Article recommendation method and device, electronic equipment and storage medium

Through the combination of interest satisfaction perception network and interest preference network, the recommendation results are dynamically adjusted, and the problems of excessive recommendation and delayed feedback in the existing recommendation system are solved, improving recommendation efficiency and accuracy.

CN118941347APending Publication Date: 2024-11-12BEIJING WODONG TIANJUN INFORMATION TECH CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310539797.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-05-12
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing recommendation system has problems of excessive recommendation and delayed feedback, which leads to inefficient recommendations and failure to effectively model the user's interest satisfaction, resulting in continued recommendation of related items to users who have satisfied their interests, wasting resources.

Method used

Adopt interest satisfaction perception network and interest preference network, by obtaining the target user's data set and candidate item set, predict the user's interest satisfaction and interest preference for each candidate item, dynamically adjust the recommendation results, suppress the satisfied interests, and reduce resources. Waste and reduce feedback delays.

Benefits of technology

It effectively reduces the waste of recommendation resources and improves recommendation efficiency. By dynamically adjusting the interest preference network, it reduces the time for the interest preference network to update, and improves the response speed and accuracy of the recommendation system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941347A_ABST
    Figure CN118941347A_ABST
Patent Text Reader

Abstract

The embodiment of the invention provides an article recommendation method and device, electronic equipment and a computer storage medium. The method comprises the steps that a data set and a candidate article set of a target user are acquired; the data set comprises target user dimension data and article dimension data corresponding to the target user dimension data; predicting the data set by adopting an interest satisfaction degree perception network to obtain a first prediction result; the first prediction result represents the interest satisfaction degree of the target user for each candidate item in the candidate item set; based on the first prediction result, adopting an interest preference network to predict the data set to obtain a second prediction result; the second prediction result represents the interest preference of the target user for each candidate item; and based on the second prediction result, determining a target candidate item recommended to the target user from the candidate item set.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an item recommendation method, device, electronic device and computer storage medium. Background Art

[0002] At present, the recommendation system is an effective way to solve information overload and plays an important role in the fields of streaming media, e-commerce, etc. The essence of the recommendation system is to model the user's interest preferences based on the user's behavior information, and then help the user extract items that meet his or her interests from a large number of candidate items.

[0003] In the related art, recommendation ranking algorithms based on deep learning have been widely used in the industry, and mainstream algorithms include the Deep Interest Network (DIN) algorithm, the DMT (Deep Multifaceted Transformers) algorithm, and the ETA (End-to-end Target Attention) algorithm. Although the above-mentioned recommendation ranking model based on deep learning has achieved good results, the inventors have found that there is at least an over-recommendation problem in the related art. Summary of the invention

[0004] The present application provides an item recommendation method, device, electronic device and computer storage medium.

[0005] The technical solution of this application is implemented as follows:

[0006] The present invention provides an item recommendation method, which includes:

[0007] Acquire a target user data set and a candidate item set; the data set includes target user dimension data and item dimension data corresponding to the target user dimension data;

[0008] Using an interest satisfaction perception network to predict the data set, to obtain a first prediction result; the first prediction result represents the interest satisfaction of the target user for each candidate item in the candidate item set;

[0009] Based on the first prediction result, the data set is predicted using an interest preference network to obtain a second prediction result; the second prediction result represents the interest preference of the target user for each candidate item;

[0010] Based on the second prediction result, a target candidate item recommended to the target user is determined from the candidate item set.

[0011] The embodiment of the present application also proposes an item recommendation device, which includes an acquisition module, a first prediction module, a second prediction module and a determination module, wherein:

[0012] An acquisition module, used to acquire a data set of a target user and a set of candidate items; the data set includes target user dimension data and item dimension data corresponding to the target user dimension data;

[0013] A first prediction module is used to predict the data set using an interest satisfaction perception network to obtain a first prediction result; the first prediction result represents the interest satisfaction of the target user for each candidate item in the candidate item set;

[0014] A second prediction module is used to predict the data set using an interest preference network based on the first prediction result to obtain a second prediction result; the second prediction result represents the interest preference of the target user for each candidate item;

[0015] A determination module is used to determine a target candidate item recommended to the target user from the candidate item set based on the second prediction result.

[0016] An embodiment of the present application provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the item recommendation method provided by one or more of the aforementioned technical solutions when executing the program.

[0017] An embodiment of the present application provides a computer storage medium, which stores a computer program; after the computer program is executed, it can implement the item recommendation method provided by one or more of the above-mentioned technical solutions.

[0018] By obtaining a data set and a set of candidate items of a target user; the data set includes target user dimension data and item dimension data corresponding to the target user dimension data; and using an interest satisfaction perception network to predict the data set, a first prediction result is obtained; the first prediction result represents the interest satisfaction of the target user for each candidate item in the candidate item set; it can be seen that the embodiment of the present application proposes an interest satisfaction perception network, which can accurately determine the interest satisfaction of the target user for each candidate item; based on the first prediction result, an interest preference network is used to predict the data set to obtain a second prediction result; the second prediction result represents the interest preference of the target user for each candidate item; based on the second prediction result, the target candidate items recommended to the target user are determined from the candidate item set. It can be seen that in the embodiment of the present application, after obtaining the target user's interest satisfaction for each candidate item, the interest preference network can adjust the target user's interest preference for each candidate item in combination with the target user's interest satisfaction for each candidate item, that is, the subsequent item recommendation results can be dynamically adjusted according to the target user's interest satisfaction, so that the user's interest that has been satisfied can be suppressed, thereby reducing the waste of recommendation slot resources and solving the problem of over-recommendation in related technologies; further, since the embodiment of the present application can effectively suppress the target user's satisfied interests each time the target user's interest preferences are determined, the interest preference network does not need to update the feedback results according to a large amount of exposure but not clicked data generated after a period of time, and can also solve the problem of delayed feedback in the recommendation system in related technologies, thereby improving recommendation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A schematic diagram of a flow chart of an item recommendation method in an embodiment of the present application;

[0020] Figure 2 A schematic diagram of a flow chart of another method for recommending items in an embodiment of the present application;

[0021] Figure 3 A schematic diagram of a network structure of an item recommendation model in an embodiment of the present application;

[0022] Figure 4 A schematic diagram of the structure of the item recommendation device according to an embodiment of the present application;

[0023] Figure 5 A schematic diagram of the structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] The present application will be further described in detail below in conjunction with the accompanying drawings and Examples. It should be understood that the embodiments provided herein are only used to explain the present application and are not intended to limit the present application. In addition, the embodiments provided below are partial embodiments for implementing the present application, rather than providing all embodiments for implementing the present application. In the absence of conflict, the technical solutions recorded in the embodiments of the present application can be implemented in any combination.

[0025] It should be noted that, in the embodiments of the present application, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a method or device including a series of elements includes not only the elements explicitly recorded, but also includes other elements not explicitly listed, or also includes elements inherent to the implementation of the method or device. In the absence of further limitations, an element defined by the sentence "includes a ..." does not exclude the presence of other related elements (such as steps in a method or units in a device, for example, a unit may be a part of a circuit, a part of a processor, a part of a program or software, etc.) in the method or device including the element.

[0026] The term "and / or" herein is only a description of the association relationship of the associated objects, indicating that there may be three relationships. For example, I and / or J can represent the three situations that I exists alone, I and J exist at the same time, and J exists alone. In addition, the term "at least one" herein represents any combination of at least two of any one or more of a plurality of. For example, including at least one of I, J, and R can represent including any one or more elements selected from the set consisting of I, J, and R.

[0027] In the related art, the working mode of the recommendation system is: first, the user's preference representation vector is obtained based on the user's behavior data; then, the correlation between the user's preference representation vector and the candidate item representation vector is calculated; finally, based on the correlation, the item that the model believes the user will be interested in is recommended to the user. However, the existing recommendation system ignores the modeling of user interest satisfaction, that is, at a certain moment, the user's interest in a certain type of item has been satisfied, but the recommendation system still recommends related items to him, which will lead to over-recommendation. In addition, the current recommendation system has the problem of delayed feedback, which reduces the recommendation efficiency.

[0028] In some embodiments of the present application, the item recommendation method can be implemented using a processor in an item recommendation device, and the above-mentioned processor can be at least one of an application specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, and a microprocessor.

[0029] It should be noted that in the technical solution of this application, the collection, use, storage, sharing and transfer of user personal information involved are in compliance with the provisions of relevant laws and regulations, and it is necessary to inform the user and obtain the user's consent or authorization. When applicable, the user's personal information is de-identified and / or anonymized and / or encrypted.

[0030] Figure 1 FIG. 1 is a flow chart of an item recommendation method in an embodiment of the present application. Figure 1 As shown, the method comprises the following steps:

[0031] Step 100: Obtain the target user's data set and candidate item set.

[0032] In the embodiment of the present application, the target user represents a user who needs to have item recommendations made, that is, a user to be recommended; the data set of the target user may include target user dimension data and item dimension data corresponding to the target user dimension data.

[0033] Here, the target user dimension data at least includes the target user's portrait data; the item dimension data at least includes item-related data of multiple items that interact with the target user within a preset time period. Exemplarily, the preset time period refers to a time period of a preset duration. Here, the length of the preset time period is not limited, for example, it can be the past two weeks, or the past week, etc.

[0034] Exemplarily, user portrait data may include information such as the target user's age, gender and identity document (Id); interactive behaviors may include clicking behaviors, browsing behaviors, collection behaviors and ordering behaviors; item-related data may include information such as item number, category, name and the store to which it belongs.

[0035] In some embodiments, in addition to target user dimension data and item dimension data, the data set may also include context dimension data; wherein the context dimension data may include information such as the target user's access time, the time when the interaction behavior occurs, the device Internet Protocol (IP) address, and the device type.

[0036] Exemplarily, the candidate item set may be a collection of some or all items obtained from various e-commerce platforms accessible to the target user, or may be a collection of some or all items provided by multiple target e-commerce platforms. The target e-commerce platform may be an e-commerce platform recommended to the target user, or may be a collection of some or all items provided by the e-commerce platform currently selected by the target user. It should be noted that the candidate item set may include related items purchased by the target user, or may include items that the target user has not purchased.

[0037] In the embodiment of the present application, there is no limitation on the method of obtaining data of each dimension in the data set. For example, it can be obtained from the multi-source behavior log stored offline in the log server, or it can be obtained in real time from the e-commerce platform. Exemplarily, the multi-source behavior log may include logs generated by different sites, such as logs generated by domestic sites and foreign sites; it may also include logs generated by different ports, such as logs generated by the mini program end and the application software (Application, APP) end. Exemplarily, the key fields in the data set include at least the user number (userId), item number (itemId), timestamp (timeStamp) of the interactive behavior, behavior category, item category (cateId) and user device ID, etc.

[0038] Step 101: Use an interest satisfaction perception network to predict a data set to obtain a first prediction result.

[0039] In the embodiment of the present application, the first prediction result represents the target user's interest satisfaction for each candidate item in the candidate item set; that is, after using the interest satisfaction perception network prediction, the target user's interest satisfaction for each candidate item in the candidate item set can be obtained; here, the interest satisfaction corresponds to a numerical value, which can represent the user's interest in any candidate item.

[0040] It should be noted that, if the interest satisfaction corresponding to a candidate item is high, it means that the target user has a weaker interest in the candidate item, that is, the target user's interest in the candidate item is basically satisfied; conversely, if the interest satisfaction corresponding to a candidate item is low, it means that the target user has a strong interest in the candidate item, that is, the target user's interest in the candidate item has not been satisfied; subsequently, based on the target user's interest satisfaction for each candidate item, the target user's interest preference for each candidate item can be adjusted to make the recommendation results more in line with user needs and improve the accuracy of the recommendation results.

[0041] In some embodiments, the above method may also include: determining the behavior sequence of the target user based on the data set of the target user; dividing the behavior sequence to obtain multiple subsequences; determining the first vector corresponding to each subsequence in the multiple subsequences; the first vector includes vectors of multiple dimensions corresponding to each item involved in each subsequence.

[0042] Exemplarily, after obtaining the target user's data set, the data set is first preprocessed to obtain a preprocessed data set. It is understandable that since the target user's behavior data can be obtained from a variety of data sources, and the original data from each source is likely to have noise data, inconsistent formats, and data dispersion, etc., after obtaining the target user's data set, it is first preprocessed to obtain a preprocessed data set.

[0043] Here, preprocessing may include operations such as extraction, cleaning and conversion, which are not specifically limited in the embodiments of the present application; it should be noted that by preprocessing the target user's data set, the scattered, messy and non-uniformly formatted data in the data set can be integrated into standardized structured data.

[0044] Exemplarily, since the data set includes data of multiple dimensions, namely, target user dimension data, item dimension data, and context dimension data; the preprocessed data set also includes data of multiple dimensions; by performing feature extraction on the data of each dimension in the preprocessed data set, a corresponding feature vector set of multiple dimensions can be obtained.

[0045] Among them, the feature vector set of each dimension includes multiple feature vectors; illustratively, taking the target user dimension data as an example, after feature extraction, the feature vector set of the target user dimension can be obtained; if the feature vector set of the target user dimension includes the gender characteristics and age characteristics of the target user, then the vector (0, 0, 1) can be used to represent males, the vector (0, 1, 0) can be used to represent females, and the vector (1, 0, 0) can be used to represent the case where the gender characteristics are unknown; for age characteristics, the age range of 0-30 can be represented by the vector (1, 0, 0, 0), and the age range of 31-60 can be represented by the vector (0, 1, 0, 0).

[0046] It should be noted that in order to reduce the impact of different feature dimensions of each dimension on the network prediction results, after obtaining the feature vector set of each dimension, each vector in each feature vector set can be normalized to unify the feature dimension before subsequent processing; here, there is no limitation on the method of achieving normalization, for example, it can be achieved by bucketing and hashing.

[0047] In some embodiments, after obtaining the data set of the target user, the behavior sequence of the target user can also be extracted from the data set, and the extracted behavior sequence can be divided according to a set time interval to obtain multiple subsequences; wherein each subsequence has a corresponding serial number; here, there is no limitation on the value of the serial number.

[0048] Exemplarily, each of the multiple subsequences involves at least one item that interacts with the target user; the target user's behavior sequence can be understood as a sequence consisting of multiple items that interact with the target user, that is, the behavior sequence includes multiple items that interact with the target user; here, there is no limitation on the length of the set time interval, for example, it can be every day, every hour, etc.

[0049] For example, suppose the subsequence can be represented by the session sequence S k , where k represents the sequence number of the subsequence, the value of k ranges from 1 to n, n is an integer greater than 1, and n represents the number of subsequences after division, that is, the number of session sequences; each session sequence includes a part of continuous items involved in the behavior sequence, and connecting all session sequences can obtain the set of items for the target user to complete the interactive behavior.

[0050] It can be understood that in order to reduce the impact of noise data, after obtaining the multiple subsequences after division and the feature vector sets of multiple dimensions after normalization, the first vector corresponding to each subsequence in the multiple subsequences can be determined based on each subsequence as a unit; here, the first vector can include vectors of multiple dimensions corresponding to each item involved in each subsequence; illustratively, the multiple dimensions can include a target user dimension, an item dimension, and a context dimension.

[0051] In some embodiments, using an interest satisfaction perception network to predict a data set to obtain a first prediction result may include: using an interest satisfaction perception network to predict a first vector corresponding to each subsequence to obtain a first prediction result.

[0052] Exemplarily, after obtaining the first vector corresponding to each subsequence, the vectors of multiple dimensions corresponding to each item involved in each subsequence can be concatenated to obtain the second vector corresponding to each subsequence; here, the second vector includes the concatenated vectors corresponding to each item involved in each subsequence.

[0053] For example, suppose that the vectors of multiple dimensions corresponding to item i (item i represents the i-th item in the subsequence) involved in a certain subsequence include the target user dimension vector E u 、Item dimension hidden vector I i And the context dimension hidden vector C i By concatenating the vectors of multiple dimensions corresponding to item i, the concatenated latent vector corresponding to item i can be obtained. By analogy, the concatenated vectors corresponding to the items involved in the subsequence can be obtained; wherein the set of these concatenated vectors is the second vector corresponding to the subsequence.

[0054] In an embodiment of the present application, the interest preference network may include an encoder network in a Transformer network; in some embodiments, taking each subsequence as a unit, the second vector corresponding to each subsequence, that is, the concatenated vector corresponding to each item involved in each subsequence is input into the encoder network, and a third vector corresponding to the first subsequence can be obtained; here, the third vector includes the updated vector corresponding to each item involved in each subsequence.

[0055] Exemplarily, the encoder network uses a self-attention mechanism to generate an updated vector X corresponding to each item involved in each subsequence i If we take the session sequence of item i as an example, the calculation method is as shown in formula (1):

[0056]

[0057] in, and represents the set of concatenated vectors corresponding to items other than item i in the items involved in the current session sequence, T represents transposition, d represents the vector dimension corresponding to the set, and softmax is the activation function; after the above formula, the encoder network finally outputs the updated vectors corresponding to each item involved in each subsequence; wherein, the set of these updated vectors is the third vector corresponding to the current session sequence; here, the set of third vectors corresponding to multiple subsequences can be represented as X.

[0058] In some embodiments, an interest satisfaction perception network is used to predict the third vector corresponding to each subsequence to obtain a first prediction result, which may include: determining a fourth vector corresponding to the candidate item set; inputting the third vector corresponding to each subsequence and the fourth vector corresponding to the candidate item set into the interest satisfaction perception network to obtain a first prediction result.

[0059] In the embodiment of the present application, the fourth vector includes the vector of each candidate item in the candidate item set; illustratively, after obtaining the candidate item set, the vector of each candidate item in the candidate item can be further determined. The determination process can refer to the above-mentioned process of determining the feature vector set of each dimension and normalizing it, which will not be repeated here; assuming that the fourth vector corresponding to the candidate item set is E tar , the output X of the above encoder network and the fourth vector E of the candidate items tar By inputting it into the interest satisfaction perception network, the target user's interest satisfaction for each candidate item can be obtained.

[0060] Exemplarily, the principle of generating interest satisfaction is as follows: First, according to the updated vectors corresponding to the items involved in each session sequence obtained by division and the multiple subsequences output by the encoder network, a neural network is used in each session sequence to model the final updated vector S of each item involved in each session sequence k (Here, take the kth session sequence as an example), then, the set S composed of each session sequence is S = {S1,...,S k ,...S n} and the fourth vector E corresponding to the candidate item set tar Here, the correlation between each session sequence and the candidate item set is calculated as β = {β1, ..., β k ,...β n}; Finally, the softmax activation function is used to activate the generated correlation β to generate the final first prediction result, that is, the target user's interest satisfaction for each candidate item in the candidate item set.

[0061] Here, there is no limitation on the type of the above neural network, for example, it can be a recurrent neural network (RNN) or a gated recurrent neural network (GRU).

[0062] In some embodiments, after obtaining the third vector corresponding to each subsequence and the fourth vector corresponding to the candidate item set, the third vector corresponding to each subsequence and the fourth vector corresponding to the candidate item set can be input into the trained interest satisfaction perception network to obtain a first prediction result.

[0063] Exemplarily, the training process of the interest satisfaction perception network may include the following steps: obtaining data sets of multiple historical users in a first historical time period and a second historical time period; preprocessing, feature extraction, normalization and splicing of data of multiple dimensions in the data sets of the two historical time periods, respectively, to obtain embedding vectors of multiple dimensions of multiple historical users in the first historical time period and the second historical time period, respectively; using the embedding vectors of multiple dimensions of multiple historical users in the first historical time period to pre-train the constructed initial interest satisfaction perception network, and then using the embedding vectors of multiple dimensions of multiple historical users in the second historical time period to fine-tune the pre-trained interest satisfaction perception network, to obtain a trained interest satisfaction perception network.

[0064] Here, the first historical time period is the time period before the second historical time period, and the length of the first historical time period is usually greater than that of the second historical time period; illustratively, assuming that the current day is the Tth day, the T-30th day to the T-8th day can be determined as the first historical time period, and the T-7th day to the T-1st day can be determined as the second historical time period; it should be noted that the difference between the data set of each historical user in the first historical time period and the data set of the second historical time period is only that the time window is different, and the data type included in the data set is the same; the method for obtaining the data set of each historical user can refer to step 100, which is not repeated here.

[0065] For example, during the training of the interest satisfaction perception network, the exposure-but-unclicked behavior data in the data set of each historical user can also be used as a supervision label to characterize the interest satisfaction of each historical user on different items.

[0066] Step 102: Based on the first prediction result, an interest preference network is used to predict the data set to obtain a second prediction result.

[0067] Exemplarily, the second prediction result may include an interest preference vector of the target user for each candidate item, where the interest preference vector is used to characterize the interest preference of the target user for each candidate item.

[0068] In an embodiment of the present application, the interest preference network also includes a decoder network in the Transformer network; wherein the encoder network and the decoder network are two important sub-networks of the Transformer network, and their internal structures are composed of a multi-head attention network layer, a forward feedback neural network layer, and a residual normalization network layer.

[0069] For example, in the training process of the interest preference network, the exposure-unclicked behavior data in the data set of each historical user can be used as a supervision label to assist in the modeling of the interest preferences of each historical user, so as to obtain a trained interest preference network. The basic idea is as follows: Based on the underlying input data set including multiple dimensions such as each historical user and item, the interest preference vector of each historical user is generated. Here, I is the item dimension vector corresponding to each historical user behavior sequence; the fourth vector E corresponding to the candidate item set tar The satisfaction score is scored (cosine distance can be used for measurement and scoring), where the category of each candidate item in the candidate item set is cid3. The supervised label generation method is as follows: according to the category cid3 of each candidate item, the exposure and click behavior time of each historical user in the most recent session sequence is set, where the setting of time is a hyperparameter and will be optimized according to the training of the network.

[0070] In some embodiments, based on the first prediction result, using an interest preference network to predict the data set to obtain a second prediction result can include: inputting the first prediction result and a third vector corresponding to each subsequence into a decoder network to obtain a second prediction result.

[0071] Exemplarily, after obtaining the first prediction result output by the interest satisfaction perception network and the third vector corresponding to each subsequence output by the encoder network according to the above steps, these two parts of information can be input into the decoder network to obtain the second prediction result; wherein the second prediction result includes the interest preference vector of the target user for each candidate item; wherein the decoder network adopts a multi-head self-attention network, and the relevant calculation formula is shown in formula (2):

[0072]

[0073] Here, K and V are both outputs of the encoder network, which include the updated vectors corresponding to each item involved in the target user's behavior sequence; α represents the target user's interest satisfaction set for each candidate item in the candidate item set, h represents the number of head self-attention networks in the decoder network, and W H is the parameter matrix; where, in the interest preference network training process, the parameter matrix W HThe parameters of are randomly initialized and will be further optimized as the network is trained. It can be seen that by concatenating the output results of each head self-attention network, E interest , that is, the target user's interest preference vector for each candidate item.

[0074] It can be seen that in the embodiment of the present application, the interest preference vector of the target user output by the interest preference network can be dynamically adjusted according to the interest satisfaction of the target user, so that the user interest that has been satisfied can be suppressed, thereby ensuring the accuracy of subsequent item recommendation results.

[0075] Step 103: Based on the second prediction result, determine a target candidate item to be recommended to the target user from the candidate item set.

[0076] In an embodiment of the present application, after obtaining the second prediction result, multiple candidate items in the candidate item set can be sorted based on the second prediction result to obtain a sorting result; based on the sorting result, a target candidate item recommended to the target user is determined from the multiple candidate items.

[0077] Exemplarily, the interest preference network may also include multiple expert networks, and sorting multiple candidate items in the candidate item set based on the second prediction result may include: inputting the second prediction result into multiple expert networks to obtain fifth vectors corresponding to the multiple expert networks; splicing the fifth vectors corresponding to the multiple expert networks to obtain a target vector; and sorting the multiple candidate items in the candidate item set based on the target vector.

[0078] It can be understood that in order to reduce the prediction error of the interest preference network, the embodiment of the present application integrates the user interest preference vector E with interest satisfaction interest Input to each expert network to learn the feature information of different dimensions in different fields, and obtain the vector results (M1,...,M n ), that is, the fifth vector; then, the fifth vectors corresponding to multiple different expert networks are concatenated to obtain the target vector M = Concate(M1,...,M n ), where n represents the number of expert networks.

[0079] Exemplarily, the interest preference network may also include a multi-task learning network, which, based on the target vector, sorts multiple candidate items in the candidate item set, and may include: inputting the target vector into the multi-task learning network to obtain the target user's preference probability value for each of the multiple candidate items; and sorting the multiple candidate items based on the preference probability value corresponding to each candidate item.

[0080] Here, there is no limitation on the type of multi-task learning network, for example, it can be a click rate prediction network or a conversion rate prediction network, etc.; illustratively, when the multi-task learning network is a click rate prediction network, the preference probability value corresponds to the click probability value; when the multi-task learning network is a conversion rate prediction network, the preference probability value corresponds to the conversion probability value; wherein, the click probability value indicates the probability value of the target user clicking on a candidate item; and the conversion probability value indicates the probability value of the target user purchasing a candidate item. The click rate prediction network is taken as an example for exemplary description.

[0081] For example, after obtaining the target vectors corresponding to multiple expert networks, they can be input into the click rate prediction network to obtain the click probability value of the target user for each candidate item in the candidate item set; then, the multiple candidate items in the candidate item set can be sorted in descending order of the click probability value to obtain a sorting result; from the sorting result, several candidate items ranked at the top are selected as target candidate items and recommended to the target user, so that personalized recommendation of candidate items can be achieved.

[0082] An embodiment of the present application proposes an item recommendation method, which includes: obtaining a data set and a candidate item set of a target user; the data set includes target user dimensional data and item dimensional data corresponding to the target user dimensional data; using an interest satisfaction perception network to predict the data set to obtain a first prediction result; the first prediction result represents the target user's interest satisfaction for each candidate item in the candidate item set; based on the first prediction result, using an interest preference network to predict the data set to obtain a second prediction result; the second prediction result represents the target user's interest preference for each candidate item; based on the second prediction result, determining a target candidate item recommended to the target user from the candidate item set.

[0083] It can be seen that the embodiment of the present application proposes an interest satisfaction perception network, which can be used to determine the target user's interest satisfaction for each candidate item; on this basis, the interest preference network can adjust the target user's interest preference vector for each candidate item in combination with the target user's interest satisfaction for each candidate item, that is, the subsequent item recommendation results can be dynamically adjusted according to the target user's interest satisfaction, so that the user's interest that has been satisfied can be suppressed, thereby reducing the waste of recommendation slot resources and solving the problem of over-recommendation in related technologies; in addition, since the embodiment of the present application can effectively suppress the target user's already satisfied interests each time the target user's interest preference vector is determined, the interest preference network does not need to update the feedback results based on a large amount of exposure but not clicked data generated after a period of time, thereby solving the problem of delayed feedback in the recommendation system in related technologies and improving recommendation efficiency.

[0084] In order to better reflect the purpose of this application, further explanation is given based on the above embodiments of this application.

[0085] Figure 2 FIG. 1 is a flow chart of another method for recommending items in an embodiment of the present application. Figure 2 As shown, the method comprises the following steps:

[0086] Step 200: Obtain the target user's data set and perform preprocessing;

[0087] Step 201: determine a feature vector set of multiple dimensions and normalize and concatenate them;

[0088] Step 202: inputting the processed vectors of multiple dimensions into an interest satisfaction perception network;

[0089] Step 203: input the processed vectors of multiple dimensions into the interest preference network, and adjust the interest preference vector of the target user using the interest satisfaction output by the interest satisfaction perception network;

[0090] Step 204: Generate recommendation results based on the interest preference vector of the target user.

[0091] It should be noted that the specific implementation of step 200 to step 204 has been described in the above embodiment and will not be repeated here.

[0092] Next, we will further explain the model's network structure. Figure 3 Schematic diagram of the network structure of an item recommendation model in an embodiment of the present application; Figure 3 As shown in the figure, the model mainly consists of seven layers, from bottom to top: feature layer of multiple dimensions, interest satisfaction perception network layer, Transformer network layer (including Figure 3 Transformer encoder network layer and Transformer decoder network layer), expert network layer, multi-task learning network layer and final scoring layer.

[0093] Among them, the feature layers of multiple dimensions are equivalent to the network layers corresponding to the feature vector sets of the above multiple dimensions; the interest satisfaction perception network layer is equivalent to the network layer corresponding to the above interest satisfaction perception network; the Transformer encoder network layer is equivalent to the network layer corresponding to the above encoder network; the Transformer decoder network layer is equivalent to the network layer corresponding to the above decoder network; the expert network layer is equivalent to the network layer corresponding to the above multiple expert networks (expert network 1 to expert network n), the multi-task learning network layer is equivalent to the network layer corresponding to the above multi-task learning network; the final scoring layer is equivalent to the above network layer for sorting multiple candidate items in the candidate item set.

[0094] In an embodiment of the present application, by modeling the user's interest satisfaction and dynamically adjusting the interest preference vector, the user's interest preferences that have been satisfied are suppressed, and the user's interest preferences that have not been satisfied are promoted, so as to alleviate the problem of over-recommendation and further avoid the waste of recommendation slots and the problem of delayed feedback.

[0095] Figure 4 Schematic diagram of the composition structure of the item recommendation device of the embodiment of the present application. Figure 4 As shown, the device includes: an acquisition module 300, a first prediction module 301, a second prediction module 302 and a determination module 303, wherein:

[0096] An acquisition module 300 is used to acquire a target user data set and a candidate item set; the data set includes target user dimension data and item dimension data corresponding to the target user dimension data;

[0097] The first prediction module 301 is used to predict the data set using an interest satisfaction perception network to obtain a first prediction result; the first prediction result represents the interest satisfaction of the target user for each candidate item in the candidate item set;

[0098] A second prediction module 302 is used to predict the data set using an interest preference network based on the first prediction result to obtain a second prediction result; the second prediction result represents the interest preference of the target user for each candidate item;

[0099] The determination module 303 is configured to determine, based on the second prediction result, a target candidate item to be recommended to the target user from the candidate item set.

[0100] In some embodiments, the first prediction module 301 is further configured to:

[0101] Determining a behavior sequence of the target user according to the data set of the target user;

[0102] Dividing the behavior sequence to obtain multiple subsequences;

[0103] Determine a first vector corresponding to each subsequence in the multiple subsequences; the first vector includes vectors of multiple dimensions corresponding to each item involved in each subsequence;

[0104] An interest satisfaction perception network is used to predict the first vector corresponding to each subsequence to obtain a first prediction result.

[0105] In some embodiments, the interest preference network includes an encoder network, and the first prediction module 301 is further used to:

[0106] Concatenate the vectors of multiple dimensions corresponding to the items involved in each subsequence to obtain a second vector corresponding to each subsequence;

[0107] Inputting the second vector corresponding to each subsequence into the encoder network to obtain a third vector corresponding to each subsequence;

[0108] The interest satisfaction perception network is used to predict the third vector corresponding to each subsequence to obtain the first prediction result.

[0109] In some embodiments, the first prediction module 301 is further configured to:

[0110] Determine a fourth vector corresponding to the candidate item set; the fourth vector includes a vector of each candidate item in the candidate item set;

[0111] The third vector corresponding to each subsequence and the fourth vector corresponding to the candidate item set are input into the interest satisfaction perception network to obtain the first prediction result.

[0112] In some embodiments, the interest preference network further includes a decoder network, and the second prediction module 302 is further configured to:

[0113] The first prediction result and the third vector corresponding to each subsequence are input into the decoder network to obtain the second prediction result.

[0114] In some embodiments, the determining module 303 is further configured to:

[0115] sorting the plurality of candidate items in the candidate item set based on the second prediction result to obtain a sorting result;

[0116] Based on the ranking result, a target candidate item recommended to the target user is determined from the multiple candidate items.

[0117] In some embodiments, the interest preference network further includes a plurality of expert networks, and the determination module 303 is further configured to:

[0118] Inputting the second prediction result into the multiple expert networks to obtain fifth vectors corresponding to the multiple expert networks;

[0119] splicing the fifth vectors corresponding to the multiple expert networks to obtain a target vector;

[0120] Based on the target vector, multiple candidate items in the candidate item set are sorted.

[0121] In some embodiments, the interest preference network further includes a multi-task learning network, and the determination module 303 is further used to:

[0122] Inputting the target vector into the multi-task learning network to obtain a preference probability value of the target user for each of the multiple candidate items;

[0123] Based on the preference probability value corresponding to each candidate item, the multiple candidate items are sorted. It can be seen that the item recommendation device provided in the embodiment of the present application proposes an interest satisfaction perception network, which can be used to determine the target user's interest satisfaction for each candidate item; on this basis, the interest preference network can adjust the target user's interest preference vector for each candidate item in combination with the target user's interest satisfaction for each candidate item, that is, the subsequent item recommendation results can be dynamically adjusted according to the target user's interest satisfaction, so that the user's interest that has been satisfied can be suppressed, thereby reducing the waste of recommendation slot resources; in addition, since the embodiment of the present application can effectively suppress the target user's already satisfied interest each time the target user's interest preference vector is determined, the interest preference network does not need to update the feedback results based on the large amount of exposure and non-click data generated after a period of time, thereby improving the recommendation efficiency.

[0124] In practical applications, the acquisition module 300, the first prediction module 301, the second prediction module 302 and the determination module 303 can all be implemented by a processor located in an electronic device, and the processor can be at least one of an ASIC, a DSP, a DSPD, a PLD, a FPGA, a CPU, a controller, a microcontroller, and a microprocessor.

[0125] Based on the same technical concept as the above embodiments, see Figure 5 , which shows an electronic device 400 provided by the present application, which may include: a memory 401 and a processor 402; wherein,

[0126] Memory 401, used for storing computer programs and data;

[0127] The processor 402 is used to execute the computer program stored in the memory to implement any one of the item recommendation methods in the aforementioned embodiments.

[0128] In practical applications, the memory 401 may be a volatile memory, such as RAM; or a non-volatile memory, such as ROM, flash memory, hard disk drive (HDD) or solid-state drive (SSD); or a combination of the above types of memory, and provide instructions and data to the processor 402.

[0129] The processor 402 may be at least one of ASIC, DSP, DSPD, PLD, FPGA, CPU, controller, microcontroller, and microprocessor. It is understandable that for different electronic devices, the electronic device used to implement the processor function may also be other, which is not specifically limited in the embodiments of the present application.

[0130] An embodiment of the present application further provides a computer storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, any one of the item recommendation methods of the aforementioned embodiments is implemented.

[0131] In addition, each functional module in this embodiment can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or software functional modules.

[0132] If the integrated unit is implemented in the form of a software function module and is not sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment is essentially or the part that contributes to the relevant technology or the whole or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) or a processor to perform all or part of the steps of the method of this embodiment. The aforementioned storage medium includes: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, and other media that can store program codes.

[0133] Specifically, the computer program instructions corresponding to an item recommendation method in this embodiment can be stored on a storage medium such as a CD, a hard disk, or a USB flash drive. When the computer program instructions corresponding to an item recommendation method in the storage medium are read or executed by an electronic device, any item recommendation method in the aforementioned embodiments is implemented.

[0134] In some embodiments, the functions or modules included in the device provided in the embodiments of the present application can be used to execute the method described in the above method embodiments. The specific implementation can refer to the description of the above method embodiments. For the sake of brevity, it will not be repeated here.

[0135] The above description of various embodiments tends to emphasize the differences between the various embodiments. The same or similar aspects can be referenced to each other, and for the sake of brevity, they will not be repeated herein.

[0136] The methods disclosed in the various method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0137] The features disclosed in the various product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0138] The features disclosed in the various method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0139] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of hardware embodiments, software embodiments, or embodiments in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage and optical storage, etc.) that contain computer-usable program code.

[0140] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0141] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0142] The above are only preferred embodiments of the present application and are not intended to limit the protection scope of the present application.

Claims

1. A method for recommending items. It is characterized in that The method comprises: Acquire a target user data set and a candidate item set; the data set includes target user dimension data and item dimension data corresponding to the target user dimension data; Using an interest satisfaction perception network to predict the data set, to obtain a first prediction result; the first prediction result represents the interest satisfaction of the target user for each candidate item in the candidate item set; Based on the first prediction result, the data set is predicted using an interest preference network to obtain a second prediction result; the second prediction result represents the interest preference of the target user for each candidate item; Based on the second prediction result, a target candidate item recommended to the target user is determined from the candidate item set.

2. The method according to claim 1, It is characterized in that The method further comprises: Determining a behavior sequence of the target user according to the data set of the target user; Dividing the behavior sequence to obtain multiple subsequences; Determine a first vector corresponding to each subsequence in the multiple subsequences; the first vector includes vectors of multiple dimensions corresponding to each item involved in each subsequence; Accordingly, the use of the interest satisfaction perception network to predict the data set to obtain a first prediction result includes: An interest satisfaction perception network is used to predict the first vector corresponding to each subsequence to obtain a first prediction result.

3. The method according to claim 2, It is characterized in that The interest preference network includes an encoder network, and the interest satisfaction perception network is used to predict the first vector corresponding to each subsequence to obtain a first prediction result, including: Concatenate the vectors of multiple dimensions corresponding to the items involved in each subsequence to obtain a second vector corresponding to each subsequence; Inputting the second vector corresponding to each subsequence into the encoder network to obtain a third vector corresponding to each subsequence; The interest satisfaction perception network is used to predict the third vector corresponding to each subsequence to obtain the first prediction result.

4. The method according to claim 3, It is characterized in that The using the interest satisfaction perception network to predict the third vector corresponding to each subsequence to obtain the first prediction result includes: Determine a fourth vector corresponding to the candidate item set; the fourth vector includes a vector of each candidate item in the candidate item set; The third vector corresponding to each subsequence and the fourth vector corresponding to the candidate item set are input into the interest satisfaction perception network to obtain the first prediction result.

5. The method according to claim 3 or 4, It is characterized in that The interest preference network further includes a decoder network. Based on the first prediction result, the interest preference network is used to predict the data set to obtain a second prediction result, including: The first prediction result and the third vector corresponding to each subsequence are input into the decoder network to obtain the second prediction result.

6. The method according to claim 1, It is characterized in that The determining, based on the second prediction result, a target candidate item recommended to the target user from the candidate item set includes: sorting the plurality of candidate items in the candidate item set based on the second prediction result to obtain a sorting result; Based on the ranking result, a target candidate item recommended to the target user is determined from the multiple candidate items.

7. The method according to claim 6, It is characterized in that The interest preference network further includes a plurality of expert networks, and the sorting of the plurality of candidate items in the candidate item set based on the second prediction result includes: Inputting the second prediction result into the multiple expert networks to obtain fifth vectors corresponding to the multiple expert networks; splicing the fifth vectors corresponding to the multiple expert networks to obtain a target vector; Based on the target vector, multiple candidate items in the candidate item set are sorted.

8. The method according to claim 7, It is characterized in that The interest preference network further includes a multi-task learning network, and the sorting of multiple candidate items in the candidate item set based on the target vector includes: Inputting the target vector into the multi-task learning network to obtain a preference probability value of the target user for each of the multiple candidate items; The multiple candidate items are sorted based on the preference probability value corresponding to each candidate item.

9. An item recommendation device, It is characterized in that The device comprises: An acquisition module, used to acquire a data set of a target user and a set of candidate items; the data set includes target user dimension data and item dimension data corresponding to the target user dimension data; A first prediction module is used to predict the data set using an interest satisfaction perception network to obtain a first prediction result; the first prediction result represents the interest satisfaction of the target user for each candidate item in the candidate item set; A second prediction module is used to predict the data set using an interest preference network based on the first prediction result to obtain a second prediction result; the second prediction result represents the interest preference of the target user for each candidate item; A determination module is used to determine a target candidate item recommended to the target user from the candidate item set based on the second prediction result.

10. An electronic device, It is characterized in that The device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the method according to any one of claims 1 to 8 is implemented when the processor executes the program.

11. A computer storage medium having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.