Data recommendation method and device, electronic equipment and computer readable medium

Abstract

The invention discloses a data recommendation method and device, electronic equipment and a computer readable medium, and relates to the technical field of computers.The method comprises the steps that user context data and user history browsing item data are obtained according to a received data recommendation request; based on the historical browsing article data of the user, obtaining a pooling vector through a pooling network, and according to the historical browsing article data of the user and the to-be-recommended article data, obtaining a deep interest network representation vector through a deep interest network; determining a first mapping matrix and a second mapping matrix based on the preset representation dimension and the preset real number space configuration data, and further determining a weighted proportionality coefficient based on the user context data; determining a target interest vector according to the weighted proportionality coefficient, the pooling vector and the deep interest network representation vector; and based on the target interest vector, determining target recommended article data in the to-be-recommended article data. The performance of the recommendation system is improved, and personalized recommendation is realized.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a data recommendation method, device, electronic device and computer-readable medium. Background Art

[0002] At present, recommendation systems play a vital role in the modern digital world. They can not only greatly improve user experience, but also bring significant economic benefits to enterprises. The more commonly used sequence networks in recommendation systems are Pooling Networks and Deep Interest Networks (DIN). There are also ultra-long sequence architectures such as SIM to process large-scale user sequences. Whether it is DIN interest network or Pooling network, it is impossible to achieve true data personalized recommendation, and the efficiency and accuracy of data personalized recommendation are low. Summary of the invention

[0003] In view of this, embodiments of the present application provide a data recommendation method, device, electronic device, and computer-readable medium, which can solve the problems of low efficiency and accuracy of existing personalized data recommendations.

[0004] To achieve the above objective, according to one aspect of an embodiment of the present application, a data recommendation method is provided, comprising:

[0005] Acquire user context data and user historical browsing item data according to the received data recommendation request;

[0006] Based on the user's historical browsing data, a pooling vector is obtained through a pooling network. Based on the user's historical browsing data and the data of items to be recommended, a deep interest network representation vector is obtained through a deep interest network.

[0007] Determine a first mapping matrix and a second mapping matrix based on a preset representation dimension and preset real number space configuration data, and further determine a weighted proportional coefficient based on the user context data;

[0008] Determine the target interest vector according to the weighted proportional coefficient, the pooling vector and the deep interest network representation vector;

[0009] Based on the target interest vector, target recommended item data in the to-be-recommended item data is determined.

[0010] Optionally, based on the user's historical browsing item data, a pooling vector is obtained through a pooling network, including:

[0011] Based on the user's historical browsing item data, determine the corresponding historical browsing item vector;

[0012] The pooled vector is obtained by summing up the historical browsed item vectors through the pooling network.

[0013] Optionally, obtaining a deep interest network representation vector through a deep interest network includes:

[0014] The similarity between the user's historical browsing item data and the item data to be recommended is calculated through the deep interest network to obtain a similarity score;

[0015] Normalizing the similarity scores to obtain a normalized similarity score;

[0016] Based on the normalized similarity score and the user sequence feature vector corresponding to the corresponding user historical browsing item data, a deep interest network representation vector is obtained.

[0017] Optionally, determining the first mapping matrix and the second mapping matrix includes:

[0018] Determining a first parameter value based on a preset representation dimension;

[0019] Obtaining shape data and dimension data in the preset real number space configuration data, and determining a second parameter value according to the shape data and the dimension data;

[0020] According to the first parameter value and the second parameter value, the dimension and shape of the mapping matrix are determined, and then the corresponding first mapping matrix and second mapping matrix are obtained.

[0021] Optionally, based on the user context data, a weighted proportionality factor is determined, including:

[0022] Convert the user context data into a user context data vector, and multiply the first mapping matrix by the user context data vector to obtain first calculation result data;

[0023] Multiplying the second mapping matrix by the first calculation result data on the left to obtain the second calculation result data;

[0024] The second calculation result data is normalized to obtain a weighted proportional coefficient.

[0025] Optionally, determining a target interest vector includes:

[0026] Calculate a first product of the pooling vector and the corresponding weighted proportional coefficient and a second product of the deep interest network representation vector and the corresponding weighted proportional coefficient;

[0027] The sum of the first product and the second product is determined as the target interest vector.

[0028] Optionally, determining target recommended item data in the to-be-recommended item data based on the target interest vector includes:

[0029] Input the target interest vector into the deep learning network model of the recommendation system to output the estimated browsing probability of the user for the corresponding recommended item data;

[0030] Based on the browsing probability, the target recommended item data is determined.

[0031] In addition, the present application also provides a data recommendation device, including:

[0032] an acquisition unit, configured to acquire user context data and user historical browsing item data according to the received data recommendation request;

[0033] The vector conversion unit is configured to obtain a pooling vector through a pooling network based on the user's historical browsing item data, and obtain a deep interest network representation vector through a deep interest network based on the user's historical browsing item data and the item data to be recommended;

[0034] A weighted proportionality coefficient determination unit is configured to determine a first mapping matrix and a second mapping matrix based on a preset representation dimension and preset real number space configuration data, and further determine a weighted proportionality coefficient based on user context data;

[0035] A target interest vector determination unit is configured to determine a target interest vector according to a weighted proportionality coefficient, a pooling vector and a deep interest network representation vector;

[0036] The target recommended item data determining unit is configured to determine target recommended item data in the to-be-recommended item data based on the target interest vector.

[0037] Optionally, the vector conversion unit is further configured to:

[0038] Based on the user's historical browsing item data, determine the corresponding historical browsing item vector;

[0039] The pooled vector is obtained by summing up the historical browsed item vectors through the pooling network.

[0040] Optionally, the vector conversion unit is further configured to:

[0041] The similarity between the user's historical browsing item data and the item data to be recommended is calculated through the deep interest network to obtain a similarity score;

[0042] Normalizing the similarity scores to obtain a normalized similarity score;

[0043] Based on the normalized similarity score and the user sequence feature vector corresponding to the corresponding user historical browsing item data, a deep interest network representation vector is obtained.

[0044] Optionally, the weighted proportional coefficient determination unit is further configured to:

[0045] Determining a first parameter value based on a preset representation dimension;

[0046] Obtaining shape data and dimension data in the preset real number space configuration data, and determining a second parameter value according to the shape data and the dimension data;

[0047] According to the first parameter value and the second parameter value, the dimension and shape of the mapping matrix are determined, and then the corresponding first mapping matrix and second mapping matrix are obtained.

[0048] Optionally, the weighted proportional coefficient determination unit is further configured to:

[0049] Convert the user context data into a user context data vector, and multiply the first mapping matrix by the user context data vector to obtain first calculation result data;

[0050] Multiplying the second mapping matrix by the first calculation result data on the left to obtain the second calculation result data;

[0051] The second calculation result data is normalized to obtain a weighted proportional coefficient.

[0052] Optionally, the target interest vector determining unit is further configured to:

[0053] Calculate a first product of the pooling vector and the corresponding weighted proportional coefficient and a second product of the deep interest network representation vector and the corresponding weighted proportional coefficient;

[0054] The sum of the first product and the second product is determined as the target interest vector.

[0055] Optionally, the target recommended item data determination unit is further configured to:

[0056] Input the target interest vector into the deep learning network model of the recommendation system to output the estimated browsing probability of the user for the corresponding recommended item data;

[0057] Based on the browsing probability, the target recommended item data is determined.

[0058] In addition, the present application also provides a data recommendation electronic device, including: one or more processors; a storage device for storing one or more programs, when the one or more programs are executed by one or more processors, the one or more processors implement the data recommendation method as described above.

[0059] In addition, the present application also provides a computer-readable medium on which a computer program is stored, and when the program is executed by a processor, the data recommendation method as described above is implemented.

[0060] To achieve the above objective, according to another aspect of the embodiments of the present application, a computer program product is provided.

[0061] A computer program product according to an embodiment of the present application includes a computer program, and when the program is executed by a processor, the data recommendation method provided by the embodiment of the present application is implemented.

[0062] One embodiment of the above invention has the following advantages or beneficial effects: the present application obtains user context data and user historical browsing item data according to the received data recommendation request; based on the user historical browsing item data, obtains a pooling vector through a pooling network, and obtains a deep interest network representation vector through a deep interest network according to the user historical browsing item data and the item data to be recommended; based on the preset representation dimension and the preset real number space configuration data, determines the first mapping matrix and the second mapping matrix, and then determines the weighted proportional coefficient based on the user context data; determines the target interest vector according to the weighted proportional coefficient, the pooling vector and the deep interest network representation vector; determines the target recommended item data in the item data to be recommended based on the target interest vector. It can effectively reflect the user interest in the user context session of the recommendation system, improve the performance of the recommendation system, realize personalized recommendation, and improve the efficiency and accuracy of personalized data recommendation.

[0063] The further effects of the above-mentioned non-conventional optional manner will be described below in conjunction with the specific implementation manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] The accompanying drawings are used to better understand the present application and do not constitute an improper limitation on the present application.

[0065] Figure 1 is a schematic diagram of the main process of a data recommendation method according to an embodiment of the present application;

[0066] Figure 2 is a schematic diagram of the main process of a data recommendation method according to an embodiment of the present application;

[0067] Figure 3 It is a schematic diagram of the main process of a data recommendation method according to an embodiment of the present application;

[0068] Figure 4 is a schematic diagram of main units of a data recommendation device according to an embodiment of the present application;

[0069] Figure 5 is an exemplary system architecture diagram to which the embodiments of the present application can be applied;

[0070] Figure 6 It is a structural diagram of a computer system of a terminal device or server suitable for implementing an embodiment of the present application. DETAILED DESCRIPTION

[0071] The following is an explanation of the exemplary embodiments of the present application in conjunction with the accompanying drawings, including various details of the embodiments of the present application to facilitate understanding, which should be considered as merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. Similarly, for the sake of clarity and conciseness, the description of known functions and structures is omitted in the following description. It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of the present application are in compliance with the relevant provisions of national laws and regulations. It should be noted that in the embodiments of the present application, certain software, components, models, and other existing solutions in the industry may be mentioned, which should be considered as exemplary, and their purpose is only to illustrate the feasibility of the implementation of the technical solution of the present application, but it does not mean that the applicant has or must have used the solution. In the technical solution of the present application, the collection, collection, update, analysis, processing, use, transmission, storage, etc. of the user personal information involved are in compliance with the provisions of relevant laws and regulations, are used for legal and reasonable purposes, and do not violate public order and good customs, are not shared, leaked or sold outside these legal uses, and are subject to supervision and management by regulatory authorities. Take necessary measures to prevent illegal access to such user personal information data, maintain the security of user personal information, network security and national security, ensure that persons who have access to personal information data comply with relevant laws and regulations, and ensure the security of user personal information. Once such user personal information data is no longer needed, the risk should be minimized by limiting or even prohibiting data collection and / or deleting data.

[0072] When used, including in certain related applications, user privacy is protected by de-identifying data, such as by removing specific identifiers when used, controlling the amount or specificity of data stored, controlling how data is stored, and / or other methods of de-identification.

[0073] Figure 1 is a schematic diagram of the main process of the data recommendation method according to an embodiment of the present application, such as Figure 1 As shown, the data recommendation method mainly includes the following steps S101 to S105.

[0074] Step S101: acquiring user context data and user historical browsing item data according to the received data recommendation request.

[0075] In this embodiment, the execution subject of the data recommendation method (for example, a server) can receive the data recommendation request through a wired connection or a wireless connection. After receiving the data recommendation request, the execution subject can determine the user context data and the user's historical browsing item data corresponding to the request. Among them, the user context data may include user geographic location data and usage time data, etc., and user context features (such as user geographic location, usage time, etc.) can be extracted from the user context data. The user's historical browsing item data may include the user's historical clicked and browsed item sequence.

[0076] Step S102: Based on the user's historical browsing data of items, a pooling vector is obtained through a pooling network, and based on the user's historical browsing data of items and the data of items to be recommended, a deep interest network representation vector is obtained through a deep interest network.

[0077] Specifically, based on the user's historical item browsing data, a pooling vector is obtained through a pooling network, including: based on the user's historical item browsing data, a corresponding historical item browsing vector is determined; and each historical item browsing vector is summed up through a pooling network to obtain a pooling vector.

[0078] The execution subject can convert the corresponding browsed item sequence into the corresponding historical browsed item vector based on the user's historical browsed item data through word embedding, and transform each historical browsed item vector e into a vector through a pooling network. i (i ranges from 1 to n) sum to accurately obtain the pooling vector E P ,

[0079] Specifically, a deep interest network representation vector is obtained through a deep interest network, including: calculating the similarity between the user's historical browsed item data and the to-be-recommended item data through a deep interest network (i.e., Deep Interest Network, DIN) to obtain a similarity score; normalizing the similarity score to obtain a normalized similarity score; and obtaining a deep interest network representation vector based on the normalized similarity score and a user sequence feature vector corresponding to the corresponding user's historical browsed item data.

[0080] The cosine similarity between the user's historical browsing item data and the recommended item data is calculated through the Deep Interest Network (DIN) to obtain the cosine similarity score; the cosine similarity score s is converted into the value by the normalized exponential function, i.e., the softmax function. i Normalize to get the normalized similarity score s i ,s i = sofrmax(s i); Based on the normalized similarity score s i = sofrmax(s i ) and the user sequence feature vector e corresponding to the corresponding user historical browsing item data i , e i YesE seq Any feature vector in can represent the feature vector corresponding to the i-th item, n is the length of the user's browsing sequence, indicating the number of items the user has browsed historically, and the deep interest network representation vector is obtained.

[0081] Step S103, determining a first mapping matrix and a second mapping matrix based on a preset representation dimension and preset real number space configuration data, and further determining a weighted proportional coefficient based on user context data.

[0082] Specifically, determining the first mapping matrix and the second mapping matrix includes: determining a first parameter value based on a preset representation dimension; obtaining shape data and dimension data in preset real space configuration data, and determining a second parameter value based on the shape data and the dimension data; determining the dimension and shape of the mapping matrix based on the first parameter value and the second parameter value, and then obtaining the corresponding first mapping matrix and second mapping matrix.

[0083] The preset representation dimension can be, for example, the user sequence feature vector e i and the recommended item vector E item The representation dimension can be 32 dimensions or 64 dimensions or other dimensions, which can be determined according to the dimensions supported by the device used. The larger the dimension, the richer the meaning expressed by the corresponding vector. For example, the weighted proportional coefficient α = σ(H(WE context )), where W∈R d×d , H∈R 1×d , σ is the sigmoid function (function used for normalization), E context is the feature vector obtained after the user context feature is embedded in the input layer. Wherein, d can be the first parameter value of the embodiment of the present application, which is used to characterize the user sequence feature vector e i and the recommended item vector E item The representation dimension can be 32 dimensions, 64 dimensions or other dimensions, which can be determined according to the dimensions supported by the device used. The larger the dimension, the richer the meaning expressed by the corresponding vector.

[0084] Obtain shape data (e.g., d×d, 1×d) and dimension data in preset real space configuration data, wherein the dimension data may include the dimension of the vector contained in the mapping matrix and the dimension of the real space where the mapping matrix is ​​located. The dimension of the real space where the mapping matrix is ​​located can be determined based on the number of “×” symbols. Specifically, the number of “×” symbols plus one can be the dimension of the real space where the mapping matrix is ​​located. According to the shape data and the dimension data, determine the second parameter value, such as a sequence of integer pairs (1,1), (0,1) consisting of the coefficients of d.

[0085] The dimension of the mapping matrix may include the dimension of the vector contained in the mapping matrix and the dimension of the real space where the mapping matrix is ​​located. Then, the dimension of the vector contained in the mapping matrix can be determined according to the value of d (i.e., the first parameter value), and the dimension of the real space where the mapping matrix is ​​located and the shape of the mapping matrix can be determined according to the sequence of integer pairs formed by the coefficients of d (i.e., the second parameter value). The sequence of integer pairs formed by the coefficients of d can be the second parameter value of the embodiment of the present application. For example, when the sequence of integer pairs formed by the coefficients of d is (1,1), (0,1), it means that the mapping matrix is ​​in a two-dimensional real space and the shapes of the mapping matrix are d×d and 1×d respectively. According to the determined dimension and shape of the mapping matrix, the corresponding first mapping matrix can be accurately determined. For example, W∈R d×d and the second mapping matrix, such as H∈R 1×d .

[0086] A weighted proportionality coefficient is determined according to the determined first mapping matrix and the second mapping matrix and the user context data.

[0087] Step S104, determining a target interest vector according to the weighted proportional coefficient, the pooling vector and the deep interest network representation vector.

[0088] Specifically, determining the target interest vector includes: calculating the pooling vector E P The first product of the corresponding weighted proportional coefficient 1-α is (1-α)×E P And the deep interest network representation vector E D The second product α×E with the corresponding weighted proportional coefficient α D ; The sum of the first product and the second product is determined as the target interest vector E int =α×E D +(1-α)×E P .

[0089] The embodiment of the present application uses a self-attention mechanism (context attention) to process user context information (corresponding to E context ), and obtain the sequence vector weighted proportional coefficients α and 1-α with a value range of 0 to 1, so as to fuse the DIN vector (i.e., ED ) and Pooling vector (i.e. E P ), as the final interest vector of the user behavior sequence, that is, the target interest vector E int . Among them, α=σ(H(WE context )), E int =α×E D +(1-α)×E P . Among them, W∈R d×d , H∈R 1×d , σ is the sigmoid function (function used for normalization).

[0090] Step S105: determining target recommended item data in the to-be-recommended item data based on the target interest vector.

[0091] Specifically, based on the target interest vector, determining the target recommended item data in the to-be-recommended item data includes: transforming the target interest vector E int Input into the recommendation system deep learning network model M to output the estimated user browsing probability of the corresponding recommended item data Based on browsing probability Determine the target recommended item data.

[0092] For example, E int 、E seq 、E contxt 、E item 、E user Input the recommendation system deep learning network model M to obtain the estimated user browsing probability of the corresponding recommended item data Among them, E int =α×E D +(1-α)×E P , α=σ(H(WE context )), where W∈R d×d , H∈R 1×d ,σ is the sigmoid function, Among them, s i =H(W[e i ||e i -E item ||e i -E item ||E item ]), where W∈R d×4d , H∈R 1×d ,s i =softmax(s i ), E seq ={e1, e2, ..., e n},e i YesE seq Any feature vector in can represent the feature vector corresponding to the i-th item. The recommended item data corresponding to the browsing probability greater than the preset threshold is determined as the target recommended item data to improve the accuracy of determining the target recommended item data. Specifically, the loss function of the deep learning network model M of the recommendation system: The embodiment of the present application adopts the cross entropy loss function as the objective function and uses the Adam gradient descent method to optimize the network parameters. in, is the estimated browsing probability of the user, y is the real label of whether the user browses the item. For example, the item browsing record data of the user in the last 30 days is obtained. If the user browses the target item, y=1, if the user does not browse the target item, y=0, wherein the target item can be any item corresponding to the item browsing record data of the user in the last 30 days (can be any number of days, which is only an example here and is not specifically limited).

[0093] This embodiment obtains user context data and user historical browsing item data according to the received data recommendation request; obtains pooling vectors through a pooling network based on the user historical browsing item data; obtains a deep interest network representation vector through a deep interest network based on the user historical browsing item data and the item data to be recommended; determines a first mapping matrix and a second mapping matrix based on a preset representation dimension and preset real number space configuration data, and then determines a weighted proportional coefficient based on the user context data; determines a target interest vector based on the weighted proportional coefficient, the pooling vector and the deep interest network representation vector; and determines the target recommended item data in the item data to be recommended based on the target interest vector. It can effectively reflect the user interests in the user context session of the recommendation system, improve the performance of the recommendation system, realize personalized recommendations, and improve the efficiency and accuracy of personalized data recommendations.

[0094] Figure 2 is a schematic diagram of the main flow of a data recommendation method according to an embodiment of the present application. Figure 2 As shown, the data recommendation method mainly includes the following steps S201 to S208.

[0095] Step S201: Obtain user context data and user historical browsing item data according to the received data recommendation request.

[0096] User context data may include user geographic location data and usage time data, etc. User context features (such as user geographic location, usage time, etc.) can be extracted from user context data. User historical browsing item data may include the sequence of items that the user has clicked and browsed in the past. Based on the sequence of items that the user has browsed in the past, items that the user has not browsed and are similar to the browsing history are recommended, so as to capture the user's personalized interests and better serve the user's life.

[0097] Step S202: Based on the user's historical browsing data of items, a pooling vector is obtained through a pooling network, and based on the user's historical browsing data of items and the data of items to be recommended, a deep interest network representation vector is obtained through a deep interest network.

[0098] The pooling network is the Pooling network, and the deep interest network is the Deep Interest Network, DIN network. The pooling vector is accurately obtained through the Pooling network, and the deep interest network representation vector is accurately obtained through the DIN network based on the user's historical browsing item data and the data of the items to be recommended.

[0099] Step S203: determining a first mapping matrix and a second mapping matrix based on a preset representation dimension and preset real number space configuration data.

[0100] The preset representation dimension can be, for example, the user sequence feature vector e i and the recommended item vector E item The representation dimension can be 32 dimensions or 64 dimensions or other dimensions, which can be accurately determined based on the recommended portrait of historical data. The preset real space configuration data may include symbolic data, wherein the symbolic data may include a multiplication sign "×" and brackets "()", and the number of multiplication signs "×" in the preset real space configuration data is obtained. The dimension of the real space where the mapping matrix is ​​located is determined according to the number of multiplication signs "×", the number of mapping matrices is determined according to the number of brackets "()" (for example, 2), and the shape of the mapping matrix is ​​determined according to the data in the brackets "()", such as 1,1, 0,1. According to the preset representation dimension, the dimension of the real space where the mapping matrix is ​​located, and the shape and number of the mapping matrix (for example, 2), the first mapping matrix and the second mapping matrix are accurately determined.

[0101] Step S204: convert the user context data into a user context data vector, and multiply the first mapping matrix by the user context data vector to obtain first calculation result data.

[0102] The user context data vector is E contex t, the first mapping matrix such as W∈R d×d , multiply the first mapping matrix by the user context data vector to obtain the first calculation result data, for example .

[0103] Step S205: multiply the second mapping matrix by the first calculation result data to obtain second calculation result data.

[0104] The second mapping matrix is ​​H∈R 1×d , multiply the second mapping matrix by the first calculation result data to obtain the second calculation result data, for example

[0105] Step S206, normalizing the second calculation result data to obtain a weighted proportional coefficient.

[0106] The second calculation result data Normalization is performed to obtain the weighted proportional coefficient α = σ(H(WE context )), σ is the sigmoid function.

[0107] Step S207, determining a target interest vector according to the weighted proportional coefficient, the pooling vector and the deep interest network representation vector.

[0108] According to the weighted proportionality coefficient α = σ(H(WE context )), Pooling vector E P and the deep interest network representation vector E D , determine the target interest vector E int =α×E D +(1-α)×E P

[0109] Step S208: determining target recommended item data in the to-be-recommended item data based on the target interest vector.

[0110] The to-be-recommended item data corresponding to the browsing probability greater than a preset threshold is determined as the target recommended item data, so as to improve the accuracy of determining the target recommended item data.

[0111] Figure 3 1 is a schematic diagram of a deep learning network model structure of a recommendation system of a data recommendation method according to an embodiment of the present application. Figure 3 As shown in the figure, in the recommendation system, the features required by the deep learning network model of the recommendation system are divided into four categories: user-side attribute features F user (e.g., user age, gender, etc.), user context features F context (e.g., user geographic location, usage time, etc.), item attribute features Fx tem (e.g., item price, duration, etc.), user sequence feature F seq(e.g., the sequence of items that users click and browse), where items generally refer to the products, videos, pictures and text posts recommended by the system. In the embodiment of the present application, user context features (e.g., user geographic location, usage time, etc.) can be obtained based on user context data, corresponding to E context Based on the user's historical browsing data, we can get the user sequence feature (the sequence of items clicked and browsed by the user), which corresponds to E seq Based on the items to be recommended, item attribute features can be obtained. Based on the user attribute data, user-side attribute features can be obtained. The user sequence feature vector corresponds to the user's historical behavior features on the items. The input layer of the recommendation system embedding layer combines these features (for example, user-side attribute features F user (e.g., user age, gender, etc.), user context features F context (e.g., user geographic location, usage time, etc.), item attribute features F item (e.g., item price, duration, etc.), user sequence feature F seq (For example, the sequence of items that users click and browse) is transformed from a discrete high-dimensional space to a continuous low-dimensional space through embedding mapping. For example, e = emb(f onehot ).

[0112] User-side attribute feature F user , user context feature F context , Item Attributes F item , user sequence feature F seq After the embedding mapping operation in the input layer Embedding Layer, we get E user , E context , E item , E seq , and then use the recommendation system deep learning network model M to learn to estimate the user's browsing probability of this item is the estimated browsing probability of users. The deep learning network model M of the recommendation system can be any model, such as the classic DeepFM, DCN, Autoint and other models.

[0113] Context information enhanced sequence network, that is, the deep learning network model M of the recommendation system in the embodiment of the present application. The user's context information (including the user's geographic location, browsing time and other information) will affect the user's expression of interest, and the user's interest preferences will be quite different in different context information environments. Whether it is the DIN network or the Pooling network, they all ignore the impact of the user's context information on the user's expression of interest, and cannot achieve truly personalized recommendations. The embodiment of the present application uses the user context information enhanced sequence network, that is, the deep learning network model M of the recommendation system, which can effectively reflect the user's interest in the user context session of the recommendation system.

[0114] Specifically, the specific structure and mechanism of the Pooling network and DIN network are as follows:

[0115] (1) Pooling network: Assume that the user sequence feature vector is E seq ={e1, e2, ..., e n}, the feature vectors are sorted by the user's historical behavior time, where n is the length of the user's browsing sequence, indicating the number of items the user has browsed historically. In the embodiment of the present application, the Pooling network adopts the most classic sum pooling method, and the sum of the embeddings of each item is expressed to obtain E P , Among them, e i Represents the feature vector corresponding to the i-th item. This method is simple and intuitive, with the advantages of small computational complexity and fast inference speed. However, the Pooling network assigns the same weight to each element in the sequence, which cannot capture the complex relationship and dependency between the elements in the sequence, and the interest modeling of different items to be estimated is not detailed enough, and cannot well capture the personalized needs of users.

[0116] (2) DIN network, namely Deep Interest Network (DIN): DIN network first transforms each user sequence feature vector e i and the recommended item vector E item Perform target attention calculation to get the similarity score s i :s i =H(W[e i ||e i -E item ||e i -E item ||E item ]). Where W∈R d×4d , H∈R 1×d, R represents two-dimensional space, W means a matrix with a shape of d×4d in two-dimensional space, H means a matrix with a shape of 1×d in two-dimensional space, and d is the user sequence feature vector e i and the recommended item vector E item The representation dimension of can be 32 dimensions or 64 dimensions or other dimensions, which can be determined according to the dimensions supported by the device used. The larger the dimension, the richer the meaning of the corresponding vector expression. Then, the similarity score s is converted into i Normalization gets the normalized s i , and finally use the normalized s i Multiply by the corresponding user sequence feature vector e i , and then sum to get the DIN representation vector (ie, deep interest network representation vector) E D Specifically, the normalized s i =softmax(s i ),

[0117] The DIN network uses an attention mechanism to assign different weights to each element in the sequence to personalize the model, but it is biased against the majority of positive historical samples of users. The DIN model and its various subsequent variant models are centered around the target item feature vector (for example, the recommended item vector in the embodiment of the present application) E item and historical behavior feature vector (eg, user sequence feature vector in the embodiment of the present application) i The similarity of the user is used to model the user's interests to predict whether the user will interact with the target item. However, the user's contextual information is also a key factor affecting the user's interests and decisions. The user's interests will change significantly in different contextual information environments. For example, in a short session time, the user tends to watch the items with the highest historical browsing frequency, while in a long session, repeated similar items are likely to cause the user's system fatigue, and the user will gradually tend to watch items with less historical browsing frequency or items that have never been seen. It can be found that the user's contextual information (including the user's geographic location, browsing time, etc.) will affect the user's browsing interest selection. Therefore, the embodiment of the present application adopts a self-attention mechanism (context attention) to process the user's contextual information (corresponding to E context ), and obtain the sequence vector weighted proportional coefficients α and 1-α with a value range of 0 to 1, so as to fuse the DIN vector (i.e., E D ) and Pooling vector (i.e. E P ), as the final interest vector of the user behavior sequence, that is, the target interest vector E int . Among them, α=σ(H(WE context)), E int =α×E D +(1-α)×E P . Among them, W∈R d×d , H∈R 1×d , σ is the sigmoid function (function used for normalization). int 、E seq 、E context 、E item 、E user Input the recommendation system deep learning network model M to get the estimated user browsing probability Among them, E int =α×E D +(1-α)×E P , α=σ(H(WE context )), where W∈R d×d , H∈R 1×d ,σ is the sigmoid function, Among them, s i =H(W[e i ||e i -E item ||e i -E item ||E item ]), where W∈R d×4d , H∈R 1×d ,s i =softmax(s i ), E seq ={e1, e2, ..., e n},e i YesE seq Any feature vector in can represent the feature vector corresponding to the i-th item. For example, the loss function of the deep learning network model M of the recommendation system: The embodiment of the present application adopts the cross entropy loss function as the objective function and uses the Adam gradient descent method to optimize the network parameters. in, is the estimated browsing probability of the user, y is the real label of whether the user browses the item. For example, the item browsing record data of the user in the last 30 days is obtained. If the user browses the target item, y=1, if the user does not browse the target item, y=0, wherein the target item can be any item corresponding to the item browsing record data of the user in the last 30 days (can be any number of days, which is only an example here and is not specifically limited).

[0118] In the embodiment of the present application, DIN network: Deep interest network, deep interest network in the recommendation system; Pooling network: pooling network in deep learning; Embedding: feature representation vector. The embodiment of the present application can realize serialization based on user context perception, use the self-attention mechanism to process the user's context information, obtain the weighted proportional coefficient of the Pooling interest vector and the DIN interest vector, and thus fuse the two interest vectors as the user's personalized interest vector expression, which can effectively reflect the user's interest in the user context session of the recommendation system.

[0119] The embodiment of the present application combines the characteristics of the Pooling network that fairly assigns the same weight to each sequence element and the DIN network that assigns different personalized weights to each element in the sequence according to the items to be estimated, combined with the fact that the contextual information of different users (including the user's geographic location, browsing time, etc.) will affect the user's expression of interest. The embodiment of the present application uses a sequence recommendation model based on user context perception and a self-attention mechanism to process the user's context information to obtain the weighted proportional coefficient of the Pooling interest vector and the DIN interest vector, thereby fusing the two interest vectors as the user's personalized interest vector expression, which can effectively reflect the user's interest in the user context session of the recommendation system, improve the performance of the recommendation system, and significantly improve the user experience.

[0120] Figure 4 is a schematic diagram of the main units of the data recommendation device according to an embodiment of the present application. Figure 4 As shown, the data recommendation device 400 includes an acquisition unit 401, a vector conversion unit 402, a weighted proportional coefficient determination unit 403, a target interest vector determination unit 404 and a target recommended item data determination unit 405.

[0121] The acquisition unit 401 is configured to acquire user context data and user historical browsing item data according to the received data recommendation request.

[0122] The vector conversion unit 402 is configured to obtain a pooling vector through a pooling network based on the user's historical browsing item data, and obtain a deep interest network representation vector through a deep interest network based on the user's historical browsing item data and the item data to be recommended.

[0123] The weighted proportionality coefficient determination unit 403 is configured to determine the first mapping matrix and the second mapping matrix based on the preset representation dimension and the preset real number space configuration data, and further determine the weighted proportionality coefficient based on the user context data.

[0124] The target interest vector determining unit 404 is configured to determine the target interest vector according to the weighted proportionality coefficient, the pooling vector and the deep interest network representation vector.

[0125] The target recommended item data determining unit 405 is configured to determine target recommended item data in the to-be-recommended item data based on the target interest vector.

[0126] In some embodiments, the vector conversion unit 402 is further configured to: determine a corresponding historically browsed item vector based on the user's historically browsed item data; and sum each historically browsed item vector through a pooling network to obtain a pooled vector.

[0127] In some embodiments, the vector conversion unit 402 is further configured to: calculate the similarity between the user's historical browsed item data and the to-be-recommended item data through a deep interest network to obtain a similarity score; normalize the similarity score to obtain a normalized similarity score; and obtain a deep interest network representation vector based on the normalized similarity score and the user sequence feature vector corresponding to the corresponding user's historical browsed item data.

[0128] In some embodiments, the weighted proportional coefficient determination unit 403 is further configured to: determine a first parameter value based on a preset representation dimension; obtain shape data and dimension data in the preset real space configuration data, and determine a second parameter value based on the shape data and dimension data; determine the dimension and shape of the mapping matrix based on the first parameter value and the second parameter value, and then obtain the corresponding first mapping matrix and second mapping matrix.

[0129] In some embodiments, the weighted proportional coefficient determination unit 403 is further configured to: convert the user context data into a user context data vector, multiply the first mapping matrix by the user context data vector on the left to obtain first calculation result data; multiply the second mapping matrix by the first calculation result data on the left to obtain second calculation result data; and normalize the second calculation result data to obtain the weighted proportional coefficient.

[0130] In some embodiments, the target interest vector determination unit 404 is further configured to: calculate a first product of the pooling vector and the corresponding weighted proportional coefficient and a second product of the deep interest network representation vector and the corresponding weighted proportional coefficient; and determine the sum of the first product and the second product as the target interest vector.

[0131] In some embodiments, the target recommended item data determination unit 405 is further configured to: input the target interest vector into the recommendation system deep learning network model to output the estimated user browsing probability of the corresponding item data to be recommended; and determine the target recommended item data based on the browsing probability.

[0132] It should be noted that the data recommendation method and the data recommendation device of the present application have a corresponding relationship in terms of specific implementation contents, so the repeated contents will not be described again.

[0133] Figure 5 An exemplary system architecture 500 is shown to which the data recommendation method or data recommendation device according to the embodiments of the present application can be applied.

[0134] like Figure 5 As shown, system architecture 500 may include terminal devices 501, 502, 503, a network 504 and a server 505. Network 504 is used to provide a medium for communication links between terminal devices 501, 502, 503 and server 505. Network 504 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0135] Users can use terminal devices 501, 502, 503 to interact with server 505 through network 504 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 501, 502, 503, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only as examples).

[0136] The terminal devices 501 , 502 , and 503 may be various electronic devices having a data recommendation processing screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, desktop computers, and the like.

[0137] The server 505 may be a server that provides various services, such as a background management server that supports the data recommendation request submitted by the user using the terminal devices 501, 502, and 503 (only as an example). The background management server may obtain the user context data and the user's historical browsing item data according to the received data recommendation request; based on the user's historical browsing item data, obtain the pooling vector through the pooling network; based on the user's historical browsing item data and the item data to be recommended, obtain the deep interest network representation vector through the deep interest network; based on the preset representation dimension and the preset real number space configuration data, determine the first mapping matrix and the second mapping matrix, and then determine the weighted proportional coefficient based on the user context data; determine the target interest vector according to the weighted proportional coefficient, the pooling vector and the deep interest network representation vector; determine the target recommended item data in the item data to be recommended based on the target interest vector. It can effectively reflect the user interest in the user context session of the recommendation system, improve the performance of the recommendation system, realize personalized recommendation, and improve the efficiency and accuracy of personalized data recommendation.

[0138] It should be noted that the data recommendation method provided in the embodiment of the present application is generally executed by the server 505 , and accordingly, the data recommendation device is generally arranged in the server 505 .

[0139] It should be understood that Figure 5 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0140] Reference below Figure 6 , which shows a schematic diagram of the structure of a computer system 600 of a terminal device suitable for implementing an embodiment of the present application. Figure 6 The terminal device shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0141] like Figure 6 As shown, the computer system 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage part 608 into a random access memory (RAM) 603. In the RAM 603, various programs and data required for the operation of the computer system 600 are also stored. The CPU 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0142] The following components are connected to the I / O interface 605: an input section 606 including a keyboard, a mouse, etc.; an output section 607 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as needed. A removable medium 611, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 610 as needed, so that a computer program read therefrom is installed into the storage section 608 as needed.

[0143] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication part 609, and / or installed from the removable medium 611. When the computer program is executed by the central processing unit (CPU) 601, the above-mentioned functions defined in the system of the present application are executed.

[0144] It should be noted that the computer-readable medium shown in the present application may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. Computer-readable storage media may include, for example, but are not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present application, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0145] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the above-mentioned module, program segment or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0146] The units involved in the embodiments described in the present application may be implemented by software or hardware. The units described may also be set in a processor, for example, it may be described as: a processor includes an acquisition unit, a vector conversion unit, a weighted proportional coefficient determination unit, a target interest vector determination unit, and a target recommended item data determination unit. The names of these units do not, in some cases, constitute limitations on the units themselves.

[0147] As another aspect, the present application also provides a computer-readable medium, which may be included in the device described in the above embodiment; or it may exist independently and not be assembled into the device. The above computer-readable medium carries one or more programs. When the above one or more programs are executed by a device, the device obtains user context data and user historical browsing item data according to the received data recommendation request; based on the user historical browsing item data, a pooling vector is obtained through a pooling network, and based on the user historical browsing item data and the item data to be recommended, a deep interest network representation vector is obtained through a deep interest network; based on the preset representation dimension and the preset real space configuration data, a first mapping matrix and a second mapping matrix are determined, and then based on the user context data, a weighted proportional coefficient is determined; based on the weighted proportional coefficient, the pooling vector and the deep interest network representation vector, a target interest vector is determined; based on the target interest vector, the target recommended item data in the item data to be recommended is determined.

[0148] The computer program product of the present application includes a computer program, and when the computer program is executed by a processor, the data recommendation method in the embodiment of the present application is implemented.

[0149] According to the technical solution of the embodiment of the present application, it is possible to effectively reflect the user interests in the user context session of the recommendation system, improve the performance of the recommendation system, realize personalized recommendations, and improve the efficiency and accuracy of data personalized recommendations.

[0150] The above specific implementations do not constitute a limitation on the protection scope of this application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of this application should be included in the protection scope of this application.

Claims

1. A data recommendation method, characterized in that: include: Obtain user context data and user historical browsing item data according to the received data recommendation request; Based on the user's historical browsing data of items, a pooling vector is obtained through a pooling network, and based on the user's historical browsing data of items and the data of items to be recommended, a deep interest network representation vector is obtained through a deep interest network; Determine a first mapping matrix and a second mapping matrix based on a preset representation dimension and preset real number space configuration data, and further determine a weighted proportional coefficient based on the user context data; Determining a target interest vector according to the weighted proportional coefficient, the pooling vector and the deep interest network representation vector; Based on the target interest vector, target recommended item data in the to-be-recommended item data is determined.

2. The method according to claim 1, characterized in that: The step of obtaining a pooling vector through a pooling network based on the user's historical browsing item data includes: Based on the user's historically browsed item data, determining a corresponding historically browsed item vector; The pooling vector is obtained by summing up each of the historical browsed item vectors through a pooling network.

3. The method according to claim 1, characterized in that The step of obtaining a deep interest network representation vector through a deep interest network includes: Calculating similarity between the user's historical browsing item data and the to-be-recommended item data through a deep interest network to obtain a similarity score; Normalizing the similarity scores to obtain normalized similarity scores; Based on the normalized similarity score and the user sequence feature vector corresponding to the corresponding user historical browsing item data, a deep interest network representation vector is obtained.

4. The method according to claim 1, characterized in that The determining of the first mapping matrix and the second mapping matrix comprises: Based on the preset representation dimension, determining a first parameter value; Acquire shape data and dimension data in the preset real number space configuration data, and determine a second parameter value according to the shape data and the dimension data; The dimension and shape of the mapping matrix are determined according to the first parameter value and the second parameter value, thereby obtaining the corresponding first mapping matrix and second mapping matrix.

5. The method according to claim 1, characterized in that The determining of a weighted proportionality coefficient based on the user context data includes: Convert the user context data into a user context data vector, and multiply the first mapping matrix by the user context data vector to obtain first calculation result data; Multiplying the second mapping matrix by the first calculation result data on the left to obtain second calculation result data; The second calculation result data is normalized to obtain a weighted proportional coefficient.

6. The method according to claim 1, characterized in that The determining of the target interest vector includes: Calculate a first product of the pooling vector and a corresponding weighted proportional coefficient and a second product of the deep interest network representation vector and a corresponding weighted proportional coefficient; A sum of the first product and the second product is determined as a target interest vector.

7. The method according to any one of claims 1 to 6, characterized in that: The determining, based on the target interest vector, target recommended item data in the to-be-recommended item data comprises: Inputting the target interest vector into the deep learning network model of the recommendation system to output the estimated browsing probability of the user for the corresponding item data to be recommended; Based on the browsing probability, target recommended item data is determined.

8. A data recommendation device, characterized in that: include: an acquisition unit, configured to acquire user context data and user historical browsing item data according to the received data recommendation request; A vector conversion unit is configured to obtain a pooling vector through a pooling network based on the user's historical browsing item data, and obtain a deep interest network representation vector through a deep interest network based on the user's historical browsing item data and the item data to be recommended; A weighted proportionality coefficient determination unit is configured to determine a first mapping matrix and a second mapping matrix based on a preset representation dimension and preset real number space configuration data, and further determine a weighted proportionality coefficient based on the user context data; a target interest vector determining unit, configured to determine a target interest vector according to the weighted proportional coefficient, the pooling vector and the deep interest network representation vector; The target recommended item data determining unit is configured to determine the target recommended item data in the to-be-recommended item data based on the target interest vector.

9. A data recommendation electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

10. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

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

Images