Information push model construction method and device, computer device, and storage medium

Abstract

The application relates to an information push model construction method and device, computer equipment and a storage medium. The method comprises the following steps: obtaining historical feature data; obtaining representation embedding data according to the historical feature data, wherein the representation embedding data comprises object representation data, push content representation data, push mode representation data and push time representation data; taking the object representation data and the push content representation data as training data to perform neural network training to obtain a push content submodel; taking the object representation data and the push mode representation data as training data to perform neural network training to obtain a push mode submodel; taking the object representation data and the push time representation data as training data to perform neural network training to obtain a push time submodel; and parallelly splicing the push content submodel, the push mode submodel and the push time submodel to obtain an information push model. The method can improve the efficiency of generating online push information.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, computer device, and storage medium for constructing an information push model. Background Technology

[0002] With the development of computer technology, the method of using intelligent algorithms to analyze data and generate online push information has been increasingly widely used, and how to generate online push information efficiently and accurately has received more and more attention.

[0003] Traditional technologies generate online push notifications based on a single decision condition, which cannot accurately and efficiently generate targeted push notifications and results in high push costs. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer device, and computer-readable storage medium for constructing an information push model to address the aforementioned technical problems and improve the efficiency of generating online push information.

[0005] A method for constructing an information push model, characterized in that the method includes:

[0006] Obtain historical feature data;

[0007] The representation embedding data is obtained based on historical feature data. The representation embedding data includes object representation data, push content representation data, push method representation data, and push time representation data.

[0008] The object representation data and the push content representation data are used as training data to train a neural network to obtain a push content sub-model;

[0009] The object representation data and the push method representation data are used as training data to train a neural network to obtain a push method sub-model;

[0010] The object representation data and the push time representation data are used as training data to train a neural network to obtain a push time sub-model.

[0011] The information push model is obtained by concatenating the push content sub-model, push method sub-model, and push time sub-model in parallel. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time based on the object representation data corresponding to the object to be pushed.

[0012] In one embodiment, after obtaining historical feature data, the method further includes:

[0013] Obtain initial historical feature data;

[0014] Correlation analysis is performed on the feature data of each dimension in the initial historical feature data to obtain the feature data of the relevant dimensions.

[0015] Importance analysis is performed on the relevant dimension feature data to determine the feature importance of the relevant dimension feature data;

[0016] Historical feature data is generated by sorting relevant dimension feature data according to feature importance.

[0017] In one embodiment, characterization embedding data is obtained based on historical feature data, including:

[0018] Retrieve target object feature data from historical feature data;

[0019] Obtain historical activity data corresponding to the feature data of the target object;

[0020] The time spent browsing objects in historical activity data is normalized to obtain the time weight parameter;

[0021] Each historical activity in the historical activity data is mapped to a corresponding historical activity vector, and the dimensions of the historical activity vectors corresponding to each historical activity are the same.

[0022] The representation embedding data is obtained by fusing the time weight parameter and the historical activity vector corresponding to each historical activity. The representation embedding data is used to represent the correspondence between the push content, push method and push time of each historical activity corresponding to the feature data of the target object.

[0023] In one embodiment, the representation embedding data is obtained by fusing time weight parameters and historical activity vectors corresponding to each historical activity, including:

[0024] The initial embedded representation data is obtained by fusing the time weight parameters and the historical activity vectors corresponding to each historical activity.

[0025] The initial embedded representation data is processed by a multilayer perceptron to obtain the representation embedded data, and the dimension of the representation embedded data is lower than that of the initial embedded representation data.

[0026] In one embodiment, object representation data and push content representation data are used as training data to train a neural network to obtain a push content sub-model, including:

[0027] The object representation data is projected through a multilayer perceptron to generate an object representation vector.

[0028] The push content representation data is projected through a multilayer perceptron to generate a push content representation vector. The dimension of the push content representation vector is the same as the dimension of the object representation vector.

[0029] The push weight is obtained by the dot product of the push content representation vector and the object representation vector;

[0030] The object representation vector, the push content representation vector, and the push weight are used as training sample data for the multilayer perceptron to obtain the push content sub-model.

[0031] In one embodiment, object representation data and push method representation data are used as training data to train a neural network to obtain a push method sub-model, including:

[0032] The object representation data is projected through a multilayer perceptron to generate an object representation vector.

[0033] Based on the application push representation data, SMS push representation data, and telephone push representation data in the push method representation data, generate application push representation vector, SMS push representation vector, and telephone push representation vector respectively.

[0034] The application push weight is obtained by the dot product of the object representation vector and the application push representation vector.

[0035] The SMS push weight is obtained by the dot product of the object representation vector and the SMS push representation vector.

[0036] The telephone push weight is obtained by the dot product of the object representation vector and the telephone push representation vector.

[0037] The push method sub-model is obtained by using the object representation vector, application push weight, SMS push weight, and telephone push weight as model training sample data.

[0038] In one embodiment, object representation data and push time representation data are used as training data to train a neural network to obtain a push time sub-model, including:

[0039] The object representation data is projected through a multilayer perceptron to generate an object representation vector.

[0040] A preset number of sub-time period characterization data are obtained by dividing the push time characterization data;

[0041] The sub-time period representation data is projected through a multilayer perceptron to generate sub-time period representation vectors.

[0042] The probability of object idle time in the sub-time period corresponding to the sub-time period representation data is obtained by using the object representation vector and the sub-time period representation vector.

[0043] The object idle time probability and object representation vector of each sub-time period are used as input sample data for training the multilayer perceptron to obtain the push time sub-model.

[0044] An information push model construction device, characterized in that the device comprises:

[0045] The data acquisition module is used to acquire historical feature data;

[0046] The representation embedding data generation module is used to obtain representation embedding data based on historical feature data. The representation embedding data includes object representation data, push content representation data, push method representation data, and push time representation data.

[0047] The model generation module is used to train a neural network using object representation data and push content representation data as training data to obtain a push content sub-model; to train a neural network using object representation data and push method representation data as training data to obtain a push method sub-model; and to train a neural network using object representation data and push time representation data as training data to obtain a push time sub-model. Based on the push content sub-model, push method sub-model, and push time sub-model, the information push model is obtained by concatenating them in parallel. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time according to the object representation data corresponding to the object to be pushed.

[0048] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:

[0049] Obtain historical feature data;

[0050] The representation embedding data is obtained based on historical feature data. The representation embedding data includes object representation data, push content representation data, push method representation data, and push time representation data.

[0051] The object representation data and the push content representation data are used as training data to train a neural network to obtain a push content sub-model;

[0052] The object representation data and the push method representation data are used as training data to train a neural network to obtain a push method sub-model;

[0053] The object representation data and the push time representation data are used as training data to train a neural network to obtain a push time sub-model.

[0054] The information push model is obtained by concatenating the push content sub-model, push method sub-model, and push time sub-model in parallel. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time based on the object representation data corresponding to the object to be pushed.

[0055] A computer-readable storage medium storing a computer program, which, when executed by a processor, performs the following steps:

[0056] Obtain historical feature data;

[0057] The representation embedding data is obtained based on historical feature data. The representation embedding data includes object representation data, push content representation data, push method representation data, and push time representation data.

[0058] The object representation data and the push content representation data are used as training data to train a neural network to obtain a push content sub-model;

[0059] The object representation data and the push method representation data are used as training data to train a neural network to obtain a push method sub-model;

[0060] The object representation data and the push time representation data are used as training data to train a neural network to obtain a push time sub-model.

[0061] The information push model is obtained by concatenating the push content sub-model, push method sub-model, and push time sub-model in parallel. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time based on the object representation data corresponding to the object to be pushed.

[0062] The aforementioned information push model construction method, apparatus, computer equipment, and storage medium acquire historical feature data, obtain representational embedding data based on the historical feature data, and the representational embedding data includes object representation data, push content representation data, push method representation data, and push time representation data. The object representation data and push content representation data are used as training data to train a neural network to obtain a push content sub-model, the object representation data and push method representation data are used as training data to train a neural network to obtain a push method sub-model, and the object representation data and push time representation data are used as training data to train a neural network to obtain a push time sub-model. The information push model is obtained by concatenating the push content sub-model, push method sub-model, and push time sub-model in parallel. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time based on the object representation data corresponding to the object to be pushed. In this way, by training neural networks based on object representation data and push content representation data, push method representation data, and push time representation data respectively, we can obtain various sub-models. Then, we can concatenate these sub-models in parallel to obtain the entire information push model. Compared with a non-concatenated single neural network model, the multi-task learning model, which is composed of multiple corresponding dimension sub-models concatenated in parallel, has better generalization ability and the training of the model is more accurate and efficient. This enables the trained information push model to improve the efficiency of generating online push information. Attached Figure Description

[0063] Figure 1This is an application environment diagram of the information push model construction method in one embodiment;

[0064] Figure 2 This is a flowchart illustrating an information push model construction method in one embodiment;

[0065] Figure 3 This is a schematic diagram of the process for generating historical feature data in one embodiment;

[0066] Figure 4 This is a schematic diagram illustrating the process of generating embedded data in one embodiment;

[0067] Figure 5 This is a schematic diagram illustrating the process of generating embedded data in one embodiment;

[0068] Figure 6 This is a schematic diagram illustrating the process of generating a push content sub-model in one embodiment;

[0069] Figure 7 This is a schematic diagram illustrating the process of generating a push method sub-model in one embodiment;

[0070] Figure 8 This is a schematic diagram of the process for generating the push time sub-model in one embodiment;

[0071] Figure 9 This is a structural block diagram of an information push model construction device in one embodiment;

[0072] Figure 10 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0074] The information push model construction method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, computer device 102 acquires historical feature data; based on the historical feature data, it obtains representation embedding data, which includes object representation data, push content representation data, push method representation data, and push time representation data; it uses the object representation data and push content representation data as training data to train a neural network to obtain a push content sub-model; it uses the object representation data and push method representation data as training data to train a neural network to obtain a push method sub-model; it uses the object representation data and push time representation data as training data to train a neural network to obtain a push time sub-model; and it concatenates the push content sub-model, push method sub-model, and push time sub-model in parallel to obtain an information push model. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time based on the object representation data corresponding to the object to be pushed. Specifically, computer device 102 may include, but is not limited to, various personal computers, laptops, servers, smartphones, tablets, smart cameras, and portable wearable devices.

[0075] In one embodiment, such as Figure 2 As shown, an information push model construction method is provided, which is applied to... Figure 1 Taking computer device 102 as an example, the following steps are included:

[0076] Step S202: Obtain historical feature data.

[0077] Among them, historical feature data is historical user behavior feature data, which is data generated by users' past participation in activities. Historical feature data includes user-side historical data and activity-side historical data. User-side historical data is user feature data, such as user's age, gender, occupation, etc. Activity-side historical data is activity feature data, such as the activity's holding time, activity content, activity push information method, push information time, etc.

[0078] Specifically, the computer equipment filters the proposed features of historical activity data from the data pool based on the business logic of the business department and expert knowledge, filters user-side feature data and activity-side feature data, and then combines the user-side feature data and activity-side feature data to form historical feature data.

[0079] Step S204: Obtain representation embedding data based on historical feature data. The representation embedding data includes object representation data, push content representation data, push method representation data, and push time representation data.

[0080] Among them, the representation embedding data is used to represent the features of each dimension of historical feature data, the object representation data is used to represent the features of each dimension of user-side feature data, the push content representation data is used to represent the various types of push content in the activity-side feature data, the push method representation data is used to represent the information push method type in the activity-side feature data, and the push time representation data is used to represent the different push time period types in the activity-side feature data.

[0081] Specifically, the computer device performs embedding representation processing on the user-side feature data from the historical feature data obtained in the aforementioned steps to obtain user historical interaction information and feature statistics. Considering that each user may participate in multiple different historical activities, corresponding activity weights are set according to the historical interaction information corresponding to the current user, and then weighted fusion is performed to obtain the object representation parameters of the historical activity information corresponding to the current user in the user-side feature data. The object representation parameters are then concatenated with the feature statistics and used as input to the neural network model to generate lower-dimensional embedded representation data, i.e., object representation data. The activity-side feature data is then processed in the same way to obtain lower-dimensional push content representation data, push method representation data, and push time representation data, thereby obtaining representation embedding data.

[0082] Step S206: Use the object representation data and the push content representation data as training data to train the neural network and obtain the push content sub-model.

[0083] Among them, the dimensions and content of object representation data and push content representation data are different, that is, the data dimensions of the two are different. The neural network can be a multilayer perceptron, a generalized neural network, a convolutional neural network, etc., without specific restrictions here.

[0084] Specifically, the computer device uses a multilayer perceptron to project object representation data and push content representation data into the same low-dimensional space to obtain the representation in the new space. Then, based on the representation of the two in the new space, the probability score between the current user's object representation data and the recommended content representation data is obtained. This probability score is used as the standard output data of the model, and the binary cross-entropy is used as the loss function of the multilayer perceptron model for training to obtain the push content sub-model.

[0085] Step S208: Use the object representation data and the push method representation data as training data to train the neural network and obtain the push method sub-model.

[0086] Among them, the dimensions and content of object representation data and push method representation data are different, that is, the data dimensions of the two are different. The neural network can be a multilayer perceptron, a generalized neural network, a convolutional neural network, etc., without specific restrictions. Push method representation data includes various types of push methods such as application push, SMS push, and telephone push.

[0087] Specifically, the computer device uses a multilayer perceptron to project the object representation data and the push method representation data of different push method types into the same low-dimensional space to obtain the representation in the new space. Then, based on the representation of the two in the new space, the probability score between the current user's object representation data and the push method type data is obtained. This probability score is used as the standard output data of the model. Since users may accept multiple push methods, a pair-wise model is used for optimization, and finally, a push method sub-model is trained.

[0088] Step S210: Use the object representation data and the push time representation data as training data to train the neural network and obtain the push time sub-model.

[0089] The object representation data and the push time representation data have different dimensions and content, that is, the data dimensions of the two are different. The neural network can be a multilayer perceptron, a generalized neural network, a convolutional neural network, etc. There are no specific restrictions here. The push time representation data divides the user's idle time from 8:00 to 20:00 into segments of two hours and labels them, and uses One-Hot encoding. At the same time, it is combined with whether it is a weekday, rest day, holiday, or promotional day as a combination feature at the event level.

[0090] Specifically, the computer device inputs the object representation data and the push time representation data into a multilayer perceptron and projects them into the same low-dimensional space. The data in the push time representation data for different time periods are assigned corresponding labels. For example, the labels used include multiple labels such as transaction after push, user access to service after push, no response after push, and complaint after push. The probability score of each label for each time period is obtained. The probability score is used as the standard output of the neural network. During the training process, multi-class cross-entropy is used as the loss function for optimization to obtain the push time sub-model.

[0091] Step S212: Based on the parallel concatenation of the push content sub-model, push method sub-model, and push time sub-model, an information push model is obtained. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time according to the object representation data corresponding to the object to be pushed.

[0092] In this embodiment, historical feature data is acquired, and representation embedding data is obtained based on the historical feature data. The representation embedding data includes object representation data, push content representation data, push method representation data, and push time representation data. The object representation data and push content representation data are used as training data to train a neural network to obtain a push content sub-model, the object representation data and push method representation data are used as training data to train a neural network to obtain a push method sub-model, and the object representation data and push time representation data are used as training data to train a neural network to obtain a push time sub-model. The information push model is obtained by concatenating the push content sub-model, push method sub-model, and push time sub-model in parallel. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time based on the object representation data corresponding to the object to be pushed. In this way, by training neural networks based on object representation data and push content representation data, push method representation data, and push time representation data respectively, we can obtain various sub-models. Then, we can concatenate these sub-models in parallel to obtain the entire information push model. Compared with a non-concatenated single neural network model, the multi-task learning model, which is composed of multiple corresponding dimension sub-models concatenated in parallel, has better generalization ability and the training of the model is more accurate and efficient. This enables the trained information push model to improve the efficiency of generating online push information.

[0093] In one embodiment, such as Figure 3 As shown, after obtaining historical feature data, the process also includes:

[0094] Step S302: Obtain initial historical feature data.

[0095] The initial historical feature data consists of feature data generated from all users' participation in historical activities.

[0096] Step S304: Perform correlation analysis on the feature data of each dimension in the initial historical feature data to obtain the relevant dimension feature data.

[0097] Specifically, the computer equipment selects features to be used based on experience (the number of features ranges from dozens to hundreds). Then, it uses the Spearman correlation coefficient to analyze the features with strong correlation among all the features to be used. Features with a correlation that reaches a preset threshold are used as the relevant dimension feature data. The correlation between features is calculated according to the formula shown in Formula 1 below:

[0098]

[0099] This indicates that for two characteristic variables (x) i ,y i The paired values ​​are ranked in ascending (or descending) order, R. i Represents xi The rank of Q i Represents y i The rank of R i -Q i For x i y i The difference in rank.

[0100] Step S306: Perform importance analysis on the relevant dimension feature data to determine the feature importance of the relevant dimension feature data.

[0101] Specifically, the computer equipment performs simple logistic regression and preliminary tree modeling based on the relevant dimensional feature data determined in the aforementioned steps, and calculates the feature importance of each feature in the relevant dimensional feature data.

[0102] Step S308: Sort the relevant dimension feature data according to feature importance to generate historical feature data.

[0103] In this embodiment, the computer device acquires initial historical feature data, performs correlation analysis on each dimension feature to obtain highly correlated related dimension feature data, and then uses simple logistic regression and tree model to perform preliminary modeling analysis to obtain the feature importance of each dimension feature in the related dimension feature data. Finally, it sorts the data according to feature importance to obtain historical feature data, effectively eliminating invalid and redundant data and enhancing the reliability of historical feature data.

[0104] In one embodiment, such as Figure 4 As shown, the representation embedding data obtained based on historical feature data includes:

[0105] Step S402: Obtain the target object feature data from the historical feature data.

[0106] Among them, the target object feature data is user data from historical activities, including basic information such as user age, occupation, and work location.

[0107] Step S404: Obtain historical activity data corresponding to the target object feature data.

[0108] Specifically, the computer device retrieves all historical activity data associated with the target object's feature data from the database.

[0109] Step S406: Normalize the object browsing dwell time in the historical activity data to obtain the time weight parameter.

[0110] Specifically, the computer device determines the historical activity data corresponding to each target object feature data in the aforementioned steps, and obtains the browsing dwell time of the target user corresponding to the target object feature data in each historical activity data, and normalizes the browsing dwell time to obtain the time weight.

[0111] Step S408: Map each historical activity in the historical activity data to a corresponding historical activity vector, with each historical activity vector having the same dimension.

[0112] Specifically, the computer equipment analyzes each historical activity in the historical activity data from various dimensions, maps the historical activities into corresponding historical activity vectors according to preset rules based on their characteristics, and ensures that the dimensions of the historical activity vectors corresponding to each historical activity are the same.

[0113] Step S410: Based on the time weight parameter and the historical activity vector corresponding to each historical activity, the representation embedding data is obtained by fusing them. The representation embedding data is used to represent the correspondence between the push content, push method and push time of each historical activity corresponding to the feature data of the target object.

[0114] Specifically, the computer device performs embedding representation mapping on each historical activity vector, and then fuses it with the time weight parameters corresponding to each activity to obtain representation embedding data.

[0115] In this embodiment, target object feature data is obtained from historical feature data, and corresponding historical activity data is obtained. The browsing dwell time of the target object in the historical activity data is normalized to obtain a time weight parameter. Each historical activity in the historical activity data is mapped to a corresponding historical activity vector, and the dimensions of the historical activity vectors corresponding to each historical activity are the same. Finally, the representation embedding data is obtained by fusing the time weight parameter and the historical activity vectors corresponding to each historical activity. The representation embedding data is used to represent the correspondence between the push content, push method and push time of each historical activity corresponding to the target object feature data. It can accurately allocate appropriate time weights according to the current user's browsing time of each historical activity, improve the weight importance of historical activities with high user relevance, and make the generated representation embedding data more reflective of the degree of correlation between different activities, thereby improving the accuracy of the generated representation embedding data.

[0116] In one embodiment, such as Figure 5 As shown, the representation embedding data is obtained by fusing the time weight parameters and the historical activity vectors corresponding to each historical activity, including:

[0117] Step S502: The initial embedded representation data is obtained by fusing the time weight parameters and the historical activity vectors corresponding to each historical activity.

[0118] Step S504: The initial embedded representation data is processed by a multilayer perceptron to obtain representation embedded data. The dimension of the representation embedded data is lower than that of the initial embedded representation data.

[0119] Specifically, the computer device calculates the representation of the embedded data User according to the following formulas 2 to 4. emb :

[0120] User history_emb =RELU(a1*EMB(sale1)+a2*EMB(sale2)+…+a n *EMB(sale n )),

[0121] Formula 2

[0122] Where ReLU is the activation function, a n The time weight can be directly defined by the user's browsing time for the corresponding activity or automatically learned by the attention mechanism. EMB stands for Embedding function, which can uniquely project discrete historical activities into a high-dimensional dense space.

[0123] Then User history_emb It is concatenated with the user's basic statistical feature, User_feature, as shown in the following formula:

[0124] User_allemb = Concat(User history_emb ,User_feature), Formula 3

[0125] The basic statistical features of users, User_feature, include the basic characteristics and behavioral characteristics of customers.

[0126] Subsequently, based on User_allemb, a low-dimensional representation embedding data User is obtained through multilayer perceptron processing. emb Specifically, as shown in Formula 4 below:

[0127] User emb =MLP(User_allemb), Formula 4

[0128] MLP stands for Multilayer Perceptron.

[0129] In this embodiment, initial embedded representation data is obtained by fusing time weight parameters and historical activity vectors corresponding to each historical activity. The initial embedded representation data is then processed by a multilayer perceptron to obtain representation embedded data. The dimension of the representation embedded data is lower than that of the initial embedded representation data. In this way, the lengthy initial embedded representation data is simplified by the multilayer perceptron to generate lower-dimensional representation embedded data, effectively saving data computational resources.

[0130] In one embodiment, such as Figure 6 As shown, the push content sub-model is obtained by training a neural network using object representation data and push content representation data as training data, including:

[0131] Step S602: Project the object representation data through a multilayer perceptron to generate an object representation vector.

[0132] Specifically, the computer device projects object representation data that can characterize the features of each dimension of the user-side feature data as input to the multilayer perceptron to generate the corresponding object representation vector. Other neural network models can also be used instead of the multilayer perceptron.

[0133] Step S604: The push content representation data is projected through a multilayer perceptron to generate a push content representation vector. The dimension of the push content representation vector is the same as the dimension of the object representation vector.

[0134] Specifically, the computer device projects the push content representation data that can characterize the feature data of the activity side as the input of the multilayer perceptron to generate the corresponding push content representation vector. The multilayer perceptron can also be replaced by other neural network models. The dimension of the push content representation vector is the same as the dimension of the object representation vector.

[0135] Step S606: Obtain the push weight based on the dot product of the push content representation vector and the object representation vector.

[0136] Among them, the push weight is used to represent the probability that the content of the current push information will be successfully pushed.

[0137] Specifically, the computer device calculates the dot product of the push content representation vector and the object representation vector according to the formula shown in Formula 5 below, which is the push weight score:

[0138] score=Sigmoid(u(x)·v(y).T), Formula 5

[0139] Where .T denotes the transpose of the vector, u(x) denotes the object representation vector, v(y) denotes the push content representation vector, u(x)·v(y) denotes the dot product of the push content representation vector and the object representation vector, and Sigmoid(u(x)·v(y).T) denotes integrating the dot product of the two into the interval (0,1).

[0140] Step S608: Use the object representation vector, the push content representation vector, and the push weight as training sample data for the multilayer perceptron to obtain the push content sub-model.

[0141] Specifically, the computer device labels the push weights of each group of object representation vectors and push content representation vectors. The labels include extremely high push success rate, relatively high push success rate, moderate push success rate, low push success rate, and extremely low push success rate. The push weights and corresponding labels are used to form training sample data, which is used as training samples for the multilayer perceptron to obtain the push content sub-model. Binary cross-entropy (BCE) is used as the loss function in the training, as shown in Formula 6 below:

[0142] L1=BCE(score,label), Formula 6

[0143] Where label is the tag identifier, score is the push weight, and L1 is the loss error.

[0144] In this embodiment, object representation data is projected through a multilayer perceptron to generate object representation vectors, and push content representation data is projected through a multilayer perceptron to generate push content representation vectors. The dimensions of the push content representation vectors are the same as those of the object representation vectors. Push weights are obtained based on the dot product of the push content representation vectors and the object representation vectors. The object representation vectors, push content representation vectors, and push weights are used as training sample data for the multilayer perceptron to train the push content sub-model. The push content representation data and object representation data are projected into vectors in the same dimensional space, and the probability of successful push is calculated using their dot product. This is then used as sample data for model training, effectively improving the accuracy of model training.

[0145] In one embodiment, such as Figure 7 As shown, the push method sub-model is obtained by training a neural network using object representation data and push method representation data as training data, including:

[0146] Step S702: Project the object representation data through a multilayer perceptron to generate an object representation vector.

[0147] Specifically, the computer device projects object representation data that can characterize the features of each dimension of the user-side feature data as input to the multilayer perceptron to generate the corresponding object representation vector. Other neural network models can also be used instead of the multilayer perceptron.

[0148] Step S704: Generate application push representation vector, SMS push representation vector, and telephone push representation vector respectively based on the application push representation data, SMS push representation data, and telephone push representation data in the push method representation data.

[0149] Specifically, the computer device projects the push method representation data, which can characterize the feature data of the activity side, as the input of the multilayer perceptron to generate the corresponding push content representation vector. Specifically, the application push representation data, SMS push representation data, and telephone push representation data in the push method representation data are used to generate application push representation vector, SMS push representation vector, and telephone push representation vector, respectively. Other neural network models can also be used instead of the multilayer perceptron. The vector dimensions of the application push representation vector, SMS push representation vector, and telephone push representation vector are the same as those of the object representation vector.

[0150] Step S706: Obtain the application push weight based on the dot product of the object representation vector and the application push representation vector.

[0151] Among them, the application push weight is used to characterize the probability of successful push when using application software to push information.

[0152] Step S708: Obtain the SMS push weight based on the dot product of the object representation vector and the SMS push representation vector.

[0153] Among them, SMS push weight is used to characterize the probability of successful push when using SMS push information.

[0154] Step S710: Obtain the telephone push weight based on the dot product of the object representation vector and the telephone push representation vector.

[0155] Among them, the telephone push weight is used to characterize the probability of successful push when using telephone to push information.

[0156] Step S712: Use the object representation vector, application push weight, SMS push weight, and telephone push weight as model training sample data to obtain the push method sub-model.

[0157] Specifically, the computer device calculates the dot product between the application push weight, SMS push weight, and telephone push weight and the current object representation vector to obtain the success probability of the push corresponding to the current object representation vector for each push method. Then, the recommendation weights of different push methods are labeled and combined into sample training data to train a multilayer perceptron to obtain a push method sub-model. Since the object representation vector corresponds to multiple push methods, a pair-wise model is used for optimization, as shown in Equation 7 below:

[0158]

[0159] Where, N S Let S represent the set of all activities that generate associations with object S, L2 be the loss error, and σ represent the activation function, which can be the sigmoid function. i For the score of the positive sample, s j The score is the score of the negative sample. Through training, the difference between the scores of the positive and negative samples is maximized.

[0160] In this embodiment, object representation data is projected through a multilayer perceptron to generate object representation vectors. Application push representation vectors, SMS push representation vectors, and telephone push representation vectors are generated based on application push representation data, SMS push representation data, and telephone push representation data from the push method representation data. Application push weights are obtained by taking the dot product of the object representation vector and the application push representation vector; SMS push weights are obtained by taking the dot product of the object representation vector and the SMS push representation vector; and telephone push weights are obtained by taking the dot product of the object representation vector and the telephone push representation vector. The object representation vectors, application push weights, SMS push weights, and telephone push weights are used as training sample data to obtain the push method sub-model. The dot product between various recommendation methods and the object representation vector is used as the probability value of successful push, improving the reliability of the model.

[0161] In one embodiment, such as Figure 8 As shown, the object representation data and push time representation data are used as training data to train a neural network to obtain a push time sub-model, including:

[0162] Step S802: Project the object representation data through a multilayer perceptron to generate an object representation vector.

[0163] Specifically, the computer device projects object representation data that can characterize the features of each dimension of the user-side feature data as input to the multilayer perceptron to generate the corresponding object representation vector. Other neural network models can also be used instead of the multilayer perceptron.

[0164] Step S804: Based on the push time representation data, a preset number of sub-time period representation data are obtained.

[0165] Specifically, the computer device segments the idle time of the objects corresponding to the object representation data in step S802 from 8:00 to 20:00 into two-hour segments, labels them, and uses One-Hot encoding. It also concatenates whether it is a weekday, rest day, holiday, or event day as a time-level combined feature (Torch_feature), and uses the following formula 8 to learn the user's time sub-segment representation data Time. emb :

[0166] Time emb =MLP(Torch_feature), formula 8

[0167] MLP stands for Multilayer Perceptron.

[0168] Step S806: Project the sub-time period representation data through a multilayer perceptron to generate a sub-time period representation vector.

[0169] The sub-time period representation vector has the same dimension as the object representation vector in step S802 above.

[0170] Step S808: Obtain the object idle time probability of the sub-time period corresponding to the sub-time period representation data based on the object representation vector and the sub-time period representation vector.

[0171] Specifically, the computer device uses a multilayer perceptron to obtain representations of the six label dimensions of the time dimension, and obtains the object idle time probability of the sub-time period corresponding to the sub-time period characterization data.

[0172] Step S810: The object idle time probability and the object representation vector of each sub-time period are used as input sample data of the multilayer perceptron to train the push time sub-model.

[0173] Specifically, the computer device uses the softmax function for activation and multi-class cross-entropy as the loss function, as shown in Equations 9 and 10 below:

[0174] p = Softmax(MLP(Concat(User) emb Time emb ))), Formula 9

[0175]

[0176] Where p = [p0, ..., p C-1 ] is a probability distribution, where each element p iLet y represent the probability that a sample belongs to the i-th class; y = [y0, ..., y1] C-1 ] is the One-Hot representation of the sample label, where y represents the sample belonging to the i-th category. i =1, otherwise y i =0; C is the sample label.

[0177] The object idle time probability and the object representation vector in each sub-time period are used as input sample data for the multilayer perceptron.

[0178] In this embodiment, object representation data is projected onto a multilayer perceptron to generate object representation vectors. Based on the push time representation data, a preset number of sub-time period representation data are obtained. The sub-time period representation data is then projected onto a multilayer perceptron to generate sub-time period representation vectors. The object idle time probability of the sub-time period corresponding to the sub-time period representation data is obtained based on the object representation vector and the sub-time period representation vector. The object idle time probability of each sub-time period and the object representation vector are used as input sample data for the multilayer perceptron to train the push time sub-model. The push time sub-model trained in time periods can provide more accurate success probability for each time period, thus improving the accuracy of the push time sub-model.

[0179] This application also provides an application scenario in which the above-mentioned information push model construction method is applied. This method is applied to the scenario of pushing marketing campaign information. Specifically, the application of the information push model construction method in this application scenario is as follows:

[0180] Computer devices collect behavioral features from historical data from both the user and marketing sides. Based on expert knowledge from business departments and methods such as feature importance calculation, input features are selected and then fed into a neural network to learn shallow embedding representations of users and marketing content.

[0181] Features to be used are selected from the non-sensitive data pool of internet financial institutions. Spearman correlation coefficient analysis is performed on all features to analyze the correlation between specific features and marketing reach tags. Invalid and redundant features are eliminated. Simple logistic regression and tree model are used for preliminary modeling, and the feature importance of each feature is calculated. Based on the corresponding feature importance, the best dozens of dimensions are selected for subsequent modeling to accelerate model training and inference speed.

[0182] After selecting the input features, embedding representations are performed on users and marketing campaigns. The user embedding representation mainly includes two sources: their historical interaction information and feature statistics.

[0183] Its formalized description is as follows:

[0184] User history_emb =RELU(a1*EMB(sale1)+a2*EMB(sale_2)+…+a n *EMB(sale n )),

[0185] Formula 11

[0186] Where ReLU is the activation function, a n The attention weights can be defined directly by the user's dwell time or learned automatically by the attention mechanism. It is sufficient to ensure that the sum of all attention weights is 1. EMB() stands for Embedding function, which can uniquely project discrete historical activity IDs into a high-dimensional dense space.

[0187] Next, the embedded representation obtained from this historical interaction is concatenated with the user's basic statistical characteristics, using the following formula:

[0188] User_allemb = Concat(User history_emb ,User_feature) Formula 12

[0189] This is then processed by a multilayer perceptron to obtain a lower-dimensional embedding representation of the user:

[0190] User emb =MLP(User_allemb) Formula 13

[0191] MLP stands for Multilayer Perceptron.

[0192] The same principle applies to the marketing campaign side, where we obtain embedded representations from both ends, which can be used by the three downstream tasks. Since these tasks are highly correlated, multi-task learning techniques are employed to learn all three tasks simultaneously.

[0193] Task 1: Since users and activities are fundamentally different, a multilayer perceptron is first used for projection, projecting the representations of users and activities (x and y) onto the same low-dimensional space to obtain the representations in the new space, u(x) and v(y). With these two representations, the formula is used:

[0194] score=Sigmoid(u(x)·v(y).T), Formula 14

[0195] Where .T represents the transpose of the vector, and the final prediction score is the dot product of the two vectors, which is then integrated into the (0,1) interval using the Sigmoid function.

[0196] Then, the binary cross-entropy (BCE) loss was used to optimize task 1.

[0197] L1=BCE(score,label), Formula 15

[0198] Task 2: Conduct a detailed evaluation of marketing outreach methods and recommend appropriate outreach methods for this marketing campaign.

[0199] In this task, the marketing reach methods (Torch_way) are mainly divided into APP marketing, SMS marketing, telephone marketing, etc., with limited options. Therefore, we use the Pair-wise model as a basis to score different reach methods.

[0200] Similarly, we perform embedding processing on different reach methods:

[0201] Torchway emb =EMB(Torch_way), Formula 16

[0202] The embedding of these outreach methods is processed through a model structure similar to Task 1 to obtain user preference scores for different outreach methods.

[0203] Since users are likely to accept multiple reach methods, a pair-wise model is used for optimization.

[0204]

[0205] Where s i For the score of the positive sample, s j The score is the score of the negative sample. Through training, the difference between the scores of the positive and negative samples is maximized.

[0206] Task 3: Learn the probability of users' free time to avoid disturbing them, while also increasing the chances of reaching them.

[0207] Users' free time often changes due to various objective factors. Task 3 divides users' free time from 8:00 to 20:00 into two-hour segments, labels them, and uses One-Hot encoding. It also concatenates whether it is a weekday, rest day, holiday, or promotional day as a time-level combined feature (Torch_feature), and learns the embedding representation of user time using the following formula:

[0208] Time emb =MLP(Torch_feature), Formula 18

[0209] MLP stands for Multilayer Perceptron.

[0210] After obtaining the embedded representation of time, it is also fed into a model structure similar to Task 1 to process and obtain the user's preference for different touch times.

[0211] During training, a multi-class loss function was used, and the labels used included various labels such as post-marketing transactions, post-marketing user access to services, post-marketing no response, and post-marketing complaints.

[0212] Instead of using vector dot product as the final score, we use MLP to obtain a six-label dimension representation, and finally use Softmax for activation and multi-class cross-entropy loss for optimization.

[0213] Its formula is as follows:

[0214] p = Softmax(MLP(Concat(User) emb Time emb ))), Formula 19

[0215]

[0216] Where p = [p0, ..., p C-1 ] is a probability distribution, where each element p i Let y represent the probability that a sample belongs to the i-th class; y = [y0, ..., y1] C-1 ] is the One-Hot representation of the sample label, where y represents the label when the sample belongs to category i. i =1, otherwise y i =0; C is the sample label.

[0217] During the optimization process, instead of using a weighted summation method to uniformly optimize the loss of different tasks, we use fixed variables to optimize tasks one, two, and three separately step by step.

[0218] Finally, the above-mentioned push model is used to generate personalized marketing strategies for each customer, reducing customer disturbance and reaching customers in the best way and at the most appropriate time.

[0219] The aforementioned information push model construction method obtains historical feature data and then uses this data to derive representation embedding data. This representation embedding data includes object representation data, push content representation data, push method representation data, and push time representation data. The object representation data and push content representation data are used as training data to train a neural network to obtain a push content sub-model. Similarly, the object representation data and push method representation data are used as training data to train a neural network to obtain a push method sub-model, and the object representation data and push time representation data are used as training data to train a neural network to obtain a push time sub-model. The information push model is then constructed by concatenating these three sub-models in parallel. This information push model is used to obtain the corresponding information push content category, information push method category, and information push time based on the object representation data corresponding to the object to be pushed to. In this way, by training neural networks based on object representation data and push content representation data, push method representation data, and push time representation data respectively, we can obtain various sub-models. Then, we can concatenate these sub-models in parallel to obtain the entire information push model. Compared with a non-concatenated single neural network model, the multi-task learning model, which is composed of multiple corresponding dimension sub-models concatenated in parallel, has better generalization ability and the training of the model is more accurate and efficient. This enables the trained information push model to improve the efficiency of generating online push information.

[0220] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0221] In one embodiment, such as Figure 9 As shown, an information push model construction device is provided. This device can employ software modules, hardware modules, or a combination of both as part of a computer device. Specifically, the device includes: a data acquisition module 902, a representation embedding data generation module 904, and a module generation module 906, wherein:

[0222] Data acquisition module 902 is used to acquire historical feature data;

[0223] The representation embedding data generation module 904 is used to obtain representation embedding data based on historical feature data. The representation embedding data includes object representation data, push content representation data, push method representation data, and push time representation data.

[0224] The model generation module 906 is used to train a neural network using object representation data and push content representation data as training data to obtain a push content sub-model; to train a neural network using object representation data and push method representation data as training data to obtain a push method sub-model; and to train a neural network using object representation data and push time representation data as training data to obtain a push time sub-model. Based on the push content sub-model, push method sub-model, and push time sub-model, the information push model is obtained by concatenating them in parallel. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time according to the object representation data corresponding to the object to be pushed.

[0225] The aforementioned information push model construction device acquires historical feature data and obtains representation embedding data based on this data. This representation embedding data includes object representation data, push content representation data, push method representation data, and push time representation data. The object representation data and push content representation data are used as training data to train a neural network to obtain a push content sub-model. Similarly, the object representation data and push method representation data are used as training data to train a neural network to obtain a push method sub-model, and the object representation data and push time representation data are used as training data to train a neural network to obtain a push time sub-model. The information push model is obtained by concatenating these three sub-models in parallel. This information push model is used to obtain the corresponding information push content category, information push method category, and information push time based on the object representation data corresponding to the object to be pushed to. In this way, by training neural networks based on object representation data and push content representation data, push method representation data, and push time representation data respectively, we can obtain various sub-models. Then, we can concatenate these sub-models in parallel to obtain the entire information push model. Compared with a non-concatenated single neural network model, the multi-task learning model, which is composed of multiple corresponding dimension sub-models concatenated in parallel, has better generalization ability and the training of the model is more accurate and efficient. This enables the trained information push model to improve the efficiency of generating online push information.

[0226] In one embodiment, the data acquisition module 902 is further configured to acquire initial historical feature data; perform correlation analysis on the feature data of each dimension in the initial historical feature data to obtain relevant dimension feature data; perform importance analysis on the relevant dimension feature data to determine the feature importance of the relevant dimension feature data; and sort the relevant dimension feature data according to the feature importance to generate historical feature data.

[0227] In one embodiment, the representation embedding data generation module 904 is further configured to: acquire target object feature data from historical feature data; acquire historical activity data corresponding to the target object feature data; normalize the object browsing dwell time in the historical activity data to obtain a time weight parameter; map each historical activity in the historical activity data to a corresponding historical activity vector, wherein the dimensions of the historical activity vectors corresponding to each historical activity are the same; and fuse the time weight parameter and the historical activity vectors corresponding to each historical activity to obtain representation embedding data, which is used to represent the correspondence between the push content, push method, and push time of each historical activity corresponding to the target object feature data.

[0228] In one embodiment, the representation embedding data generation module 904 is further configured to fuse the time weight parameters and the historical activity vectors corresponding to each historical activity to obtain initial embedded representation data; and to process the initial embedded representation data through a multilayer perceptron to obtain representation embedding data, wherein the dimension of the representation embedding data is lower than the dimension of the initial embedded representation data.

[0229] In one embodiment, the model generation module 906 is further configured to project object representation data through a multilayer perceptron to generate object representation vectors; project push content representation data through a multilayer perceptron to generate push content representation vectors, wherein the dimension of the push content representation vectors is the same as the dimension of the object representation vectors; obtain push weights based on the dot product of the push content representation vectors and the object representation vectors; and train the push content sub-model using the object representation vectors, push content representation vectors, and push weights as training sample data for the multilayer perceptron.

[0230] In one embodiment, the model generation module 906 is further configured to project the object representation data through a multilayer perceptron to generate an object representation vector; generate application push representation vectors, SMS push representation vectors, and telephone push representation vectors respectively based on the application push representation data, SMS push representation data, and telephone push representation data in the push method representation data; obtain the application push weight based on the dot product of the object representation vector and the application push representation vector; obtain the SMS push weight based on the dot product of the object representation vector and the SMS push representation vector; obtain the telephone push weight based on the dot product of the object representation vector and the telephone push representation vector; and obtain the push method sub-model by using the object representation vector, application push weight, SMS push weight, and telephone push weight as model training sample data.

[0231] In one embodiment, the model generation module 906 is further configured to: project object representation data through a multilayer perceptron to generate object representation vectors; divide the push time representation data into a preset number of sub-time period representation data; project the sub-time period representation data through a multilayer perceptron to generate sub-time period representation vectors; obtain the object idle time probability of the sub-time period corresponding to the sub-time period representation data according to the object representation vector and the sub-time period representation vector; and train the push time sub-model by using the object idle time probability and object representation vector of each sub-time period as input sample data of the multilayer perceptron.

[0232] Specific limitations regarding the information push model construction device can be found in the limitations of the information push model construction method described above, and will not be repeated here. Each module in the aforementioned information push model construction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0233] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements an information push model construction method. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0234] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0235] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0236] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0237] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps in the above method embodiments.

[0238] 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.

[0239] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0240] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0241] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for constructing an information push model, characterized in that, The method includes: Obtain historical feature data; Based on the historical feature data, representation embedding data is obtained, which includes object representation data, push content representation data, push method representation data, and push time representation data. The object representation data and the push content representation data are used as training data to train a neural network to obtain a push content sub-model. The object representation data and the push method representation data are used as training data to train a neural network to obtain a push method sub-model; The object representation data and the push time representation data are used as training data to train a neural network to obtain a push time sub-model. An information push model is obtained by concatenating the push content sub-model, the push method sub-model, and the push time sub-model in parallel. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time based on the object representation data corresponding to the object to be pushed. The step of obtaining the characterization embedding data based on the historical feature data includes: Obtain the target object feature data from the historical feature data; Obtain historical activity data corresponding to the feature data of the target object; The time spent browsing objects in the historical activity data is normalized to obtain the time weight parameter; Each historical activity in the historical activity data is mapped to a corresponding historical activity vector, and the dimensions of the historical activity vectors corresponding to each historical activity are all the same. The representation embedding data is obtained by fusing the time weight parameter and the historical activity vector corresponding to each historical activity. The representation embedding data is used to represent the correspondence between the push content, push method and push time of each historical activity corresponding to the target object feature data.

2. The method according to claim 1, characterized in that, After obtaining the historical feature data, the process also includes: Obtain initial historical feature data; Correlation analysis is performed on the feature data of each dimension in the initial historical feature data to obtain the relevant dimension feature data; The importance of the relevant dimension feature data is determined by performing an importance analysis on the relevant dimension feature data. The historical feature data is generated by sorting the relevant dimension feature data according to the importance of the features.

3. The method according to claim 1, characterized in that, The process of fusing the time weight parameters and the historical activity vectors corresponding to each historical activity to obtain the representation embedding data includes: Initial embedded representation data is obtained by fusing the time weight parameters and the historical activity vectors corresponding to each historical activity. The initial embedded representation data is processed by a multilayer perceptron to obtain representation embedded data, wherein the dimension of the representation embedded data is lower than that of the initial embedded representation data.

4. The method according to claim 1, characterized in that, The step of obtaining the characterization embedding data based on the historical feature data further includes: Embedded and represented user-side feature data from historical feature data to obtain user historical interaction information and feature statistics. Based on the historical interaction information of the current user, set the corresponding activity weights and perform weighted fusion to obtain the object representation parameters of the historical activity information of the current user in the user-side feature data. The object representation parameters are concatenated with feature statistics and used as input to a neural network model to generate lower-dimensional embedded representation data, which serves as the object representation data. The activity-side feature data is embedded and represented to obtain lower-dimensional push content representation data, push method representation data, and push time representation data, thus obtaining the representation embedding data.

5. The method according to claim 1, characterized in that, The step of using the object representation data and the push content representation data as training data to train a neural network to obtain a push content sub-model includes: The object representation data is projected through a multilayer perceptron to generate an object representation vector. The push content representation data is projected through a multilayer perceptron to generate a push content representation vector, and the dimension of the push content representation vector is the same as the dimension of the object representation vector. The push weight is obtained based on the dot product of the push content representation vector and the object representation vector; The object representation vector, the push content representation vector, and the push weight are used as training sample data for the multilayer perceptron to obtain the push content sub-model.

6. The method according to claim 1, characterized in that, The step of using the object representation data and the push method representation data as training data to train a neural network to obtain a push method sub-model includes: The object representation data is projected through a multilayer perceptron to generate an object representation vector. Based on the application push representation data, SMS push representation data, and telephone push representation data in the push method representation data, respectively generate application push representation vector, SMS push representation vector, and telephone push representation vector; The application push weight is obtained based on the dot product of the object representation vector and the application push representation vector. The SMS push weight is obtained based on the dot product of the object representation vector and the SMS push representation vector. The telephone push weight is obtained based on the dot product of the object representation vector and the telephone push representation vector. The object representation vector, application push weight, SMS push weight, and telephone push weight are used as model training sample data to obtain the push method sub-model.

7. The method according to claim 1, characterized in that, The step of using the object representation data and the push time representation data as training data to train a neural network to obtain a push time sub-model includes: The object representation data is projected through a multilayer perceptron to generate an object representation vector. Based on the push time representation data, a preset number of sub-time period representation data are obtained; The sub-time period representation data is projected through a multilayer perceptron to generate a sub-time period representation vector. The object idle time probability of the sub-time period corresponding to the sub-time period representation data is obtained by using the object representation vector and the sub-time period representation vector. The object idle time probability and the object representation vector in each sub-time period are used as input sample data for training a multilayer perceptron to obtain the push time sub-model.

8. An information push model construction device, characterized in that, The device includes: The data acquisition module is used to acquire historical feature data; The representation embedding data generation module is used to obtain representation embedding data based on the historical feature data. The representation embedding data includes object representation data, push content representation data, push method representation data, and push time representation data. The model generation module is used to train a neural network using the object representation data and the push content representation data as training data to obtain a push content sub-model; to train a neural network using the object representation data and the push method representation data as training data to obtain a push method sub-model; and to train a neural network using the object representation data and the push time representation data as training data to obtain a push time sub-model. Based on the push content sub-model, the push method sub-model, and the push time sub-model, an information push model is obtained by concatenating them in parallel. The information push model is used to obtain the corresponding information push content category, information push method category, and information push time according to the object representation data corresponding to the object to be pushed. The representation embedding data generation module is further configured to: acquire target object feature data from the historical feature data; acquire historical activity data corresponding to the target object feature data; normalize the object browsing dwell time in the historical activity data to obtain a time weight parameter; map each historical activity in the historical activity data to a corresponding historical activity vector, wherein the dimensions of the historical activity vectors corresponding to each historical activity are all the same; and fuse the time weight parameter and the historical activity vectors corresponding to each historical activity to obtain representation embedding data, wherein the representation embedding data is used to represent the correspondence between the push content, push method and push time of each historical activity corresponding to the target object feature data.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images