Information recommendation method and device

Abstract

The present application provides an information recommendation method and device, wherein the information recommendation method comprises: obtaining a behavior feature sequence of a target user and feature data of each information to be recommended, wherein the behavior feature sequence comprises a plurality of behavior feature data; respectively splicing the behavior feature sequence and the feature data of each information to be recommended, inputting the pre-trained model, and obtaining predicted recommendation data for each information to be recommended, wherein the pre-trained model is obtained by training based on the judgment of the temporal relationship of feature data in a sample sequence, and the sample sequence comprises: sample behavior feature data and sample recommendation information feature data collected in time sequence; according to the predicted recommendation data of each information to be recommended, determining the target recommendation information from each information to be recommended, and recommending the target recommendation information to the target user. This scheme can improve the accuracy of information recommendation.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and more particularly to an information recommendation method, an information recommendation device, a computing device, and a computer-readable storage medium. Background Art

[0002] In order to improve user experience, targeted information recommendations for different users are widely used.

[0003] In the related art, user characteristics are usually obtained based on the single behavior data of each user, using a pre-trained neural network model, and then the target information matching the user characteristics is determined, so as to achieve targeted information recommendation.

[0004] However, the user characteristics reflected by a single behavior data may not be accurate enough, resulting in the target information not matching the user well. Therefore, a more accurate solution is needed. Summary of the invention

[0005] In view of this, an embodiment of the present application provides an information recommendation method. The present application also relates to an information recommendation device, a computing device, and a computer-readable storage medium to solve the problem of inaccurate information recommendation in the prior art.

[0006] According to a first aspect of an embodiment of the present application, there is provided an information recommendation method, comprising:

[0007] Acquire a behavior feature sequence of a target user and feature data of each piece of information to be recommended, wherein the behavior feature sequence includes a plurality of behavior feature data;

[0008] The behavior feature sequence and the feature data of each piece of information to be recommended are respectively concatenated and input into a pre-trained model to obtain predicted recommendation data for each piece of information to be recommended, wherein the pre-trained model is obtained by judging and training the time series relationship of feature data in a sample sequence, and the sample sequence includes: sample behavior feature data and sample recommendation information feature data collected in time series;

[0009] According to the predicted recommendation data of each piece of information to be recommended, target recommended information is determined from each piece of information to be recommended, and the target recommended information is recommended to the target user.

[0010] According to a second aspect of an embodiment of the present application, there is provided an information recommendation device, including:

[0011] A feature data acquisition unit is configured to acquire a behavior feature sequence of a target user and feature data of each piece of information to be recommended, wherein the behavior feature sequence includes a plurality of behavior feature data;

[0012] The recommendation data prediction unit is configured to respectively concatenate the behavior feature sequence and the feature data of each piece of information to be recommended, input the concatenated data into a pre-trained model, and obtain predicted recommendation data for each piece of information to be recommended, wherein the pre-trained model is obtained by training based on the judgment of the time sequence relationship of the feature data in the sample sequence, and the sample sequence includes: sample behavior feature data and sample recommendation information feature data collected in time sequence;

[0013] The target information recommendation unit is configured to determine target recommended information from the information to be recommended according to the predicted recommendation data of the information to be recommended, and recommend the target recommended information to the target user.

[0014] According to a third aspect of an embodiment of the present application, a computing device is provided, comprising a memory, a processor, and computer instructions stored in the memory and executable on the processor, wherein the processor implements the steps of the information recommendation method when executing the computer instructions.

[0015] According to a fourth aspect of an embodiment of the present application, a computer-readable storage medium is provided, which stores computer instructions, and when the computer instructions are executed by a processor, the steps of the information recommendation method are implemented.

[0016] An embodiment of the present application realizes obtaining a behavior feature sequence of a target user and feature data of each information to be recommended, wherein the behavior feature sequence includes multiple behavior feature data; the behavior feature sequence and the feature data of each information to be recommended are spliced ​​respectively, input into a pre-trained model, and the predicted recommendation data for each information to be recommended is obtained, wherein the pre-trained model is obtained by training based on the judgment of the temporal relationship of feature data in a sample sequence, and the sample sequence includes: sample behavior feature data and sample recommendation information feature data collected in time sequence; according to the predicted recommendation data of each information to be recommended, the target recommendation information is determined from each information to be recommended, and the target recommendation information is recommended to the target user. Among them, the sample sequence includes: sample behavior feature data and sample recommendation information feature data collected in time sequence. Moreover, the pre-trained model is obtained by training based on the judgment of the temporal relationship of feature data in the sample sequence. Therefore, the pre-trained model can obtain the predicted recommendation data for the sample recommendation information spliced ​​by the behavior feature sequence based on the temporal relationship of each behavior feature data in the behavior feature sequence. In this way, the predicted recommendation data not only corresponds to the user characteristics reflected by each behavior feature data in the behavior feature sequence, but also corresponds to the user's temporal characteristics reflected by the temporal relationship of each behavior feature data. Therefore, compared with the target recommendation information determined by the user characteristics at a certain moment reflected by a single behavior data, more accurate predicted recommendation data can be obtained through richer user characteristics, thereby ensuring that the target recommendation information determined based on the predicted recommendation data of each information to be recommended is more in line with the user characteristics, achieving the effect of improving the accuracy of information recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a flow chart of an information recommendation method provided by an embodiment of the present application;

[0018] Figure 2 This is an example diagram of a behavior feature sequence in an information recommendation method provided in an embodiment of the present application;

[0019] Figure 3 is an example diagram of local feature data in an information recommendation method provided in an embodiment of the present application;

[0020] Figure 4 This is an example diagram of temporal feature data in an information recommendation method provided in an embodiment of the present application;

[0021] Figure 5 This is an example diagram of a processing flow of an information recommendation method provided by an embodiment of the present application;

[0022] Figure 6 This is an example diagram of an application scenario of an information recommendation method provided by an embodiment of the present application;

[0023] Figure 7is a structural diagram of an information recommendation device provided by an embodiment of the present application;

[0024] Figure 8 It is a structural block diagram of a computing device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0025] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.

[0026] The terms used in one or more embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit one or more embodiments of the present application. The singular forms of "a", "said" and "the" used in one or more embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in one or more embodiments of the present application refers to and includes any or all possible combinations of one or more associated listed items.

[0027] It should be understood that, although the terms first, second, etc. may be used to describe various information in one or more embodiments of the present application, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of one or more embodiments of the present application, the first may also be referred to as the second, and similarly, the second may also be referred to as the first. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0028] First, the terms involved in one or more embodiments of the present application are explained.

[0029] Pre-training: Instead of directly training for specific downstream tasks, the model is trained by using sample data for unsupervised training in advance.

[0030] Bidirectional Encoder Representations from Transformer (BERT) model: A bidirectional encoder representation model based on Transformer. The foundation of the BERT model is the Transformer model. Among them, "bidirectional" means that when the BERT model processes a piece of data, it can combine the context information of the data to obtain the information reflected by the data and the context information as a whole. For example, for a certain word being processed, the information of the previous word and the next word of the word can be combined to obtain the semantics expressed by these three words as a whole. Transformer model: A model architecture in natural language processing. The core layer of this model is a multi-head attention mechanism. The multi-head attention mechanism uses multiple attention mechanisms to perform separate calculations to obtain more levels of semantic information, and then splices and combines the results obtained by each attention mechanism to obtain the final result. The attention mechanism is a special structure embedded in the machine learning model to automatically learn and calculate the contribution of the input data to the output data, so as to select the input data based on the contribution and obtain the input data that can be used to achieve the task goal.

[0031] Deep Neural Networks (DNN): A neural network model for deep learning. The neural network layers inside a DNN can be divided into three categories: input layer, hidden layer, and output layer. Usually, the first layer of a DNN is the input layer, the last layer is the output layer, and the layers in between are all hidden layers. The layers are fully connected, that is, any neuron in the i-th layer must be connected to any neuron in the i+1-th layer.

[0032] Confidence: The confidence interval of a probability sample shows that the true value of a parameter of this probability sample has a certain probability of falling within the credibility range around the measurement result. Confidence is the confidence level, that is, the above-mentioned "certain probability".

[0033] Logistic Regression (LR) model: refers to the use of the predicted value of the linear regression model to approximate the logarithmic probability of the true label of the classification task. Although it is called regression, it is actually a classification model and is often used for binary classification.

[0034] eXtreme Gradient Boosting (XGBoost) model: an algorithm or engineering implementation based on the Gradient Boosting Decision Tree (GBDT). The basic idea of ​​XGBoost is the same as that of GBDT, but some optimizations have been made, such as the second-order derivative makes the loss function more accurate and the regularization term avoids tree overfitting. Among them, GBDT is an additive model based on the idea of ​​boosting ensemble. In the process of training GBDT, the forward distribution algorithm is used for greedy learning. In each iteration, a CART tree is learned to fit the residual of the prediction results of the previous t-1 CART trees and the true value of the training sample. CART is a binary tree, and each split will produce two child nodes. Boosting, also known as the boosting method, is a machine learning algorithm that can be used to reduce the bias in supervised learning. The idea of ​​boosting ensemble means that the model contains multiple weak learning classifiers used iteratively, and the results of each weak learning classifier are added to a strong learning classifier. In the process of adding, different weights are usually given according to the classification accuracy of each weak learning classifier to strengthen the classification of previously misclassified data, thereby achieving strong learning classification. A weak learning classifier is a classifier whose classification result is more accurate than random classification, but the difference between it and the true value is greater than the difference threshold; a strong learning classifier is a classifier whose classification result is less than the difference threshold, that is, the classification result is very close to the true value.

[0035] DeepFM: A model that replaces Wide in the Wide&Deep model with FM. FM refers to Factorization Machines, which is an improved version of the logistic regression model. It aims to solve the problem of difficult training of model parameters in scenarios with sparse data, and considers the second-order interaction of features, making up for the poor expressiveness of logistic regression. The Wide&Deep model is a combination of a linear model and a deep model. The linear model and the deep model are fused in a parallel structure. The linear part is intended to extract low-order interactive information, and the deep part extracts high-order interactive information.

[0036] In specific applications, the following three methods can usually be used to achieve the effect of recommending information to target users: The first method includes: the operator manually sets the recommendation rules for the information to be recommended based on his or her understanding of the business to which the information recommendation belongs, and then recommends the target recommended information in the information to be recommended to the target user according to the recommendation rules. The second method includes: based on the behavioral characteristic data of the target user, using machine learning models such as LR model, XGBoost model, etc., to obtain the target user's need for the recommended information, and recommending the target recommended information to the target user whose need reaches the degree threshold in the information to be recommended. The third method includes: based on the behavioral characteristic data of the target user, using deep learning such as DeepFM model, to obtain the target user's need for the recommended information, and recommending the target recommended information to the target user whose need reaches the degree threshold in the information to be recommended.

[0037] However, among the above three methods, the first method relies on manual settings, which is not convenient enough, and the rules set by manual subjective settings cannot adaptively adjust the recommendation results for different target users and user behaviors at different times. The machine learning model in the second method cannot obtain the high-order information implied in the behavioral feature data. Among them, high-order information is opposite to low-order information, which refers to implicit information that cannot be directly observed. Low-order information refers to intuitive information that can be directly observed. Taking the classic beer and diaper incident as an example, the low-order information includes: the trend of beer sales and the trend of diaper sales; and the high-order information is: the sales of beer and diapers rise at the same time every Friday night. The DeepFM model in the third method can only process the behavioral feature data at a certain moment separately, and cannot take advantage of the temporal nature of the target user's behavioral feature data.

[0038] In order to achieve more convenient and accurate information recommendation, in the present application, an information recommendation method is provided. The present application also involves an information recommendation device, a computing device, and a computer-readable storage medium, which are described in detail one by one in the following embodiments.

[0039] Figure 1 A flowchart of an information recommendation method provided according to an embodiment of the present application is shown, which specifically includes the following steps:

[0040] S102, obtaining a behavior feature sequence of a target user and feature data of each piece of information to be recommended, wherein the behavior feature sequence includes a plurality of behavior feature data.

[0041] In a specific application, the target user may include a user of a target service that provides information recommendation. In the process of the target user operating the target application used to realize the target service, the operation data generated can characterize the behavior characteristics of the target user. In addition, the operations performed by the target user on the target application have a time sequence relationship, for example, the time sequence relationship of first opening the mall interface and then selecting the product. In this way, the behavior characteristics represented by the operation data have a time sequence relationship, and multiple behavior characteristics can also form a behavior characteristic sequence. Therefore, the multiple behavior characteristic data included in the behavior characteristic sequence of the target user can be obtained based on the operation data of the target user in the target application. For example, the operation data may include: data characterizing the operations of game users in upgrading, wearing equipment, purchasing equipment, etc. in the game application, data characterizing the operations of video users in browsing, clicking, collecting, etc. in the video application, and so on. In addition, based on the operation data of the target user in the target application, the method of obtaining the behavior characteristic data may specifically include: collecting the operation data of the target user in the target application to obtain the behavior characteristic data; or extracting the characteristics of the collected operation data of the target user in the target application to obtain the behavior characteristic data, which are all reasonable and are not limited in this embodiment. Among them, the timing of collecting the operation data of the target user in the target application can be multiple. Exemplarily, the collection time may be when the target user performs a specified operation on the target application, so that the operation data can be collected in real time, and then the effect of real-time recommendation can be achieved, ensuring that the information recommendation is more accurate. Alternatively, the collection time may be when a preset period is reached, so that the historical operation data of the target user is collected. In one case, in order to ensure that the information recommendation is more accurate, information recommendation can be performed at a target time corresponding to the generation time of the historical operation data for the target user. For example, the generation time of the historical operation data is 20:00 on February 1, then the target time can be 20:00 on February 2 ± a preset offset. That is, the user's operations at the same time point are likely to be the same, and the impact of the timing relationship on the accuracy of the target recommendation information is also likely to be the same. Therefore, recommending information at a target time corresponding to the generation time of the historical operation data can ensure that the information recommendation is more accurate.

[0042] Moreover, for different target applications, that is, different application scenarios, the information to be recommended can be multiple. Exemplarily, the information to be recommended may include: game packages for users in game applications, videos, advertisements, music, etc. in video applications. For specific information to be recommended, it can be set according to application requirements, and this embodiment does not limit this. Therefore, the method for obtaining the characteristic data of each information to be recommended may include: searching for the characteristic data of each information to be recommended that is pre-stored in the database, or, performing feature extraction on each information to be recommended, respectively, to obtain the characteristic data of each information to be recommended, which are all reasonable, and this embodiment does not limit this. In addition, for different quantities of behavioral characteristic data at the same time, different methods can be used to obtain the behavioral characteristic sequence of the target user. In order to facilitate understanding and reasonable layout, it will be specifically explained in the form of optional embodiments later.

[0043] S104, concatenate the behavior feature sequence and the feature data of each information to be recommended respectively, input them into the pre-trained model, and obtain the predicted recommendation data for each information to be recommended, wherein the pre-trained model is obtained based on the judgment training of the time series relationship of the feature data in the sample sequence, and the sample sequence includes: sample behavior feature data and sample recommendation information feature data collected in time series.

[0044] In specific applications, the temporal relationship of feature data in a sample sequence may be of various kinds. Exemplarily, the temporal relationship of feature data in a sample sequence may be at least one of the temporal relationship between multiple sample behavior feature data and the temporal relationship between sample behavior feature data and sample recommendation information feature data. Furthermore, the method of obtaining a pre-trained model based on the judgment training of the temporal relationship of feature data in a sample sequence may be supervised training, unsupervised training or self-supervised training. Among them, in supervised training, the sample sequence may also include: a temporal relationship label. For example, a temporal relationship label between multiple sample behavior feature data and at least one of the temporal relationship labels between sample behavior feature data and the feature data of the recommendation information.

[0045] S106, determining target recommended information from each piece of information to be recommended according to the predicted recommendation data of each piece of information to be recommended, and recommending the target recommended information to the target user.

[0046] In a specific application, the predicted recommendation data may be the probability that each piece of information to be recommended belongs to the recommended information required by the target user, that is, the target recommended information. The information is sorted in descending order according to the probability, and the recommended information with the highest probability in the top designated position is selected as the target recommended information for recommendation. Among them, the designated position may be one or more. Alternatively, the predicted recommendation data may be whether each piece of information to be recommended belongs to the recommended information required by the target user, that is, the classification result of the target recommended information. The target information is determined based on the classification result, and the classification result is the information to be recommended that belongs to the recommended information required by the target user, that is, the target recommended information. Exemplarily, in a game application scenario, the information to be recommended is a game gift package, and the game gift package recommendation is to recommend a suitable gift package (including one or more props) to the player based on the player's status in the game, that is, the behavioral characteristic data, to help the player to pass the level.

[0047] In one embodiment of the present application, the sample sequence includes: sample behavior feature data and sample recommendation information feature data collected in time series. In addition, the pre-training model is trained based on the judgment of the time series relationship of the feature data in the sample sequence. Therefore, the pre-training model can obtain the predicted recommendation data for the sample recommendation information spliced ​​by the behavior feature sequence based on the time series relationship of each behavior feature data in the behavior feature sequence. In this way, the predicted recommendation data not only corresponds to the user characteristics reflected by each behavior feature data itself in the behavior feature sequence, but also corresponds to the user's temporal characteristics reflected by the time series relationship of each behavior feature data. Therefore, compared with the target recommendation information determined by the user characteristics at a certain moment reflected by a single behavior data, more accurate predicted recommendation data can be obtained through richer user characteristics, thereby ensuring that the target recommendation information determined based on the predicted recommendation data of each information to be recommended is more in line with the user characteristics, thereby achieving the effect of improving the accuracy of information recommendation.

[0048] In an optional implementation, the above-mentioned obtaining the behavior feature sequence of the target user may specifically include the following steps:

[0049] Collect multiple behavioral feature data of target users in time sequence;

[0050] Based on the plurality of behavior feature data, a behavior feature sequence is obtained.

[0051] This embodiment collects multiple behavior feature data of the target user in time sequence, and obtains a behavior sequence based on the multiple behavior feature data, so as to ensure that the behavior feature sequence can reflect the user's behavior characteristics in time sequence and ensure that the subsequent target recommendation information is more accurate. The time sequence may be the same as the time sequence used to obtain the sample behavior feature data. For example, the time sequence may be a time sequence generated by user operation data. Moreover, there may be multiple ways to obtain the behavior feature sequence based on multiple behavior feature data. Exemplarily, if the number of feature data at each moment in the time sequence is one, the multiple behavior feature data collected in time sequence may be arranged in time sequence to obtain a behavior feature sequence. Alternatively, in an optional embodiment, the number of behavior feature data collected at at least one moment in the time sequence is multiple;

[0052] Accordingly, the above-mentioned obtaining of the behavior feature sequence based on the multiple behavior feature data may specifically include the following steps:

[0053] dividing the behavior feature data with the same collection time among the multiple behavior feature data into an element in the behavior feature sequence;

[0054] Arrange the elements in the order of collection time to obtain the behavior feature sequence.

[0055] For example, the user's operations on the application program are sequential in time, for example, the player's behavior in the game is sequential in time. Therefore, constructing the player's temporal features, that is, the behavioral feature sequence, helps the subsequent pre-training model obtain high-level information from the behavioral feature sequence: the user's behavioral information that cannot be directly observed in the time period corresponding to the temporal sequence, so as to more fully analyze the player's characteristics, that is, the user's traits. For example, Figure 2An example diagram of a behavior feature sequence in an information recommendation method provided by an embodiment of the present application is shown as follows: the behavior of the player in the game includes multiple behaviors such as "upgrade", "wear equipment", "mall purchase", etc., and multiple behavior feature data "upgrade", "wear equipment", "mall purchase", etc., which are collected at 9:00, are divided into feature 1; multiple behavior feature data "upgrade", "wear equipment", "mall purchase", etc., which are collected at 9:30, are divided into feature 2, etc.; and so on, until the division is completed, multiple features, that is, multiple behavior feature data, are obtained. In this way, in the process of constructing the time series feature, each behavior feature data of the player at the same time is taken as a behavior feature data. Among them, the acquisition of each behavior feature data may include: obtaining at least one operation behavior of the player at the same time; digitizing each operation behavior: for example, for 9:00, the data corresponding to the upgrade behavior is 1, the data corresponding to the wear equipment behavior is 1, and the data corresponding to the mall purchase behavior is 1; and there are a total of five operation behaviors, and the latter two operation behaviors are specifically that the user does not perform an operation, then the behavior feature data at 9:00 is 11100. On this basis, the behavior feature data are arranged in chronological order to obtain a feature sequence such as "feature 1, feature 2, ..." Similarly, in the process of obtaining prediction recommendation data through the pre-trained model that has completed training, the sample behavior feature sequence can be obtained in a similar manner, except that the behavior feature data is adaptively sample behavior feature data.

[0056] In an optional implementation, the pre-trained model includes: a local feature extraction module and a global feature fusion module;

[0057] Accordingly, the above-mentioned steps of concatenating the behavior feature sequence and the feature data of each information to be recommended, inputting them into the pre-trained model, and obtaining the predicted recommendation data for each information to be recommended may specifically include the following steps:

[0058] The behavior feature sequence and the feature data of each piece of information to be recommended are respectively concatenated to obtain concatenated feature data;

[0059] Based on the behavior feature sequence in the spliced ​​feature data, using the local feature extraction module, extract the local feature data of the moment to which each behavior feature data belongs;

[0060] Based on each local feature data, a global feature fusion module is used to determine the temporal relationship of each local feature data, and the local feature data are fused according to the determination result to obtain temporal feature data. Based on the temporal feature data and the feature data of the information to be recommended in each spliced ​​feature data corresponding to the temporal feature data, predicted recommendation data for each information to be recommended is obtained.

[0061] In a specific application, based on the behavior feature sequence in the spliced ​​feature data, a local feature extraction module is used to extract the local feature data of the moment to which each behavior feature data belongs, which can be multiple. Exemplarily, the behavior feature sequence in the spliced ​​feature data can be input into the local feature extraction module to extract the local feature data of the moment to which each behavior feature data belongs, that is, direct extraction. Or, exemplarily, the data belonging to the discrete behavior feature data in each behavior feature data can be mapped and extracted. In order to facilitate understanding and reasonable layout, the second example will be specifically described in the form of an optional embodiment.

[0062] Moreover, based on each local feature data, a global feature fusion module is used to determine the temporal relationship of each local feature data, and each local feature data is fused according to the determination result. The method of obtaining temporal feature data may include: when the temporal relationship is learned through the NSP task, the global feature fusion module may add a temporal relationship label to each local feature data, and according to the temporal relationship label, each local feature data is fused to obtain the temporal feature data. Among them, the NSP task refers to determining whether local feature data B is the feature data of the next operation of local feature data A in the training of BERT. In order to facilitate understanding and reasonable layout, the method of obtaining temporal feature data will be specifically described in the form of an optional embodiment in the following.

[0063] This embodiment obtains the local feature data of the moment to which each behavioral feature data belongs, and then fuses each local feature data according to the temporal relationship of each local feature data to obtain the temporal feature data. This can ensure that the temporal feature data retains the user characteristics reflected by the local feature data while accurately reflecting the temporal relationship through fusion according to the temporal relationship, which is conducive to the subsequent acquisition of high-order information.

[0064] In an optional implementation, the plurality of behavior feature data include: discrete behavior feature data and continuous behavior feature data, wherein the continuous behavior feature data is a vector;

[0065] Accordingly, the above-mentioned method of extracting the local feature data of the moment to which each behavior feature data belongs based on the behavior feature sequence in the spliced ​​feature data by using the local feature extraction module may specifically include the following steps:

[0066] The behavioral feature sequence in the concatenated feature data is input into the local feature extraction module, so that the local feature extraction module performs the following steps:

[0067] According to the preset dimension, the discrete behavior feature data contained in the behavior feature sequence is mapped into a feature vector;

[0068] The feature vectors and continuous behavior feature data belonging to the same moment are spliced, and each splicing result is input into the local feature extraction module respectively to obtain the local feature data of the moment to which each behavior feature data belongs.

[0069] For example, Figure 3 An example diagram of local feature data in an information recommendation method provided by an embodiment of the present application is shown as follows: the player behavior features include two types of behavior feature data, namely, discrete behavior feature data and continuous behavior feature data. Before being input into the DNN model, the discrete behavior feature data is mapped into a vector of fixed dimension according to the idea of ​​DeepFM, and then the continuous behavior feature data and the vectorized discrete behavior feature data, namely, the feature vector, are spliced. Finally, each spliced ​​vector embedding, namely, the splicing result, is input into the DNN model respectively, and the ability of the DNN model to obtain deep-level information is used to obtain high-order information of the features at each moment, which is called local feature data. Discrete behavior feature data refers to behavior feature data whose values ​​are limited. For example, the value of "gender" feature data can only be "male" or "female". Similarly, the value of "purchase equipment" behavior feature data is "purchased" or "not purchased". Figure 3 The discrete behavior feature data in feature 1 shown are: "0.99" and "1300". Continuous behavior feature data means that the value of the feature data is infinite, for example, the value of the "height" feature data is infinite. Figure 3 The continuous behavior feature data in feature 1 is shown as: "0.0254, -0.0254, -0.1665, ...". In addition, the step of mapping the discrete behavior feature data into a feature vector may specifically include: mapping the discrete behavior feature data into a vector with a preset dimension in the form of dimension filling, for example Figure 3 The vector "level" in . The values ​​of dimension filling in the mapping process can be random, and the obtained vector elements can all be real numbers.

[0070] In an optional implementation, the above-mentioned global feature fusion module is used based on each local feature data to determine the temporal relationship of each local feature data, and each local feature data is fused according to the judgment result to obtain the temporal feature data, which may specifically include the following steps:

[0071] Each local feature data is input into the global feature fusion module, so that the global feature fusion module performs the following steps:

[0072] Determine the temporal relationship of each local feature data, and obtain a determination result of whether a temporal relationship exists;

[0073] The local feature data are weighted using the weights corresponding to the judgment results to obtain the temporal feature data.

[0074] For example, Figure 4 An example of temporal feature data in an information recommendation method provided by an embodiment of the present application is shown in the figure: after the DNN model obtains the local features, multiple local features with temporal characteristics, such as the local features at time point T1 and the local features at time point T2, are input into the pre-trained BERT model to make full use of the player's historical behavior information, so as to more accurately determine the player's current needs. The composition of each local feature can be referred to in the above Figure 3 Description of the embodiment. The BERT model uses the target local behavior feature data as the query, the local behavior features of the context of the target local behavior feature data as the key, and the similarity between the query and each key as the weight, and integrates the value of each local behavior feature in the context into the original value of the target local behavior feature. Among them, the context refers to the temporal relationship between the behavior feature data and the occurrence of each local behavior feature in the context. Specifically, the Attention mechanism takes the semantic vector representation of the target local behavior feature and each local behavior feature in the context as input, first obtains the Query vector representation of the target local behavior feature, the Key vector representation of each local behavior feature in the context, and the original Value representation of the target local behavior feature and each local behavior feature in the context through linear transformation, and then calculates the similarity between the Query vector and each Key vector as the weight, and weightedly fuses the Value vector of the target local behavior feature and the Value vector of each local behavior feature in the context as the output of Attention, that is: the enhanced semantic vector representation of the target local behavior feature.

[0075] In an optional implementation, using a global feature fusion module, based on the temporal feature data and the feature data of the information to be recommended in each concatenated feature data corresponding to the temporal feature data, obtaining predicted recommendation data for each information to be recommended may specifically include the following steps:

[0076] The global feature fusion module is used to process the temporal feature data and the feature data of the information to be recommended in each concatenated feature data corresponding to the temporal feature data as follows:

[0077] For each concatenated feature data corresponding to the temporal feature data, the concatenated feature data corresponding to the temporal feature data and the confidence of the target information to be recommended belonging to the temporal feature data are calculated to obtain the data to be predicted.

[0078] Exemplarily, the global feature fusion module calculates the spliced ​​feature data corresponding to the temporal feature data, the confidence of the target information to be recommended belonging to the temporal feature data, and obtains the specific steps of the data to be predicted, which may include: calculating the score matrix between the temporal feature data and the target information to be recommended in the corresponding spliced ​​feature data through a multi-head attention mechanism, and normalizing the score matrix to obtain the above confidence. The normalization process may be implemented by a softmax function.

[0079] In an optional embodiment, the pre-trained model includes a bidirectional language representation BERT model;

[0080] Accordingly, before obtaining the target user's behavior feature sequence and the feature data of each piece of information to be recommended, the information recommendation method provided in the embodiment of the present application may further include the following steps:

[0081] Collect sample behavior feature data and sample recommendation information feature data in time sequence to obtain a sample sequence;

[0082] Based on the sample sequence and the collection timing of each sample behavior feature data in the sample sequence, the bidirectional language representation BERT model is used to perform pre-training to determine the temporal relationship between each sample behavior feature data in the sample sequence. When the pre-training meets the preset training conditions, a pre-training model is obtained.

[0083] In specific applications, the DNN and BERT models are pre-trained together during the pre-training process. The original BERT model includes two pre-training tasks, MLM and NSP, during pre-training. The MLM task is to first mask a certain amount of words and then predict the original content. This task is not applicable to the player behavior feature data set. Therefore, the NSP task is retained to perform the first pre-training on the BERT model and the DNN model at the same time, that is, input two segments of player behavior features to determine whether the second segment is a subsequent behavior feature of the first segment. The model parameters of the BERT model and the DNN model are updated during the back propagation process. In addition, the above-mentioned first pre-training process refers to the process in which the pre-training model learns how to judge the temporal relationship between the behavior feature data of each sample in the sample sequence, and the pre-training model that has not been trained for downstream tasks is obtained. Therefore, the feature data of the sample recommendation information may not be used in the above-mentioned first pre-training, such as not using the feature data of the game gift package. On this basis, the second pre-training is performed. The second pre-training refers to the pre-training of the pre-trained model obtained by the first pre-training and not trained for downstream tasks before applying the pre-trained model. At this time, the sample sequence also contains the feature data of the sample recommendation information. The second pre-training can splice the feature data of the sample recommendation information and the corresponding sample behavior feature data, and use the spliced ​​feature data for training to achieve the learning of the mapping relationship between the sample behavior feature data and the post data of the sample recommendation information by the pre-trained model.

[0084] In an optional implementation, based on the sample sequence and the collection timing of each sample behavior feature data in the sample sequence, the bidirectional language representation BERT model is used to perform pre-training for judging the timing relationship between each sample behavior feature data in the sample sequence. When the pre-training reaches the preset training conditions, a pre-training model is obtained, which may specifically include the following steps:

[0085] According to the sample sequence and the collection time sequence of each sample behavior feature data in the sample sequence, a time sequence relationship label is marked, wherein the time sequence relationship label indicates whether each sample behavior feature data has a time sequence relationship of being collected one after another;

[0086] Using the bidirectional language representation BERT model, we can obtain the prediction and judgment results of the temporal relationship between the behavioral feature data of each sample in the sample sequence;

[0087] Input the time series relationship label and the prediction judgment result into the preset loss function to obtain the loss value;

[0088] If the loss value reaches the preset training condition, the pre-trained model is obtained.

[0089] In specific applications, the cross entropy loss function can be used to calculate the loss function value of each batch. The specific calculation method is as follows:

[0090]

[0091] Among them, N represents the total number of batches, y i Represents the temporal relationship label of the sample, y of the positive and negative samples i The values ​​are 1 and 0 respectively. The positive samples have a temporal relationship, while the negative samples do not have a temporal relationship. i Represents the probability that the model predicts the sample as a positive sample.

[0092] In addition, if the loss value does not meet the preset training conditions, the model parameters of the pre-trained model are adjusted, the pre-trained model is updated, and the bidirectional language representation model BERT is used to return to execute and obtain the prediction and judgment results of the temporal relationship between the behavioral feature data of each sample in the sample sequence.

[0093] For example, Figure 5 An example of a processing flow of an information recommendation method provided by an embodiment of the present application is shown in the figure: (1) Constructing temporal features. The behavior of players in the game is sequential in time. Constructing temporal features of players can help the model learn the long-term behavior information of players and thus more fully understand the characteristics of players. For example, feature 1 at time point 1, feature 2 at time point 2, ... and feature n at time point n. (2) Fusion of local features. Player behavior features include two types of features: discrete behavior feature data and continuous behavior feature data. Before inputting into the DNN model, the discrete behavior feature data is mapped into a vector of fixed dimension according to the idea of ​​DeepFM, and then the continuous behavior feature data and the vectorized discrete behavior feature data are spliced. Finally, the spliced ​​vector is input into the DNN model. The ability of the DNN model to obtain deep information is used to obtain high-order information of the features at each moment, which is called local features. The DNN model is a fully connected model. (3) Fusion of global features. After the DNN model obtains the local features, multiple local features with temporal sequence are input into the pre-trained BERT model to make full use of the player's historical behavior information, so as to more accurately judge the player's current needs.

[0094] The following combination Figure 6 , taking the application of the information recommendation method provided by this application as an example, the information recommendation method is further explained. Figure 6 An example diagram of an application scenario of an information recommendation method provided by an embodiment of the present application is shown, which specifically includes the following steps:

[0095] Exemplarily, the player's temporal characteristics, such as T1 behavior characteristics and T2 behavior characteristics, are spliced ​​with the gift package 1 characteristics as the input of the model (DNN+BERT), and the player's temporal characteristics, such as T1 behavior characteristics and T2 behavior characteristics, can be spliced ​​with the gift package 2 characteristics as the input of the model (DNN+BERT), and the model is used to calculate the probability that each gift package belongs to the gift package required by the player, for example, the probability of gift package 1, the probability of gift package 2, and so on, and then sorted in descending order according to the probability, and the gift packages with higher probabilities are selected for recommendation to the players. The model (DNN+BERT) is the pre-trained model provided by the solution of the embodiment of the present application, which is equivalent to providing an information recommendation method based on a pre-trained model. Exemplarily, in the game application scenario, the purpose of reducing the workload of game operators and improving the gift package recommendation effect can be achieved, and the accuracy rates are increased by 48% and 29% respectively compared with the manual design of game operators and machine learning methods.

[0096] Corresponding to the above method embodiment, the present application also provides an information recommendation device embodiment, Figure 7 FIG. 1 is a schematic diagram showing the structure of an information recommendation device provided by an embodiment of the present application. Figure 7 As shown, the device comprises:

[0097] The feature data acquisition unit 702 is configured to acquire a behavior feature sequence of a target user and feature data of each piece of information to be recommended, wherein the behavior feature sequence includes a plurality of behavior feature data;

[0098] The recommendation data prediction unit 704 is configured to respectively concatenate the behavior feature sequence and the feature data of each piece of information to be recommended, and input them into a pre-trained model to obtain predicted recommendation data for each piece of information to be recommended, wherein the pre-trained model is obtained by judging and training the time series relationship of feature data in a sample sequence, and the sample sequence includes: sample behavior feature data and sample recommendation information feature data collected in time series;

[0099] The target information recommendation unit 706 is configured to determine target recommended information from the information to be recommended according to the predicted recommendation data of the information to be recommended, and recommend the target recommended information to the target user.

[0100] In one embodiment of the present application, the sample sequence includes: sample behavior feature data and sample recommendation information feature data collected in time series. In addition, the pre-training model is trained based on the judgment of the time series relationship of the feature data in the sample sequence. Therefore, the pre-training model can obtain the predicted recommendation data for the sample recommendation information spliced ​​by the behavior feature sequence based on the time series relationship of each behavior feature data in the behavior feature sequence. In this way, the predicted recommendation data not only corresponds to the user characteristics reflected by each behavior feature data itself in the behavior feature sequence, but also corresponds to the user's temporal characteristics reflected by the time series relationship of each behavior feature data. Therefore, compared with the target recommendation information determined by the user characteristics at a certain moment reflected by a single behavior data, more accurate predicted recommendation data can be obtained through richer user characteristics, thereby ensuring that the target recommendation information determined based on the predicted recommendation data of each information to be recommended is more in line with the user characteristics, thereby achieving the effect of improving the accuracy of information recommendation.

[0101] In an optional implementation, the pre-trained model includes: a local feature extraction module and a global feature fusion module;

[0102] The recommended data prediction unit 704 is further configured to:

[0103] Separately concatenating the behavior feature sequence and the feature data of each piece of information to be recommended to obtain concatenated feature data;

[0104] Based on the behavior feature sequence in the spliced ​​feature data, using the local feature extraction module, extracting the local feature data at the time to which each behavior feature data belongs;

[0105] Based on each local feature data, the global feature fusion module is used to determine the temporal relationship of each local feature data, and the local feature data are fused according to the determination result to obtain temporal feature data, and based on the temporal feature data and the feature data of the information to be recommended in each spliced ​​feature data corresponding to the temporal feature data, predicted recommendation data for each information to be recommended is obtained.

[0106] In an optional implementation, the plurality of behavior feature data include: discrete behavior feature data and continuous behavior feature data, wherein the continuous behavior feature data is a vector;

[0107] The recommended data prediction unit 704 is further configured to:

[0108] The behavior feature sequence in the spliced ​​feature data is input into the local feature extraction module, so that the local feature extraction module performs the following steps:

[0109] According to a preset dimension, the discrete behavior feature data contained in the behavior feature sequence is mapped into a feature vector;

[0110] The feature vector and the continuous behavior feature data belonging to the same moment are spliced, and each splicing result is input into the local feature extraction module respectively to obtain the local feature data of the moment to which each behavior feature data belongs.

[0111] In an optional implementation, the recommended data prediction unit 704 is further configured to:

[0112] The local feature data are input into the global feature fusion module, so that the global feature fusion module performs the following steps:

[0113] Determine the time series relationship of each local feature data to obtain a determination result of whether a time series relationship exists;

[0114] The local feature data are weighted using the weight corresponding to the judgment result to obtain the temporal feature data.

[0115] In an optional implementation, the recommended data prediction unit 704 is further configured to:

[0116] The global feature fusion module performs the following processing on the temporal feature data and the feature data of the information to be recommended in each concatenated feature data corresponding to the temporal feature data:

[0117] For each concatenated feature data corresponding to the temporal feature data, the confidence of the target to-be-recommended information belonging to the concatenated feature data corresponding to the temporal feature data is calculated to obtain the data to be predicted.

[0118] In an optional embodiment, the pre-trained model includes a bidirectional language representation BERT model;

[0119] The device further comprises: a training unit configured to:

[0120] Collect sample behavior feature data and sample recommendation information feature data in time sequence to obtain a sample sequence;

[0121] Based on the sample sequence and the collection timing of each sample behavior feature data in the sample sequence, the bidirectional language representation BERT model is used to perform pre-training to determine the timing relationship between each sample behavior feature data in the sample sequence, and when the pre-training meets the preset training conditions, the pre-training model is obtained.

[0122] In an optional implementation, the training unit is further configured to:

[0123] According to the sample sequence and the collection timing of each sample behavior feature data in the sample sequence, a time sequence relationship label is marked, wherein the time sequence relationship label indicates whether each sample behavior feature data has a time sequence relationship of being collected successively;

[0124] Using the bidirectional language representation BERT model, a prediction result of the temporal relationship between the behavior feature data of each sample in the sample sequence is obtained;

[0125] Inputting the temporal relationship label and the prediction judgment result into a preset loss function to obtain a loss value;

[0126] If the loss value reaches the preset training condition, the pre-trained model is obtained.

[0127] In an optional implementation, the feature data acquisition unit 702 is further configured to:

[0128] Collecting the plurality of behavior feature data of the target user according to the time sequence;

[0129] Based on the plurality of behavior feature data, the behavior feature sequence is obtained.

[0130] In an optional implementation, the number of behavior feature data collected at at least one moment in the time series is multiple;

[0131] The characteristic data acquisition unit 702 is further configured to:

[0132] dividing the behavior feature data with the same collection time among the plurality of behavior feature data into an element in the behavior feature sequence;

[0133] The elements are arranged in the order of the collection moments to obtain the behavior feature sequence.

[0134] The above is a schematic scheme of an information recommendation device of this embodiment. It should be noted that the technical scheme of the information recommendation device and the technical scheme of the above information recommendation method belong to the same concept, and the details not described in detail in the technical scheme of the information recommendation device can be referred to the description of the technical scheme of the above information recommendation method.

[0135] Figure 8 The block diagram of a computing device according to an embodiment of the present application is shown. The components of the computing device 800 include but are not limited to a memory 810 and a processor 820. The processor 820 is connected to the memory 810 via a bus 830, and the database 850 is used to store data.

[0136] The computing device 800 also includes an access device 840 that enables the computing device 800 to communicate via one or more networks 860. Examples of these networks include a public switched telephone network (PSTN), a local area network (LAN), a wide area network (WAN), a personal area network (PAN), or a combination of communication networks such as the Internet. The access device 840 may include one or more of any type of network interface (e.g., a network interface card (NIC)) of wired or wireless, such as an IEEE 802.11 wireless local area network (WLAN) wireless interface, a world-wide interoperability for microwave access (Wi-MAX) interface, an Ethernet interface, a universal serial bus (USB) interface, a cellular network interface, a Bluetooth interface, a near field communication (NFC) interface, and the like.

[0137] In one embodiment of the present application, the above components of the computing device 800 and Figure 8 Other components not shown in the figure may also be connected to each other, for example, via a bus. It should be understood that Figure 8 The computing device structure block diagram shown is only for the purpose of illustration, and is not intended to limit the scope of the present application. Those skilled in the art may add or replace other components as needed.

[0138] The computing device 800 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., a tablet computer, a personal digital assistant, a laptop computer, a notebook computer, a netbook, etc.), a mobile phone (e.g., a smart phone), a wearable computing device (e.g., a smart watch, smart glasses, etc.), or other types of mobile devices, or a stationary computing device such as a desktop computer or PC. The computing device 800 may also be a mobile or stationary server.

[0139] When the processor 820 executes the computer instructions, the steps of the information recommendation method are implemented.

[0140] The above is a schematic scheme of a computing device of this embodiment. It should be noted that the technical scheme of the computing device and the technical scheme of the above information recommendation method belong to the same concept, and the details not described in detail in the technical scheme of the computing device can be referred to the description of the technical scheme of the above information recommendation method.

[0141] An embodiment of the present application further provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the steps of the information recommendation method as described above.

[0142] The above is a schematic scheme of a computer-readable storage medium of this embodiment. It should be noted that the technical scheme of the storage medium and the technical scheme of the above information recommendation method belong to the same concept, and the details not described in detail in the technical scheme of the storage medium can be referred to the description of the technical scheme of the above information recommendation method.

[0143] The above describes specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0144] The computer instructions include computer program codes, which may be in source code form, object code form, executable files or some intermediate forms, etc. The computer readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0145] It should be noted that, for the above-mentioned method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the present application is not limited by the described action sequence, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0146] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0147] The preferred embodiments of the present application disclosed above are only used to help explain the present application. The optional embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of the present application. The present application selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present application, so that those skilled in the art can understand and use the present application well. The present application is only limited by the claims and their full scope and equivalents.

Claims

1. An information recommendation method, characterized in that: include: Acquire a behavior feature sequence of a target user and feature data of each piece of information to be recommended, wherein the behavior feature sequence includes a plurality of behavior feature data; The behavior feature sequence and the feature data of each piece of information to be recommended are respectively concatenated and input into a pre-trained model to obtain predicted recommendation data for each piece of information to be recommended, wherein the pre-trained model is obtained by judging and training the time series relationship of feature data in a sample sequence, and the sample sequence includes: sample behavior feature data and sample recommendation information feature data collected in time series; Determining target recommended information from the information to be recommended according to the predicted recommendation data of the information to be recommended, and recommending the target recommended information to the target user; Wherein, the pre-training model includes: a local feature extraction module and a global feature fusion module; The step of respectively concatenating the behavior feature sequence and the feature data of each piece of information to be recommended, inputting the concatenated data into a pre-trained model, and obtaining predicted recommendation data for each piece of information to be recommended includes: Separately concatenating the behavior feature sequence and the feature data of each piece of information to be recommended to obtain concatenated feature data; Based on the behavior feature sequence in the spliced ​​feature data, using the local feature extraction module, extracting the local feature data at the time to which each behavior feature data belongs; Based on each local feature data, the global feature fusion module is used to determine the temporal relationship of each local feature data, and the local feature data are fused according to the determination result to obtain temporal feature data, and based on the temporal feature data and the feature data of the information to be recommended in each spliced ​​feature data corresponding to the temporal feature data, predicted recommendation data for each information to be recommended is obtained.

2. The method according to claim 1, characterized in that The plurality of behavior feature data include: discrete behavior feature data and continuous behavior feature data, wherein the continuous behavior feature data is a vector; The extracting of local feature data at the time of each behavior feature data based on the behavior feature sequence in the spliced ​​feature data using the local feature extraction module includes: The behavior feature sequence in the spliced ​​feature data is input into the local feature extraction module, so that the local feature extraction module performs the following steps: According to a preset dimension, the discrete behavior feature data contained in the behavior feature sequence is mapped into a feature vector; The feature vector and the continuous behavior feature data belonging to the same moment are spliced, and each splicing result is input into the local feature extraction module respectively to obtain the local feature data of the moment to which each behavior feature data belongs.

3. The method according to claim 1, characterized in that The method of determining the temporal relationship of each local feature data by using the global feature fusion module based on each local feature data, and fusing each local feature data according to the determination result to obtain temporal feature data includes: The local feature data are input into the global feature fusion module, so that the global feature fusion module performs the following steps: Determine the time series relationship of each local feature data to obtain a determination result of whether a time series relationship exists; The local feature data are weighted using the weight corresponding to the judgment result to obtain the temporal feature data.

4. The method according to any one of claims 1 to 3, characterized in that: Using the global feature fusion module, based on the temporal feature data and the feature data of the information to be recommended in each concatenated feature data corresponding to the temporal feature data, obtaining predicted recommendation data for each information to be recommended, including: The global feature fusion module is used to process the temporal feature data and the feature data of the information to be recommended in each concatenated feature data corresponding to the temporal feature data as follows: For each concatenated feature data corresponding to the temporal feature data, the confidence of the target to-be-recommended information belonging to the concatenated feature data corresponding to the temporal feature data is calculated to obtain the data to be predicted.

5. The method according to any one of claims 1 to 3, characterized in that: The pre-trained model includes a bidirectional language representation BERT model; Before obtaining the behavior feature sequence of the target user and the feature data of each piece of information to be recommended, the method further includes: Collect sample behavior feature data and sample recommendation information feature data in time sequence to obtain a sample sequence; Based on the sample sequence and the collection timing of each sample behavior feature data in the sample sequence, the bidirectional language representation BERT model is used to perform pre-training to determine the timing relationship between each sample behavior feature data in the sample sequence, and when the pre-training meets the preset training conditions, the pre-training model is obtained.

6. The method according to claim 5, characterized in that The method of performing pre-training for determining the temporal relationship between the behavior feature data of each sample in the sample sequence based on the sample sequence and the collection timing of the behavior feature data of each sample in the sample sequence by using the bidirectional language representation BERT model, and obtaining the pre-training model when the pre-training meets the preset training conditions, includes: According to the sample sequence and the collection timing of each sample behavior feature data in the sample sequence, a time sequence relationship label is marked, wherein the time sequence relationship label indicates whether each sample behavior feature data has a time sequence relationship of being collected successively; Using the bidirectional language representation BERT model, a prediction result of the temporal relationship between the behavior feature data of each sample in the sample sequence is obtained; Inputting the temporal relationship label and the prediction judgment result into a preset loss function to obtain a loss value; If the loss value reaches the preset training condition, the pre-trained model is obtained.

7. The method according to any one of claims 1 to 3 or 6, characterized in that The step of obtaining a target user's behavioral feature sequence includes: Collecting the plurality of behavior feature data of the target user according to the time sequence; Based on the plurality of behavior feature data, the behavior feature sequence is obtained.

8. The method according to claim 7, characterized in that The number of behavior feature data collected at least at one moment in the time series is multiple; The step of obtaining the behavior feature sequence based on the plurality of behavior feature data comprises: dividing the behavior feature data with the same collection time among the plurality of behavior feature data into an element in the behavior feature sequence; The elements are arranged in the order of the collection moments to obtain the behavior feature sequence.

9. An information recommendation device, characterized in that: include: A feature data acquisition unit is configured to acquire a behavior feature sequence of a target user and feature data of each piece of information to be recommended, wherein the behavior feature sequence includes a plurality of behavior feature data; The recommendation data prediction unit is configured to respectively concatenate the behavior feature sequence and the feature data of each piece of information to be recommended, input the concatenated data into a pre-trained model, and obtain predicted recommendation data for each piece of information to be recommended, wherein the pre-trained model is obtained by judging and training the temporal relationship of the feature data in a sample sequence, and the sample sequence includes: Sample behavior feature data and sample recommendation information feature data collected in time series; a target information recommendation unit, configured to determine target recommended information from each piece of information to be recommended according to the predicted recommendation data of each piece of information to be recommended, and recommend the target recommended information to the target user; Wherein, the recommended data prediction unit is also used for: The step of respectively concatenating the behavior feature sequence and the feature data of each piece of information to be recommended, inputting the concatenated data into a pre-trained model, and obtaining predicted recommendation data for each piece of information to be recommended includes: Separately concatenating the behavior feature sequence and the feature data of each piece of information to be recommended to obtain concatenated feature data; Based on the behavior feature sequence in the spliced ​​feature data, using a local feature extraction module, extracting local feature data at the time to which each behavior feature data belongs; Based on each local feature data, a global feature fusion module is used to determine the temporal relationship of the local feature data, and the local feature data are fused according to the determination result to obtain temporal feature data. Based on the temporal feature data and the feature data of the information to be recommended in each spliced ​​feature data corresponding to the temporal feature data, predicted recommendation data for each information to be recommended is obtained.

10. A computing device comprising a memory, a processor, and computer instructions stored in the memory and executable on the processor, characterized in that: When the processor executes the computer instructions, the steps of the method according to any one of claims 1 to 8 are implemented.

11. A computer-readable storage medium storing computer instructions, characterized in that: When the computer instruction is executed by a processor, the steps of the method described in any one of claims 1 to 8 are implemented.

Images