out target user historical behavior sequences that match the spatiotemporal environment features from the sequence, the spatiotemporal environment features and the target user historical behavior sequences that match the spatiotemporal environment features are concatenated to obtain spatiotemporal historical behavior features. Based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters, spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters are obtained. Based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters, the original semantic features are spatiotemporally semantically transformed to obtain spatiotemporal semantic features. Based on the spatiotemporal semantic features, the fused output features are obtained.

6. A user behavior prediction method, characterized in that, include: Obtain the target user characteristics of the target user, the target operation characteristics of the specified operation, and the spatiotemporal environment characteristics corresponding to the target user; The target user features, the target operation features, and the spatiotemporal environment features are input into the user operation prediction model to obtain prediction data of the target user performing the specified operation; In the user operation prediction model, weights are assigned to the target user features and the target operation features based on the spatiotemporal environment features, and the prediction data is obtained based on the target user feature weights and the target operation feature weights. The step of obtaining the predicted data based on the target user feature weights and target operation feature weights includes: obtaining spatiotemporally weighted target user features with spatiotemporal weights based on the target user feature weights and the target user features; obtaining spatiotemporally weighted target operation features with spatiotemporal weights based on the target operation feature weights and the target operation features; obtaining original semantic features based on the spatiotemporally weighted target user features and the spatiotemporally weighted target operation object features; filtering out target user historical behavior sequences that match the spatiotemporal environment features from the target user's original historical behavior sequences; and assigning the spatiotemporal environment features... The system performs connection processing on the target user's historical behavior sequence that matches the spatiotemporal environment features to obtain spatiotemporal historical behavior features. Based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters, it obtains spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters. Based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters, it performs spatiotemporal semantic transformation on the original semantic features to obtain spatiotemporal semantic features. Based on the spatiotemporal semantic features, it obtains fused output features. Based on the fused output features, it obtains the predicted data.

7. A user behavior prediction method, characterized in that, include: Obtain the target user characteristics of the target user, the target operation characteristics of the specified operation, and the spatiotemporal environment characteristics corresponding to the target user; The target user features, the target operation features, and the spatiotemporal environment features are input into the user operation prediction model to obtain prediction data of the target user performing the specified operation; In the user operation prediction model, spatiotemporal semantic features are obtained based on the spatiotemporal environment features, the target user features, and the target operation features, and the prediction data is obtained based on the spatiotemporal semantic features. The step of obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the target operation features includes: obtaining original semantic features based on the target user features and the target operation features; filtering out target user historical behavior sequences that match the spatiotemporal environment features from the target user's original historical behavior sequences; concatenating the spatiotemporal environment features and the target user historical behavior sequences that match the spatiotemporal environment features to obtain spatiotemporal historical behavior features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters; and transforming the original semantic features through spatiotemporal semantic transformation to obtain the spatiotemporal semantic features.

8. A user behavior prediction method, characterized in that, include: Obtain the target user characteristics of the target user, the target operation characteristics of the specified operation, and the spatiotemporal environment characteristics corresponding to the target user; The target user features, the target operation features, and the spatiotemporal environment features are input into the user operation prediction model to obtain prediction data of the target user performing the specified operation; In the user operation prediction model, a fusion output feature with spatiotemporal deviation characteristics is obtained based on the spatiotemporal environment features, the target user features, and the target operation features, and the prediction data is obtained based on the fusion output feature. Based on the spatiotemporal environment features, the target user features, and the target operation features, a fusion output feature with spatiotemporal deviation characteristics is obtained, including: obtaining original semantic features based on the target user features and the target operation features; filtering target user historical behavior sequences that match the spatiotemporal environment features from the target user's original historical behavior sequences; concatenating the spatiotemporal environment features and the target user historical behavior sequences that match the spatiotemporal environment features to obtain spatiotemporal historical behavior features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal historical behavior features, meta-network weight parameters, and meta-network bias parameters; performing spatiotemporal semantic transformation on the original semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters to obtain spatiotemporal semantic features; and obtaining the fusion output feature based on the spatiotemporal semantic features.

9. A method for obtaining model output results, characterized in that, include: Obtain the spatiotemporal environmental characteristics of the target user and multiple basic input features of the user operation prediction model; Based on the spatiotemporal environmental characteristics, weights are assigned to the multiple basic input features to obtain the spatiotemporal weights corresponding to the multiple basic input features respectively; Based on the multiple basic input features and the spatiotemporal weights corresponding to the multiple basic input features, multiple spatiotemporal weighted basic input features are obtained; Based on the multiple spatiotemporal weighted basic input features, the output result of the user operation prediction model is obtained; The step of obtaining the output result of the user operation prediction model based on the multiple spatiotemporal weighted basic input features includes: obtaining original semantic features based on the multiple spatiotemporal weighted basic input features; filtering out target user historical behavior sequences that match the spatiotemporal environment features from the target user's original historical behavior sequences; concatenating the spatiotemporal environment features and the target user historical behavior sequences that match the spatiotemporal environment features to obtain spatiotemporal historical behavior features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters; obtaining spatiotemporal semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters; obtaining fused output features based on the spatiotemporal semantic features; and obtaining the output result of the user operation prediction model based on the fused output features.

10. A method for obtaining model output results, characterized in that, include: Obtain the spatiotemporal environmental characteristics of the target user and multiple basic input features of the user operation prediction model; Based on the spatiotemporal environmental features and the multiple basic input features, spatiotemporal semantic features are obtained; Based on the spatiotemporal semantic features, the output of the user operation prediction model is obtained; The step of obtaining spatiotemporal semantic features based on the spatiotemporal environmental features and the plurality of basic input features includes: obtaining original semantic features based on the spatiotemporal weighted basic input features corresponding to the plurality of basic input features respectively; filtering out the target user's historical behavior sequence that matches the spatiotemporal environmental features from the target user's original historical behavior sequence; concatenating the spatiotemporal environmental features and the target user's historical behavior sequence that matches the spatiotemporal environmental features to obtain spatiotemporal historical behavior features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters; and transforming the original semantic features through spatiotemporal semantic transformation based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters to obtain the spatiotemporal semantic features.

11. A method for obtaining model output results, characterized in that, include: Obtain the spatiotemporal environmental characteristics of the target user and multiple basic input features of the user operation prediction model; Based on the spatiotemporal environmental features and the multiple basic input features, a fused output feature with spatiotemporal deviation characteristics is obtained; Based on the fused output features, the output result of the user operation prediction model is obtained; The step of obtaining a fusion output feature with spatiotemporal deviation features based on the spatiotemporal environment features and the multiple basic input features includes: obtaining original semantic features based on the spatiotemporal weighted basic input features corresponding to the multiple basic input features; filtering out target user historical behavior sequences that match the spatiotemporal environment features from the target user's original historical behavior sequences; concatenating the spatiotemporal environment features and the target user historical behavior sequences that match the spatiotemporal environment features to obtain spatiotemporal historical behavior features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal historical behavior features, meta-network weight parameters, and meta-network bias parameters; performing spatiotemporal semantic transformation on the original semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters to obtain spatiotemporal semantic features; and obtaining the fusion output feature based on the spatiotemporal semantic features.

12. A model training method applied to a user operation prediction model, characterized in that, include: The method involves obtaining prediction data of a target user performing a specified operation on a candidate target object and actual result data of the target user performing the specified operation on the candidate target object. The method of obtaining prediction data of a target user performing a specified operation on a candidate target object includes: assigning weights to target user features and candidate target object features based on spatiotemporal environmental characteristics, and obtaining the prediction data based on the target user feature weights and candidate target object feature weights. Based on the spatiotemporal environment features, the target user features, and the candidate target object features, spatiotemporal semantic features are obtained, and the prediction data is obtained based on the spatiotemporal semantic features. Based on the spatiotemporal environment features, the target user features, and the candidate target object features, a fusion output feature with spatiotemporal deviation features is obtained, and the prediction data is obtained based on the fusion output feature; The step of obtaining spatiotemporal semantic features based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: obtaining original semantic features based on the target user features and the candidate target object features; filtering out target user historical behavior sequences that match the spatiotemporal environment features from the target user's original historical behavior sequences; concatenating the spatiotemporal environment features and the target user historical behavior sequences that match the spatiotemporal environment features to obtain spatiotemporal historical behavior features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters; performing spatiotemporal semantic transformation on the original semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters to obtain the spatiotemporal semantic features; and obtaining the prediction data based on the spatiotemporal semantic features. The deviation between the predicted data and the actual result data is calculated. If the deviation between the predicted data and the actual result data does not exceed a predetermined deviation threshold, then the current user operation prediction model will be used as the user operation prediction model recommended for the object. If the deviation between the predicted data and the actual result data is greater than the predetermined deviation threshold, the current user operation prediction model will be trained until the deviation data meets the predetermined condition of not exceeding the predetermined deviation threshold.

13. An object recommendation method, characterized in that, include: In response to the fulfillment of the triggering conditions for displaying information, obtain the recommended objects to be recommended to the target user; The recommended objects are displayed on the display interface used to show information; The recommended object is determined based on prediction data of the target user performing a specified operation on the candidate target object; The prediction data is obtained based on spatiotemporal semantic features derived from spatiotemporal environment features, target user features, and candidate target object features. This process includes: obtaining original semantic features based on the target user features and candidate target object features; filtering target user historical behavior sequences that match the spatiotemporal environment features from the target user's original historical behavior sequences; concatenating the spatiotemporal environment features and the target user historical behavior sequences that match the spatiotemporal environment features to obtain spatiotemporal historical behavior features; obtaining spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters; and performing spatiotemporal semantic transformation on the original semantic features based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters to obtain the spatiotemporal semantic features.

14. An object recommendation method, characterized in that, include: In response to the fulfillment of the triggering conditions for displaying information, obtain the recommended objects to be recommended to the target user; The recommended objects are displayed on the display interface used to show information; The recommended object is determined based on prediction data of the target user performing a specified operation on the candidate target object; The predicted data is obtained by assigning target user feature weights to the target user's target user features based on the spatiotemporal environmental characteristics corresponding to the target user, and assigning candidate target object feature weights to the candidate target object's candidate target object features. This includes: obtaining spatiotemporal weighted target user features based on the target user feature weights and the target user features; obtaining spatiotemporal weighted candidate target object features based on the candidate target object feature weights and the candidate target object features; obtaining original semantic features based on the spatiotemporal weighted target user features and the spatiotemporal weighted candidate target object features; and filtering out features from the target user's original historical behavior sequence that correspond to the spatiotemporal environmental characteristics. The target user's historical behavior sequence is matched with the spatiotemporal environment features. The spatiotemporal environment features and the target user's historical behavior sequence that matches the spatiotemporal environment features are concatenated to obtain spatiotemporal historical behavior features. Based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters, spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters are obtained. Based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters, the original semantic features are spatiotemporally semantically transformed to obtain spatiotemporal semantic features. Based on the spatiotemporal semantic features, fused output features are obtained. Based on the fused output features, the predicted data is obtained.

15. An object recommendation method, characterized in that, include: In response to the fulfillment of the triggering conditions for displaying information, obtain the recommended objects to be recommended to the target user; The recommended objects are displayed on the display interface used to show information; The recommended object is determined based on prediction data of the target user performing a specified operation on the candidate target object; The prediction data is obtained by fusion output features with spatiotemporal deviation characteristics based on spatiotemporal environment features, target user features, and candidate target object features; The process of obtaining fused output features with spatiotemporal deviation characteristics based on the spatiotemporal environment features, the target user features, and the candidate target object features includes: assigning weights to the target user features and the candidate target object features based on the spatiotemporal environment features; obtaining spatiotemporally weighted target user features based on the target user feature weights and the target user features; obtaining spatiotemporally weighted candidate target object features based on the candidate target object feature weights and the candidate target object features; and obtaining original semantic features based on the spatiotemporally weighted target user features and the spatiotemporally weighted candidate target object features, and from the target user's original historical data... To filter out target user historical behavior sequences that match the spatiotemporal environment features from the sequence, the spatiotemporal environment features and the target user historical behavior sequences that match the spatiotemporal environment features are concatenated to obtain spatiotemporal historical behavior features. Based on the spatiotemporal historical behavior features, metanetwork weight parameters, and metanetwork bias parameters, spatiotemporal semantic fully connected layer weight parameters and spatiotemporal semantic fully connected layer bias parameters are obtained. Based on the original semantic features, the spatiotemporal semantic fully connected layer weight parameters, and the spatiotemporal semantic fully connected layer bias parameters, the original semantic features are spatiotemporally semantically transformed to obtain spatiotemporal semantic features. Based on the spatiotemporal semantic features, the fused output features are obtained.

Images