Information recommendation model determination method, information recommendation method and device

Abstract

The present application provides a method for determining an information recommendation model, an information recommendation method and a device, and the determination method includes: obtaining sample behavior information of sample users for pushed content; constructing a sample feature vector sequence for characterizing sample user behavior; inputting the 1st to K-1th sample feature vectors in the sample feature vector sequence into an encoder to obtain a sample encoding vector, and outputting a state vector from a preset network layer of the encoder; determining the predicted relevance corresponding to the Kth sample feature vector; training the encoder based on the predicted relevance and the actual relevance of the Kth sample feature vector; and determining the information recommendation model based on the encoder obtained by training. This process learns the characteristics of different behaviors of different sample users, and then predicts the probability of them performing a certain future behavior based on their historical behaviors, so that the information recommendation model has higher accuracy, and thus can improve the accuracy of the pushed information preferred by the target user.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an information recommendation model determination method, an information recommendation method and a device. Background Art

[0002] In the process of recommending information to target users, current information recommendation methods usually use the target user's previous historical selection information to determine the push information to be pushed. In addition, current information recommendation methods largely determine the push information to be pushed based on the adjacent historical selection information. The closer the historical selection information is, the greater the influence on the determined push information.

[0003] However, in actual applications, the adjacent historical selection information will not be considered as the push information that the target user wants to select at the moment. Therefore, when using the current information recommendation method to recommend information, the push information determined based on the target user's preferences is less accurate. Summary of the invention

[0004] In view of this, the purpose of the present application is to provide an information recommendation model determination method, an information recommendation method and an apparatus, so as to improve the accuracy of pushing information that meets the preferences of target users.

[0005] In a first aspect, an embodiment of the present application provides a method for determining an information recommendation model, the method for determining an information recommendation model comprising:

[0006] Obtain sample behavior information of sample users for pushed content;

[0007] Based on the sequence of the sample behavior information, construct a sample feature vector sequence for characterizing the sample user behavior; the sample feature vector sequence includes K sample feature vectors;

[0008] Inputting the 1st to K-1th sample feature vectors in the sample feature vector sequence into an encoder to obtain a sample encoding vector and a state vector output by a preset network layer of the encoder;

[0009] Based on the sample coding vector, the state vector and the Kth sample feature vector in the sample feature vector sequence, determining the prediction correlation corresponding to the Kth sample feature vector; wherein the prediction correlation is used to characterize the probability that the sample user performs the behavior corresponding to the Kth sample feature vector after performing the behaviors corresponding to the 1st to K-1th sample feature vectors;

[0010] Training the encoder based on the predicted correlation and the actual correlation of the K-th sample feature vector;

[0011] The information recommendation model is determined based on the trained encoder.

[0012] In an optional implementation, the encoder includes: a first neural network and a second neural network;

[0013] Then the step of inputting the 1st to K-1th sample feature vectors in the sample feature vector sequence into the encoder to obtain a sample encoding vector and a state vector output by a preset network layer of the encoder specifically includes:

[0014] Inputting the 1st to K-2th sample feature vectors in the sample feature vector sequence into a first neural network to obtain sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively; and after inputting the K-2th sample feature vector sequence into the first neural network, obtaining an intermediate state vector from a preset network layer of the first neural network;

[0015] Inputting the sample intermediate feature vectors, the intermediate state vector, and the K-1th sample feature vector corresponding to the 1st to K-2th sample feature vectors into a second neural network to obtain weights corresponding to the respective sample intermediate feature vectors; and obtaining the state vector from a preset network layer of the second neural network;

[0016] The sample encoding vector is obtained based on each of the sample intermediate feature vectors and the weights respectively corresponding to each of the sample intermediate feature vectors.

[0017] In an optional implementation, the first neural network includes: a first sub-network and a second sub-network;

[0018] Then, the step of inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the first neural network to obtain sample intermediate feature vectors corresponding to the first to K-2th sample feature vectors, respectively, specifically includes:

[0019] Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the first sub-network in positive order, to obtain first intermediate vectors corresponding to the first to K-2th sample feature vectors respectively;

[0020] Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the second sub-network in reverse order, to obtain second intermediate vectors corresponding to the first to K-2th sample feature vectors respectively;

[0021] For each of the 1st to K-2th sample feature vectors, a sample intermediate feature vector corresponding to the sample feature vector is generated based on the first intermediate vector and the second intermediate vector corresponding to the sample feature vector.

[0022] In an optional implementation manner, the second neural network includes: a third sub-network and a fourth sub-network;

[0023] Then, the first to the K-2th sample feature vectors correspond to the sample intermediate feature vectors, the intermediate state vector, and the K-1th sample feature vector respectively, and input into the second neural network to obtain the weights corresponding to each sample intermediate feature vector, specifically including:

[0024] Inputting the intermediate state vector and the K-1th sample feature vector into a third subnetwork to obtain a third intermediate vector corresponding to the K-1th sample feature vector;

[0025] The third intermediate vector and the 1st to K-2th sample feature vectors are respectively input into the fourth sub-network as sample intermediate feature vectors to obtain weights corresponding to each of the sample intermediate feature vectors.

[0026] In a second aspect, an embodiment of the present application provides an information recommendation method, comprising:

[0027] Obtain historical behavior information of the user to be pushed for multiple pushed contents;

[0028] Based on the sequence of the historical behavior information and at least one content to be pushed, constructing a feature vector sequence corresponding to each content to be pushed;

[0029] For each content to be pushed, input the feature vector sequence corresponding to the content to be pushed into the information recommendation model to obtain the relevance corresponding to the content to be pushed;

[0030] Determine at least one target push content from the plurality of contents to be pushed based on the magnitude of the relevance corresponding to each of the contents to be pushed, and push the target push content to the user to be pushed;

[0031] The information recommendation model is determined based on the method for determining the information recommendation model described in any one of the first aspects.

[0032] In a third aspect, an embodiment of the present application further provides an information recommendation model determination device, the information recommendation model determination device comprising: a first acquisition module, a first construction module, a first input module, a first determination module, a training module and a second determination module, wherein:

[0033] The first acquisition module is used to acquire sample behavior information of sample users for pushed content;

[0034] The first construction module is used to construct a sample feature vector sequence for characterizing the sample user behavior based on the sequence of the sample behavior information; the sample feature vector sequence includes K sample feature vectors;

[0035] The first input module is used to input the 1st to K-1th sample feature vectors in the sample feature vector sequence into the encoder to obtain a sample encoding vector and a state vector output by a preset network layer of the encoder;

[0036] The first determination module is used to determine the prediction correlation corresponding to the Kth sample feature vector based on the sample coding vector, the state vector and the Kth sample feature vector in the sample feature vector sequence; wherein the prediction correlation is used to characterize the probability that the sample user performs the behavior corresponding to the Kth sample feature vector after performing the behaviors corresponding to the 1st to K-1th sample feature vectors;

[0037] The training module is used to train the encoder based on the predicted correlation and the actual correlation of the K-th sample feature vector;

[0038] The second determination module is used to determine the information recommendation model based on the encoder obtained through training.

[0039] In an optional implementation, the encoder includes: a first neural network and a second neural network;

[0040] The first input module, when inputting the 1st to K-1th sample feature vectors in the sample feature vector sequence into the encoder to obtain the sample encoding vector and the state vector output by the preset network layer of the encoder, is used to:

[0041] Inputting the 1st to K-2th sample feature vectors in the sample feature vector sequence into a first neural network to obtain sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively; and after inputting the K-2th sample feature vector sequence into the first neural network, obtaining an intermediate state vector from a preset network layer of the first neural network;

[0042] Inputting the sample intermediate feature vectors, the intermediate state vector, and the K-1th sample feature vector corresponding to the 1st to K-2th sample feature vectors into a second neural network to obtain weights corresponding to the respective sample intermediate feature vectors; and obtaining the state vector from a preset network layer of the second neural network;

[0043] The sample encoding vector is obtained based on each of the sample intermediate feature vectors and the weights respectively corresponding to each of the sample intermediate feature vectors.

[0044] In an optional implementation, the first neural network includes: a first sub-network and a second sub-network;

[0045] The first input module, when inputting the 1st to K-2th sample feature vectors in the sample feature vector sequence into the first neural network to obtain sample intermediate feature vectors respectively corresponding to the 1st to K-2th sample feature vectors, is used to:

[0046] Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the first sub-network in positive order, to obtain first intermediate vectors corresponding to the first to K-2th sample feature vectors respectively;

[0047] Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the second sub-network in reverse order, to obtain second intermediate vectors corresponding to the first to K-2th sample feature vectors respectively;

[0048] For each of the 1st to K-2th sample feature vectors, a sample intermediate feature vector corresponding to the sample feature vector is generated based on the first intermediate vector and the second intermediate vector corresponding to the sample feature vector.

[0049] In an optional implementation manner, the second neural network includes: a third sub-network and a fourth sub-network;

[0050] The first input module, when inputting the sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors, the intermediate state vector, and the K-1th sample feature vector into the second neural network, and obtaining the weights corresponding to the respective sample intermediate feature vectors, is used to:

[0051] Inputting the intermediate state vector and the K-1th sample feature vector into a third subnetwork to obtain a third intermediate vector corresponding to the K-1th sample feature vector;

[0052] The third intermediate vector and the 1st to K-2th sample feature vectors are respectively input into the fourth sub-network as sample intermediate feature vectors to obtain weights corresponding to each of the sample intermediate feature vectors.

[0053] In a fourth aspect, an embodiment of the present application further provides an information recommendation device, the information recommendation device comprising: a second acquisition module, a second construction module, a second input module and a third determination module, wherein:

[0054] The second acquisition module is used to acquire historical behavior information of the user to be pushed for multiple pushed contents;

[0055] The second construction module is used to construct a feature vector sequence corresponding to each content to be pushed based on the sequence of the historical behavior information and at least one content to be pushed;

[0056] The second input module is used to input the feature vector sequence corresponding to each content to be pushed into the information recommendation model to obtain the relevance corresponding to the content to be pushed;

[0057] The third determination module is used to determine at least one target push content from the plurality of contents to be pushed based on the magnitude of the relevance corresponding to each of the contents to be pushed, and push the target push content to the user to be pushed;

[0058] The information recommendation model is determined based on the method for determining the information recommendation model described in any one of the first aspects.

[0059] In a fifth aspect, an embodiment of the present application further provides a computer device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, the steps in the first aspect or any possible implementation of the first aspect are performed;

[0060] Or execute the steps in the above-mentioned second aspect, or any possible implementation manner of the second aspect.

[0061] In a sixth aspect, an embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the first aspect or any possible implementation of the first aspect are executed;

[0062] Or execute the steps in the above-mentioned second aspect, or any possible implementation manner of the second aspect.

[0063] When determining the information recommendation model, the embodiment of the present application obtains the sample behavior information of the sample user for the pushed content, and then constructs a sample feature vector sequence for characterizing the sample user behavior based on the sequence of the sample behavior information; then inputs the 1st to K-1th sample feature vectors in the sample feature vector sequence into the encoder to obtain the sample coding vector and the state vector output by the preset network layer of the encoder; based on the sample coding vector, the state vector and the Kth sample feature vector in the sample feature vector sequence, determines the predicted relevance corresponding to the Kth sample feature vector, and trains the encoder based on the predicted relevance and the actual relevance of the Kth sample feature vector; determines the information recommendation model based on the encoder obtained by training. This process learns the characteristics of different behaviors of different sample users, and then predicts the probability of them making a certain future behavior based on their historical behaviors, so that the final information recommendation model has higher accuracy, and thus can improve the accuracy of the pushed information preferred by the target user.

[0064] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work. Figure 1 A flowchart of a method for determining an information recommendation model provided in an embodiment of the present application is shown;

[0066] Figure 2 A schematic diagram of the model structure provided in an embodiment of the present application is shown;

[0067] Figure 3 A flowchart showing a specific method for obtaining a sample coding vector and a state vector provided in an embodiment of the present application is shown;

[0068] Figure 4 A flowchart of an information recommendation method provided by an embodiment of the present application is shown;

[0069] Figure 5 A schematic diagram showing the structure of an information recommendation model determination device provided in an embodiment of the present application is shown;

[0070] Figure 6 A schematic diagram showing the structure of an information recommendation device provided in an embodiment of the present application is shown;

[0071] Figure 7 A schematic diagram of the structure of a computer device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0072] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0073] Considering that in the prior art, in the process of recommending information to the target user, the target user's previous historical selection information is usually used to determine the push information to be pushed. The closer the historical selection information is, the greater the influence on the determined push information. However, in actual application, the adjacent historical selection information will not be used by the target user as the push information that he / she wants to select at present. Therefore, using the current information recommendation method for information recommendation cannot accurately determine the push information that meets the preferences of the target user.

[0074] When determining the information recommendation model, the embodiment of the present application obtains the sample behavior information of the sample user for the pushed content, and then constructs a sample feature vector sequence for characterizing the sample user behavior based on the sequence of the sample behavior information; then inputs the 1st to K-1th sample feature vectors in the sample feature vector sequence into the encoder to obtain the sample coding vector and the state vector output by the preset network layer of the encoder; based on the sample coding vector, the state vector and the Kth sample feature vector in the sample feature vector sequence, determines the predicted relevance corresponding to the Kth sample feature vector, and trains the encoder based on the predicted relevance and the actual relevance of the Kth sample feature vector; determines the information recommendation model based on the encoder obtained by training. This process learns the characteristics of different behaviors of different sample users, and then predicts the probability of them making a certain future behavior based on their historical behaviors, so that the final information recommendation model has higher accuracy, and thus can improve the accuracy of the pushed information preferred by the target user.

[0075] The defects existing in the above solutions are the results obtained by the inventor after practice and careful research. Therefore, the process of discovering the above problems and the solutions proposed by this application for the above problems below should be the contributions made by the inventor to this application during the application process.

[0076] The technical solutions in the present application will be clearly and completely described below in conjunction with the drawings in the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the present application generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the present application for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0077] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0078] The execution subject of the information recommendation model determination method and the information recommendation method provided in the embodiments of the present disclosure is generally a computer device with certain computing capabilities, and the computer device includes, for example: a terminal device or a server or other processing device, and the terminal device can be a user equipment (User Equipment, UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (Personal Digital Assistant, PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. In some possible implementations, the information recommendation model determination method and the information recommendation method can be implemented by a processor calling a computer-readable instruction stored in a memory.

[0079] The following first describes the information recommendation model determination method provided by the embodiment of the present disclosure by taking a computer device as an example of an execution subject.

[0080] See also Figure 1-2 ,in Figure 1 A flowchart showing a method for determining an information recommendation model provided by an embodiment of the present application is shown; Figure 2 A schematic diagram showing the structure of an information recommendation model provided by an embodiment of the present application is shown;

[0081] The method comprises steps S101 to S103, wherein:

[0082] S101: Obtain sample behavior information of sample users for pushed content.

[0083] S102: constructing a sample feature vector sequence for characterizing the sample user behavior based on the sequence of the sample behavior information; the sample feature vector sequence includes K sample feature vectors.

[0084] S103: Inputting the 1st to K-1th sample feature vectors in the sample feature vector sequence into an encoder to obtain a sample encoding vector and a state vector output by a preset network layer of the encoder.

[0085] S104: Determine the prediction correlation corresponding to the Kth sample feature vector based on the sample coding vector, the state vector and the Kth sample feature vector in the sample feature vector sequence; wherein the prediction correlation is used to characterize the probability that the sample user will perform the behavior corresponding to the Kth sample feature vector after performing the behaviors corresponding to the 1st to K-1th sample feature vectors.

[0086] S105: Training the encoder based on the predicted correlation and the actual correlation of the Kth sample feature vector.

[0087] S106: Determine the information recommendation model based on the encoder obtained through training.

[0088] The above S101 to S106 are described in detail below.

[0089] 1. In the above S101, when the method provided in the embodiment of the present application is applied to different types of information recommendation fields, the information to be pushed is different and the corresponding sample behavior information is also different.

[0090] For example, the method is applied to the field of video push, and the content to be pushed is the video; the corresponding sample behavior information includes, for example: the sample user's viewing time, viewing duration, video operations (such as clicks, favorites, fast forward, comments), video type, information such as the actors and directors involved in the video, user ratings of the video, comprehensive ratings of the video by others, etc. In addition, it can also include the target audience of the video, and the sample user's gender, age, occupation, and other information.

[0091] For example, the type of video can be any one of the various types such as movies, TV series, variety shows, etc., among which the types of movies can be divided into science fiction films, romance films, horror films, etc. Similarly, TV series can be divided into family ethics dramas, costume dramas, idol dramas, etc., and variety shows can be divided into music, sports, adventure, etc.

[0092] The user's behavior information for the content to be pushed may be the choice of watching one of the above types of video information, as well as the time and duration of watching the video information, for example: Mr. A watched a science fiction movie at 5:10 pm for one hour; Mr. A watched a family ethics TV series at 8:15 pm for 40 minutes, etc.

[0093] The user's behavior information on the content to be pushed can also be the user's rating or comment information on the pushed content. For example, Mr. A rated the science fiction movie he watched as 4 points, and his comment information was "The plot is a bit far-fetched, but the actors' acting skills are still good!"; Mr. A rated the family ethics TV series he watched as 3 points, and his comment information was "The plot is too bloody!!!" and so on.

[0094] The user's behavior information on the content to be pushed may also include the user's own information, such as the user's age, gender, occupation, including some hobbies, region and other information, to facilitate subsequent processing.

[0095] For another example, the method is applied to the field of product push, and the content to be pushed is the product; the corresponding sample behavior information includes, for example: user operations on the product, such as clicking, adding to shopping cart, purchasing, deleting from shopping cart, returning, etc.; the sample behavior information can also include the attributes of the product itself, such as the price range of the product, the type of product, the purpose of the product, the specifications of the product, the name of the product, the target group of the product, etc.; the sample behavior information can also include attribute information of the sample user, such as the age of the sample user, purchase frequency, average purchase price of different types of products, annual purchase expenditure, monthly purchase expenditure, and types of frequently purchased products.

[0096] Specifically, specific settings are made according to the actual application field.

[0097] 2. In the above S102, a sample feature vector sequence is constructed according to the sequence of the sample behavior information obtained in the above step S101.

[0098] Here, the order of sample behavior information refers to the time sequence in which the user has performed specific actions on each content to be pushed. For example, if a user clicks to view product A, purchases product B, and adds product C to the shopping cart within a month, the generated sample feature vector sequence includes the following in the order of operation time:

[0099] The sample feature vector generated based on the click-to-view operation on product A;

[0100] The sample feature vector generated based on the purchase operation of product B;

[0101] And the sample feature vector generated based on the add-to-cart operation of product C.

[0102] For another example, the sample behavior information is sorted according to the time information in the sample behavior information, where the sample behavior information includes the sample user's behavior with respect to the content to be pushed, and feature information of the sample user and the content to be pushed. For example, at x:x on December 11, 2019, Mr. A watched xx variety show for xx minutes and xx seconds. The variety show was a competitive show, and Mr. A rated it 4 points and commented that it was "very good". Mr. A is 30 years old. Based on this information, the first sample feature vector is generated.

[0103] At x:x on December 12, 2019, Mr. A watched xx movie for xx minutes and xx seconds. The genre of the movie is action movie. Mr. A rated it 2 points and commented “average”. Mr. A is 30 years old. Based on this information, the second sample feature vector is generated.

[0104] The feature vector sequence finally generated includes: the first sample feature vector and the second sample feature vector.

[0105] It should be noted here that the above is only an example of generating a sample feature vector. In the actual process, the number of sample feature vectors included in the sample feature vector sequence can be specifically set according to actual needs.

[0106] In a specific implementation, each sample feature vector sequence is used as a sample; thus, multiple samples for determining an information recommendation model can be obtained.

[0107] Samples include: positive samples and negative samples.

[0108] Among them, the positive sample is directly constructed based on the sample behavior information. For example, the sample behavior information of a sample user at time t1 to time t10 includes: A1, A2, A3, A4, A5, ..., A9, A10 in sequence. When K is 5, the positive sample may include: the sample feature vector constructed based on A1, the sample feature vector constructed based on A2, the sample feature vector constructed based on A3, the sample feature vector constructed based on A4, and the sample feature vector constructed based on A5 in sequence.

[0109] Negative samples are constructed based on the sample behavior information of the sample user and pseudo-behavior information; wherein pseudo-behavior information refers to behavior information determined by the sample user's behavior that did not actually occur. The pseudo-behavior information can be, for example, randomly constructed, or sample behavior information of other sample users at any time, or sample behavior information of the sample user at other times. Negative samples can be obtained by replacing the last sample feature vector in the positive sample after the positive sample is constructed; or K-1 sample feature vectors can be determined first, and then the Kth sample feature vector is constructed based on the pseudo-behavior information, and then the K-1 sample feature vectors determined first and the Kth sample feature vector together constitute the K sample feature vectors in the negative sample.

[0110] For example, the sample behavior information of a sample user from time t1 to time t10 includes: A1, A2, A3, A4, A5, ..., A9, A10 in sequence. When K is 5, the negative samples may include: the sample feature vector constructed based on A1, the sample feature vector constructed based on A2, the sample feature vector constructed based on A3, the sample feature vector constructed based on A4, and the sample feature vector constructed based on M1 in sequence.

[0111] Wherein, M1 and M5 are different. Wherein, M5 may be sample behavior information of another sample user at a certain moment, or randomly constructed sample behavior information, or any sample behavior information of the sample user except A5.

[0112] In a possible implementation, only one sample feature vector sequence may be generated for a sample user. For example, sample behavior information of multiple sample users on pushed content at multiple different times may be first obtained, and then for each sample user, sample behavior information corresponding to consecutive K moments is selected from the sample behavior information of the sample user on pushed content at multiple different times, so as to generate a sample behavior feature vector sequence corresponding to the sample user based on the sample behavior information corresponding to the selected K moments. The selection may be a random selection, or the sample behavior information corresponding to the K moments closest to the current time may be determined as the sample behavior information used to generate the sample behavior feature vector sequence.

[0113] For example, the sample behavior information of a sample user at time t1 to time t10 includes: A1, A2, A3, A4, A5, ..., A9, A10 in sequence. When K is 5 and the sample is a positive sample, the 5 randomly determined sample behavior information include, for example, A1, A2, A3, A4, A5. When the sample behavior information corresponding to the 5 moments closest to the current time is determined as the sample behavior information used to generate the sample behavior feature vector sequence, the 5 determined sample behavior information include: A6, A7, A8, A9, A10. When the sample is a negative sample, the 5 randomly determined sample behavior information include, for example, A1, A2, A3, A4, A5.

[0114] In another possible implementation, multiple sample feature vector sequences may be generated for one sample user.

[0115] The sample behavior information of the sample user at time t1 to time t10 includes: A1, A2, A3, A4, A5, ..., A9, A10. When K is 5 and the sample is a positive sample, the determined sample behavior information may include, for example, 6 groups, namely:

[0116] A1, A2, A3, A4, A5;

[0117] A2, A3, A4, A5, A6;

[0118] A3, A4, A5, A6, A7;

[0119] A4, A5, A6, A7, A8;

[0120] A5, A6, A7, A8, A9;

[0121] A6, A7, A8, A9, A10.

[0122] Finally, based on the six groups of sample behavior information determined above, six sample feature vector sequences can be constructed.

[0123] 3: In the above S103, the sample coding vector is generated by the encoder encoding the 1st to K-1th sample feature vectors in the sample feature vector sequence.

[0124] The preset network layer of the encoder can output a state vector.

[0125] Among them, the encoding vector and the state vector both contain the features of the 1st to K-1th sample feature vectors.

[0126] Specifically, an embodiment of the present application provides a specific structure of an encoder, including: a first neural network and a second neural network.

[0127] Among them, the first neural network is used to extract the features contained in the 1st to K-2th sample feature vectors to obtain the sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively; the second neural network is used to extract the feature information of the second feature vector based on the influence information of the 1st to K-2th sample feature vectors on the K-1th sample feature vector (that is, the intermediate state vector), and based on the feature information, perform attention processing on the sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively to obtain the weights corresponding to the 1st to K-2th sample feature vectors respectively.

[0128] Finally, the feature information of the 1st to K-1th sample feature vectors can be extracted.

[0129] Specifically, see Figure 3 As shown, step S103 may specifically be:

[0130] S301: Input the 1st to K-2th sample feature vectors in the sample feature vector sequence into the first neural network to obtain sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively; and after inputting the K-2th sample feature vector sequence into the first neural network, obtain an intermediate state vector from a preset network layer of the first neural network.

[0131] Specifically: the first neural network includes: a first sub-network and a second sub-network;

[0132] Then the first to K-2th sample feature vectors in the sample feature vector sequence are input into the first neural network to obtain sample intermediate feature vectors corresponding to the first to K-2th sample feature vectors, respectively, including:

[0133] Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the first sub-network in positive order, to obtain first intermediate vectors corresponding to the first to K-2th sample feature vectors respectively;

[0134] Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the second sub-network in reverse order, to obtain second intermediate vectors corresponding to the first to K-2th sample feature vectors respectively;

[0135] For each of the 1st to K-2th sample feature vectors, a sample intermediate feature vector corresponding to the sample feature vector is generated based on the first intermediate vector and the second intermediate vector corresponding to the sample feature vector.

[0136] Here, for each sample feature vector from the 1st to the K-2th sample feature vectors, the first intermediate vector and the second intermediate vector corresponding to each sample feature vector may be concatenated or summed to obtain a sample intermediate feature vector corresponding to each sample feature vector.

[0137] The first sub-network includes any one of a recurrent neural network (RNN), a long short-term memory network (LSTM), and a gated recurrent unit (GRU).

[0138] The second sub-network also includes any one of a recurrent neural network (RNN), a long short-term memory network (LSTM), and a gated recurrent unit (GRU).

[0139] In a possible implementation, the first subnetwork and the second subnetwork may be different neural networks, so that different neural networks can extract different features from the 1st to K-2th sample feature vectors, so that the sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively have richer information in the 1st to K-2th sample feature vectors, thereby improving the model accuracy.

[0140] The intermediate state vector is obtained from the preset network layer of the first sub-network during the process of the first sub-network processing the K-2th sample feature vector; the preset network layer is, for example, any hidden layer in the first sub-network. At this time, since the 1st to K-2th sample feature vectors are input into the first sub-network in positive order, the characteristic relationship of the first sub-network, when extracting features from the sample feature vectors at a later time, will be affected by the feature extraction process of the sample feature vectors at an earlier time, that is, the intermediate state vector finally obtained actually contains the features in the 1st to K-2th sample feature vectors.

[0141] S302: Input the sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors, the intermediate state vector, and the K-1th sample feature vector into the second neural network to obtain weights corresponding to each sample intermediate feature vector; and obtain the state vector from a preset network layer of the second neural network.

[0142] Specifically, the second neural network includes: a third sub-network and a fourth sub-network;

[0143] Then, the first to the K-2th sample feature vectors correspond to the sample intermediate feature vectors, the intermediate state vector, and the K-1th sample feature vector respectively, and input into the second neural network to obtain the weights corresponding to each sample intermediate feature vector, specifically including:

[0144] The intermediate state vector and the K-1th sample feature vector are input into a third sub-network to obtain a third intermediate vector corresponding to the K-1th sample feature vector.

[0145] The third intermediate vector and the 1st to K-2th sample feature vectors are respectively input into the fourth sub-network as sample intermediate feature vectors to obtain weights corresponding to each of the sample intermediate feature vectors.

[0146] Here, the third sub-network includes, for example, any one of a recurrent neural network (RNN), a long short-term memory network (LSTM), and a gated recurrent unit (GRU).

[0147] In a possible implementation manner, the parameters of the third sub-network may be the same as the parameters of the first sub-network. Here, the parameters include, for example: network structure, specific parameters of each network layer, and the like.

[0148] The fourth sub-network, for example, can be a feed-forward neural network layer with a single hidden layer, and a computing module; when the sample intermediate feature vectors and the third intermediate vector corresponding to the 1st to K-2th sample feature vectors are respectively input into the fourth sub-network, each sample intermediate feature vector is input into the feed-forward neural network in sequence.

[0149] For any sample intermediate feature vector h among the sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors, input it and the third intermediate vector into the feedforward neural network together to obtain e(h,s)=Utanh(Vh+Ws);

[0150] Among them, U, V, W are the parameters of the feedforward neural network. Tanh(Vh+Ws) is the activation function.

[0151] Then the calculation module calculates the weights corresponding to the sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors based on the following formula:

[0152]

[0153] Among them, ai is the weight corresponding to the i-th sample feature vector among the 1st to K-2th sample feature vectors. i represents the sample intermediate feature vector corresponding to the i-th sample feature vector; s K-1 Represents the third intermediate vector corresponding to the K-1th sample feature vector.

[0154] Finally, the weights corresponding to the sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors are obtained.

[0155] S303: Obtain the sample encoding vector based on each of the sample intermediate feature vectors and the weights corresponding to each of the sample intermediate feature vectors.

[0156] Here, for example, based on the weights corresponding to the sample intermediate feature vectors, the sample encoding vector may be obtained by performing weighted summation on the sample intermediate feature vectors.

[0157] For example, the specific operation process is as follows:

[0158] Assuming that there are 10 sample feature vectors, the first to ninth sample feature vectors are used as input data of the encoder, and the predicted relevance of the tenth sample feature vector is determined.

[0159] First, the 1st to 8th sample feature vectors are input into the first subnetwork of the first neural network in positive order to obtain the first intermediate vectors corresponding to the 1st to 8th sample feature vectors, and the 1st to 8th sample feature vectors are input into the second subnetwork of the first neural network in reverse order to obtain the second intermediate vectors corresponding to the 1st to 8th sample feature vectors. For the 1st sample feature vector, the first intermediate vector and the second intermediate vector corresponding to the 1st sample feature vector are summed to obtain the sample intermediate feature vector corresponding to the 1st sample feature vector. And this process is performed sequentially for the 2nd, 3rd, ..., 8th sample feature vectors to obtain the sample intermediate feature vectors corresponding to the 2nd to 8th sample feature vectors.

[0160] After the eighth sample feature vector sequence is input into the first neural network, an intermediate state vector is obtained from a preset network layer of the first neural network.

[0161] Afterwards, the intermediate state vector and the 9th sample feature vector are input into the third sub-network in the second neural network to obtain a third intermediate vector corresponding to the 9 sample feature vectors; the third intermediate vector and the 1st to 8th sample feature vectors are respectively input into the fourth sub-network corresponding to the sample intermediate feature vectors, and the weights corresponding to the 1st to 8th sample intermediate feature vectors can be obtained.

[0162] Based on the weights corresponding to the intermediate feature vectors of the 1st to the 8th samples, the intermediate feature vectors of the 1st to the 8th samples are weighted summed to obtain encoding vectors for encoding the feature vectors of the 1st to the 9th samples, and a state vector is output from the preset network layer of the third subnetwork for subsequent operations.

[0163] Fourth: In the above S104, for example, a preset gating function may be used to perform a gating function operation on the sample coding vector, the state vector, and the third sample feature vector to obtain a predicted correlation corresponding to the third sample feature vector.

[0164] The specific gate function can be set according to actual needs and will not be described in detail here.

[0165] Five: In the above S105 and S106, the encoder is trained based on the predicted correlation obtained in step S104 and the actual correlation of the Kth sample feature vector.

[0166] The actual correlation is the probability of generating the Kth sample feature information after generating the 1st to K-1th sample feature information among the K sample feature information.

[0167] Among them, in the positive sample, the actual correlation of the Kth sample feature vector is 1; in the negative sample, the actual correlation of the Kth sample feature vector is 0.

[0168] Based on the predicted correlation and the actual correlation of the Kth sample feature vector, when training the encoder, the cross entropy loss of correlation can be determined according to the predicted correlation and the actual correlation corresponding to each sample feature vector sequence, and then the parameters of the encoder are adjusted based on the cross entropy loss.

[0169] After multiple rounds of adjustments to the encoder parameters, when the encoder parameters no longer change or the cross entropy no longer decreases, that is, when the encoder converges, an information recommendation model is determined based on the encoder.

[0170] Here, it should be noted that in the information recommendation model, in addition to the encoder, there is also a calculation unit. The encoder is used to generate sample intermediate feature vectors and weights corresponding to the 1st to K-2th sample feature vectors, respectively, and the calculation unit is used to generate the predicted relevance based on the sample intermediate feature vectors and weights corresponding to the 1st to K-2th sample feature vectors, respectively, the state vector output by the encoder, and the Kth sample feature vector.

[0171] When determining the information recommendation model, the embodiment of the present application obtains the sample behavior information of the sample user for the pushed content, and then constructs a sample feature vector sequence for characterizing the sample user behavior based on the sequence of the sample behavior information; then inputs the 1st to K-1th sample feature vectors in the sample feature vector sequence into the encoder to obtain the sample coding vector and the state vector output by the preset network layer of the encoder; based on the sample coding vector, the state vector and the Kth sample feature vector in the sample feature vector sequence, determines the predicted relevance corresponding to the Kth sample feature vector, and trains the encoder based on the predicted relevance and the actual relevance of the Kth sample feature vector; determines the information recommendation model based on the encoder obtained by training. This process learns the characteristics of different behaviors of different sample users, and then predicts the probability of them making a certain future behavior based on their historical behaviors, so that the final information recommendation model has higher accuracy, and thus can improve the accuracy of the pushed information preferred by the target user.

[0172] In addition, in the second neural network of the embodiment of the present application, the K-1th sample feature vector is utilized to determine the weights of the first K-2 sample feature vectors, so that the model can learn the "skip level" effect of the first K-2 sample feature vectors on the Kth sample feature vector, thereby further improving the accuracy of the model.

[0173] See also Figure 4 FIG. 4 is a flowchart of an information recommendation method provided in Embodiment 2 of the present application, wherein the method includes steps S401 to S404, wherein:

[0174] S401: Obtain historical behavior information of the user to be pushed for multiple pushed contents.

[0175] S402: Based on the sequence of the historical behavior information and at least one content to be pushed, construct a feature vector sequence corresponding to each content to be pushed.

[0176] S403: For each content to be pushed, input the feature vector sequence corresponding to the content to be pushed into the information recommendation model to obtain the relevance corresponding to the content to be pushed.

[0177] S404: determining at least one target push content from the plurality of contents to be pushed based on the relevances of the contents to be pushed, and pushing the target push content to the user to be pushed;

[0178] The information recommendation model is determined based on the method for determining the information recommendation model described in any one of the first embodiments.

[0179] The above S401 to S404 are described in detail below.

[0180] The specific implementation of the above steps S401 and S402 is similar to that of the above steps S101 and S102, and will not be repeated here.

[0181] It should be noted here that the historical behavior information of the user to be pushed for multiple pushed contents is determined based on the current moment. For example, if you want to construct a feature vector sequence including K feature vectors, you need to select the K-1 behavior information closest to the current moment to determine K-1 feature vectors, and based on the M content to be pushed, determine the feature vectors of the M content to be pushed. For each content to be pushed, use the feature vector of the content to be pushed as the Kth feature vector in the feature vector sequence, and form a feature vector sequence with the determined K-1 feature vectors. Finally, you get a feature vector sequence of M. Where M is an integer greater than 0.

[0182] Then, the M feature vector sequences are sequentially input into the information recommendation model to obtain the relevances corresponding to the M contents to be pushed.

[0183] Then, according to the relevances respectively corresponding to the M contents to be pushed, at least one target push content is determined from the M contents to be pushed, and the target push content is pushed to the user to be pushed.

[0184] Exemplarily, the content to be pushed with the greatest relevance may be used as the target pushed content.

[0185] In addition, multiple target push contents may be determined from the M contents to be pushed in descending order of relevance, and each target push content may be pushed to the user to be pushed in sequence according to the relevance corresponding to each target push content.

[0186] Reference Figure 5 As shown, it is a schematic diagram of an information recommendation model determination device provided in an embodiment of the present application, and the information recommendation model determination device includes: a first acquisition module 51, a first construction module 52, a first input module 53, a first determination module 54, a training module 55 and a second determination module 56, wherein:

[0187] A first acquisition module 51 is used to acquire sample behavior information of sample users for pushed content;

[0188] A first construction module 52 is used to construct a sample feature vector sequence for characterizing the sample user behavior based on the sequence of the sample behavior information; the sample feature vector sequence includes K sample feature vectors;

[0189] A first input module 53 is used to input the 1st to K-1th sample feature vectors in the sample feature vector sequence into an encoder to obtain a sample encoding vector and a state vector output by a preset network layer of the encoder;

[0190] A first determination module 54 is used to determine the prediction correlation corresponding to the Kth sample feature vector based on the sample coding vector, the state vector and the Kth sample feature vector in the sample feature vector sequence; wherein the prediction correlation is used to characterize the probability that the sample user performs the behavior corresponding to the Kth sample feature vector after performing the behaviors corresponding to the 1st to K-1th sample feature vectors;

[0191] A training module 55, configured to train the encoder based on the predicted relevance and the actual relevance of the K-th sample feature vector;

[0192] The second determination module 56 is used to determine the information recommendation model based on the encoder obtained through training.

[0193] In a possible implementation, the encoder includes: a first neural network and a second neural network;

[0194] The first input module 53, when inputting the 1st to K-1th sample feature vectors in the sample feature vector sequence into the encoder to obtain the sample encoding vector and the state vector output by the preset network layer of the encoder, is used to:

[0195] Inputting the 1st to K-2th sample feature vectors in the sample feature vector sequence into a first neural network to obtain sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively; and after inputting the K-2th sample feature vector sequence into the first neural network, obtaining an intermediate state vector from a preset network layer of the first neural network;

[0196] Inputting the sample intermediate feature vectors, the intermediate state vector, and the K-1th sample feature vector corresponding to the 1st to K-2th sample feature vectors into a second neural network to obtain weights corresponding to the respective sample intermediate feature vectors; and obtaining the state vector from a preset network layer of the second neural network;

[0197] The sample encoding vector is obtained based on each of the sample intermediate feature vectors and the weights respectively corresponding to each of the sample intermediate feature vectors.

[0198] In a possible implementation manner, the first neural network includes: a first sub-network and a second sub-network;

[0199] The first input module 53, when inputting the 1st to K-2th sample feature vectors in the sample feature vector sequence into the first neural network to obtain sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively, is used to:

[0200] Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the first sub-network in positive order, to obtain first intermediate vectors corresponding to the first to K-2th sample feature vectors respectively;

[0201] Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the second sub-network in reverse order, to obtain second intermediate vectors corresponding to the first to K-2th sample feature vectors respectively;

[0202] For each of the 1st to K-2th sample feature vectors, a sample intermediate feature vector corresponding to the sample feature vector is generated based on the first intermediate vector and the second intermediate vector corresponding to the sample feature vector.

[0203] In a possible implementation manner, the second neural network includes: a third sub-network and a fourth sub-network;

[0204] The first input module 53, when inputting the sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors, the intermediate state vector, and the K-1th sample feature vector into the second neural network, and obtaining the weights corresponding to the respective sample intermediate feature vectors, is used to:

[0205] Inputting the intermediate state vector and the K-1th sample feature vector into a third subnetwork to obtain a third intermediate vector corresponding to the K-1th sample feature vector;

[0206] The third intermediate vector and the 1st to K-2th sample feature vectors are respectively input into the fourth sub-network as sample intermediate feature vectors to obtain weights corresponding to each of the sample intermediate feature vectors.

[0207] Reference Figure 6As shown, an information recommendation device provided in an embodiment of the present application includes: a second acquisition module 61, a second construction module 62, a second input module 63 and a third determination module 64, wherein:

[0208] The second acquisition module 61 is used to acquire historical behavior information of the user to be pushed for multiple pushed contents;

[0209] A second construction module 62 is used to construct a feature vector sequence corresponding to each content to be pushed based on the sequence of the historical behavior information and at least one content to be pushed;

[0210] The second input module 63 is used to input the feature vector sequence corresponding to each content to be pushed into the information recommendation model to obtain the relevance corresponding to the content to be pushed;

[0211] A third determination module 64 is used to determine at least one target push content from the plurality of contents to be pushed based on the magnitude of the relevance corresponding to each of the contents to be pushed, and push the target push content to the user to be pushed;

[0212] The information recommendation model is determined based on any one of the above-mentioned methods for determining the information recommendation model.

[0213] The present disclosure also provides a computer device 10, such as Figure 7 As shown, it is a schematic diagram of the structure of a computer device 10 provided in an embodiment of the present disclosure, including: a processor 11, a memory 12, and a bus 13. The memory 12 stores machine-readable instructions executable by the processor 11. When the computer device 10 is running, the processor 11 communicates with the memory 12 through the bus 13. When the machine-readable instructions are executed by the processor 11, the steps of the information recommendation model determination method provided in the embodiment of the present disclosure are executed; or the steps of the information recommendation method provided in the embodiment of the present application are executed.

[0214] An embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the computer program executes the steps of the information recommendation model determination method described in the above method embodiment, or executes the steps of the information recommendation method provided in the embodiment of the present application.

[0215] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here. In the several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0216] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0217] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0218] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium that is executable by a processor. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0219] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present application, which are used to illustrate the technical solutions of the present application, rather than to limit them. The protection scope of the present application is not limited thereto. Although the present application is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed in the present application, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A method for determining an information recommendation model, characterized in that: The information recommendation model determination method includes: Obtain sample behavior information of sample users for pushed content; Based on the sequence of the sample behavior information, construct a sample feature vector sequence for characterizing the sample user behavior; the sample feature vector sequence includes K sample feature vectors; Inputting the 1st to K-1th sample feature vectors in the sample feature vector sequence into an encoder to obtain a sample encoding vector and a state vector output by a preset network layer of the encoder; Based on the sample coding vector, the state vector and the Kth sample feature vector in the sample feature vector sequence, determining the prediction correlation corresponding to the Kth sample feature vector; wherein the prediction correlation is used to characterize the probability that the sample user performs the behavior corresponding to the Kth sample feature vector after performing the behaviors corresponding to the 1st to K-1th sample feature vectors; Training the encoder based on the predicted correlation and the actual correlation of the K-th sample feature vector; The information recommendation model is determined based on the trained encoder.

2. The determination method according to claim 1, characterized in that: The encoder includes: a first neural network and a second neural network; Then the step of inputting the 1st to K-1th sample feature vectors in the sample feature vector sequence into the encoder to obtain a sample encoding vector and a state vector output by a preset network layer of the encoder specifically includes: Inputting the 1st to K-2th sample feature vectors in the sample feature vector sequence into a first neural network to obtain sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively; and after inputting the K-2th sample feature vector sequence into the first neural network, obtaining an intermediate state vector from a preset network layer of the first neural network; Inputting the sample intermediate feature vectors, the intermediate state vector, and the K-1th sample feature vector corresponding to the 1st to K-2th sample feature vectors into a second neural network to obtain weights corresponding to the respective sample intermediate feature vectors; and obtaining the state vector from a preset network layer of the second neural network; The sample encoding vector is obtained based on each of the sample intermediate feature vectors and the weights respectively corresponding to each of the sample intermediate feature vectors.

3. The determination method according to claim 2, characterized in that: The first neural network includes: a first sub-network and a second sub-network; Then, the step of inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the first neural network to obtain sample intermediate feature vectors corresponding to the first to K-2th sample feature vectors, respectively, specifically includes: Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the first sub-network in positive order, to obtain first intermediate vectors corresponding to the first to K-2th sample feature vectors respectively; Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the second sub-network in reverse order, to obtain second intermediate vectors corresponding to the first to K-2th sample feature vectors respectively; For each of the 1st to K-2th sample feature vectors, a sample intermediate feature vector corresponding to the sample feature vector is generated based on the first intermediate vector and the second intermediate vector corresponding to the sample feature vector.

4. The determination method according to claim 2, characterized in that: The second neural network includes: a third sub-network and a fourth sub-network; Then, the first to the K-2th sample feature vectors correspond to the sample intermediate feature vectors, the intermediate state vector, and the K-1th sample feature vector respectively, and input into the second neural network to obtain the weights corresponding to each sample intermediate feature vector, specifically including: Inputting the intermediate state vector and the K-1th sample feature vector into a third subnetwork to obtain a third intermediate vector corresponding to the K-1th sample feature vector; The third intermediate vector and the 1st to K-2th sample feature vectors are respectively input into the fourth sub-network as sample intermediate feature vectors to obtain weights corresponding to each of the sample intermediate feature vectors.

5. An information recommendation method, characterized in that: include: Obtain historical behavior information of the user to be pushed for multiple pushed contents; Based on the sequence of the historical behavior information and at least one content to be pushed, constructing a feature vector sequence corresponding to each content to be pushed; For each content to be pushed, input the feature vector sequence corresponding to the content to be pushed into the information recommendation model to obtain the relevance corresponding to the content to be pushed; Determine at least one target push content from the plurality of contents to be pushed based on the magnitude of the relevance corresponding to each of the contents to be pushed, and push the target push content to the user to be pushed; The information recommendation model is determined based on the method for determining the information recommendation model according to any one of claims 1 to 4.

6. An information recommendation model determination device, characterized in that: The information recommendation model determination device comprises: The first acquisition module is used to obtain sample behavior information of sample users for pushed content; A first construction module is used to construct a sample feature vector sequence for characterizing the sample user behavior based on the sequence of the sample behavior information; the sample feature vector sequence includes K sample feature vectors; A first input module, used for inputting the 1st to K-1th sample feature vectors in the sample feature vector sequence into an encoder to obtain a sample encoding vector and a state vector output by a preset network layer of the encoder; A first determination module is used to determine the prediction correlation corresponding to the Kth sample feature vector based on the sample coding vector, the state vector and the Kth sample feature vector in the sample feature vector sequence; wherein the prediction correlation is used to characterize the probability that the sample user performs the behavior corresponding to the Kth sample feature vector after performing the behaviors corresponding to the 1st to K-1th sample feature vectors; A training module, configured to train the encoder based on the predicted relevance and the actual relevance of the K-th sample feature vector; The second determination module is used to determine the information recommendation model based on the encoder obtained through training.

7. The determination device according to claim 6, characterized in that: The encoder includes: a first neural network and a second neural network; The first input module, when inputting the 1st to K-1th sample feature vectors in the sample feature vector sequence into the encoder to obtain the sample encoding vector and the state vector output by the preset network layer of the encoder, is used to: Inputting the 1st to K-2th sample feature vectors in the sample feature vector sequence into a first neural network to obtain sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively; and after inputting the K-2th sample feature vector sequence into the first neural network, obtaining an intermediate state vector from a preset network layer of the first neural network; Inputting the sample intermediate feature vectors, the intermediate state vector, and the K-1th sample feature vector corresponding to the 1st to K-2th sample feature vectors into a second neural network to obtain weights corresponding to the respective sample intermediate feature vectors; and obtaining the state vector from a preset network layer of the second neural network; The sample encoding vector is obtained based on each of the sample intermediate feature vectors and the weights respectively corresponding to each of the sample intermediate feature vectors.

8. The determination device according to claim 7, characterized in that: The first neural network includes: a first sub-network and a second sub-network; The first input module, when inputting the 1st to K-2th sample feature vectors in the sample feature vector sequence into the first neural network to obtain sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors respectively, is used to: Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the first sub-network in positive order, to obtain first intermediate vectors corresponding to the first to K-2th sample feature vectors respectively; Inputting the first to K-2th sample feature vectors in the sample feature vector sequence into the second sub-network in reverse order, to obtain second intermediate vectors corresponding to the first to K-2th sample feature vectors respectively; For each of the 1st to K-2th sample feature vectors, a sample intermediate feature vector corresponding to the sample feature vector is generated based on the first intermediate vector and the second intermediate vector corresponding to the sample feature vector.

9. The determination device according to claim 7, characterized in that: The second neural network includes: a third sub-network and a fourth sub-network; The first input module, when inputting the sample intermediate feature vectors corresponding to the 1st to K-2th sample feature vectors, the intermediate state vector, and the K-1th sample feature vector into the second neural network, and obtaining the weights corresponding to the respective sample intermediate feature vectors, is used to: Inputting the intermediate state vector and the K-1th sample feature vector into a third subnetwork to obtain a third intermediate vector corresponding to the K-1th sample feature vector; The third intermediate vector and the 1st to K-2th sample feature vectors are respectively input into the fourth sub-network as sample intermediate feature vectors to obtain weights corresponding to each of the sample intermediate feature vectors.

10. An information recommendation device, characterized in that: include: The second acquisition module is used to acquire historical behavior information of the user to be pushed for multiple pushed contents; A second construction module is used to construct a feature vector sequence corresponding to each content to be pushed based on the sequence of the historical behavior information and at least one content to be pushed; The second input module is used to input the feature vector sequence corresponding to each content to be pushed into the information recommendation model to obtain the relevance corresponding to the content to be pushed; A third determination module is used to determine at least one target push content from the plurality of contents to be pushed based on the magnitude of the relevance corresponding to each of the contents to be pushed, and push the target push content to the user to be pushed; The information recommendation model is determined based on the method for determining the information recommendation model according to any one of claims 1 to 4.

11. A computer device, characterized in that: include: A processor, a memory and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor and the memory communicate via the bus, and when the machine-readable instructions are executed by the processor, the steps of any method described in claims 1 to 5 are performed.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are executed.

Images