User recommendation method, device, electronic device and computer storage medium

Abstract

The embodiment of the present application provides a user recommendation method, device, electronic device and computer storage medium, the method comprising: determining the user product attribute value data of the user to be recommended and the store product attribute value data of the target store; using a preset product feature model to perform feature analysis on the user product attribute value data and the store product attribute value data respectively to obtain user feature information and store feature information; performing similarity calculation on the user feature information and the store feature information, determining a target user set from the users to be recommended according to the calculation result, and recommending the target user set to the target store. In this way, by determining the user feature information and the store feature information through the preset product feature model, the correlation between user behavior and store products can be accurately calculated, and then the target user set recommended to the target store can be determined, thereby improving the accuracy of user recommendation.

Description

Technical Field

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

[0002] In sales scenarios, stores often need to develop different marketing strategies for different user groups to achieve overall business growth. Therefore, for e-commerce platforms, recommending potential new users (also called new customers) to stores is an important part of helping stores achieve user growth, especially for unpopular stores with low sales.

[0003] In the related art, a collaborative filtering algorithm can be used to find similar users of old store users (or old users of similar stores), and the found users can be used as potential new users; a prediction model can also be established for users and stores to predict potential new users. However, the technical solutions provided by the above related technologies have the problems of large computational complexity and low prediction accuracy, and the actual application effect cannot meet expectations. Summary of the invention

[0004] The present application provides a user recommendation method, device, electronic device and computer storage medium, which generates store feature information and user feature information by presetting a product feature model, and can accurately calculate the correlation between user behavior and store products, thereby improving the prediction accuracy.

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

[0006] In a first aspect, an embodiment of the present application provides a user recommendation method, the method comprising:

[0007] Determine the user product attribute value data of the user to be recommended and the store product attribute value data of the target store;

[0008] Using the preset product feature model, feature analysis is performed on the user product attribute value data and the store product attribute value data to obtain user feature information and store feature information;

[0009] The similarity between the user feature information and the store feature information is calculated, and the target user set is determined from the users to be recommended based on the calculation results, and the target user set is recommended to the target store.

[0010] In a second aspect, an embodiment of the present application provides a user recommendation device, which includes a determination unit, an analysis unit, and a recommendation unit; wherein:

[0011] A determination unit configured to determine the user commodity attribute value data of the user to be recommended and the store commodity attribute value data of the target store;

[0012] An analysis unit configured to perform feature analysis on the user's product attribute value data and the store's product attribute value data respectively using a preset product feature model to obtain user feature information and store feature information;

[0013] The recommendation unit is configured to perform similarity calculation on the user feature information and the store feature information, determine a target user set from the users to be recommended according to the calculation result, and recommend the target user set to the target store.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, the electronic device comprising a memory and a processor; wherein:

[0015] A memory for storing computer programs that can be run on the processor;

[0016] A processor is used to execute the steps of the method of the first aspect when running a computer program.

[0017] In a fourth aspect, an embodiment of the present application provides a computer storage medium, which stores a computer program, and when the computer program is executed, the steps of the method of the first aspect are implemented.

[0018] The embodiment of the present application provides a user recommendation method, device, electronic device and computer storage medium, which determines the user product attribute value data of the user to be recommended and the store product attribute value data of the target store; uses the preset product feature model to perform feature analysis on the user product attribute value data and the store product attribute value data respectively to obtain user feature information and store feature information; performs similarity calculation on the user feature information and the store feature information, determines the target user set from the users to be recommended according to the calculation results, and recommends the target user set to the target store. In this way, by determining the user feature information and the store feature information through the preset product feature model, the correlation between the user behavior and the store product can be accurately calculated, so as to determine the potential new users recommended to the target store; at the same time, the technical solution of the present application can realize the update of the user feature information and the store feature information only by updating the preset product feature model, and then can update the potential new users recommended to the target store in real time, thereby not only improving the accuracy of user recommendation, but also ensuring the real-time nature of user recommendation, thereby improving the operation efficiency of the store. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 A flowchart of a user recommendation method provided in an embodiment of the present application;

[0020] Figure 2 A flowchart of another user recommendation method provided in an embodiment of the present application;

[0021] Figure 3A schematic diagram of the working process of a user recommendation method provided in an embodiment of the present application;

[0022] Figure 4 A flowchart of another user recommendation method provided in an embodiment of the present application;

[0023] Figure 5 A schematic diagram of the working process of another user recommendation method provided in an embodiment of the present application;

[0024] Figure 6 A schematic diagram of the working process of another user recommendation method provided in an embodiment of the present application;

[0025] Figure 7 A schematic diagram of the working process of another user recommendation method provided in an embodiment of the present application;

[0026] Figure 8 A schematic diagram of the working process of another user recommendation method provided in an embodiment of the present application;

[0027] Fig. 9 A schematic diagram of the working process of another user recommendation method provided in an embodiment of the present application;

[0028] Fig.10 A schematic diagram of the structure of a user recommendation device provided in an embodiment of the present application;

[0029] Fig.11 A schematic diagram of the composition structure of another user recommendation device provided in an embodiment of the present application;

[0030] Fig.12 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application;

[0031] Fig.13 A schematic diagram of the structure of another electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0032] The following will be combined with the accompanying drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. It is understood that the specific embodiments described herein are only used to explain the related applications, rather than to limit the applications. It should also be noted that, for the convenience of description, only the parts related to the related applications are shown in the accompanying drawings.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0034] In the following description, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0035] It should be pointed out that the terms "first\second\third" involved in the embodiments of the present application are only used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that "first\second\third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0036] Before further describing the embodiments of the present application in detail, the related technologies involved in the embodiments of the present application are first described:

[0037] In the field of Internet e-commerce, stores often need to develop different marketing strategies for different user (or customer) groups to achieve overall business growth. Therefore, recommending potential new users to stores (a process also known as store acquisition) is an important part of helping stores achieve user growth.

[0038] In particular, for unpopular stores with low sales, there are few users who have interacted with them, so the number of potential new customers is large, and there are a large number of such unpopular stores on e-commerce platforms. These stores rely solely on operators to manually find target new customers for marketing activities (such as issuing advertisements / coupons), which is often too inefficient. Therefore, it is necessary to use algorithms to help unpopular stores find their potential new customers and improve the operating efficiency of stores and merchants.

[0039] In the related technology, the following methods can be used to attract new customers to a store: (1) Based on a collaborative filtering algorithm, find users who are similar to the old users of the target store and regard them as potential new users of the store; (2) Based on a collaborative filtering algorithm, find stores similar to the target store and regard the old users of similar stores as potential new users of the store; (3) Modeling based on user-store behavioral characteristics / user statistical characteristics / store statistical characteristics, and directly predict potential new users who may make purchases through the model.

[0040] However, on the one hand, for finding similar users of old users of a store based on collaborative filtering and for finding users of similar stores based on collaborative filtering, there are problems of too large a recall of users and too low prediction accuracy; on the other hand, for directly using user-store behavior samples for modeling, if users who have interactive behaviors with the store (called interactive users) are used as training samples, then since a large number of unpopular stores do not have enough interactive user samples, it is easy to lead to insufficient model training, and the expected effect cannot be achieved for unpopular stores; if users who have no behavior with the store are used as training samples, due to the missing / sparse characteristics of some behaviors, it will lead to problems of too large a user recall and too low prediction accuracy.

[0041] Based on this, the embodiment of the present application provides a user recommendation method, the basic idea of ​​which is: determine the user commodity attribute value data of the user to be recommended and the store commodity attribute value data of the target store; use the preset commodity feature model to perform feature analysis on the user commodity attribute value data and the store commodity attribute value data respectively, and obtain user feature information and store feature information; perform similarity calculation on the user feature information and the store feature information, determine the target user set from the users to be recommended according to the calculation results, and recommend the target user set to the target store. In this way, by determining the user feature information and the store feature information through the preset commodity feature model, the correlation between the user behavior and the store commodity can be accurately calculated, so as to determine the potential new users recommended to the target store, and there is no disadvantage of excessive user recall; at the same time, the technical solution of the present application can realize the update of the user feature information and the store feature information only by updating the preset commodity feature model, and then the potential new users recommended to the target store can be updated in real time, thereby not only improving the accuracy of user recommendation, but also ensuring the real-time nature of user recommendation, thereby improving the operation efficiency of the store.

[0042] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0043] In one embodiment of the present application, see Figure 1 , which shows a flow chart of a user recommendation method provided by an embodiment of the present application. Figure 1 As shown, the method may include:

[0044] S101: Determine the user product attribute value data of the user to be recommended and the store product attribute value data of the target store.

[0045] It should be noted that the user recommendation method provided in the embodiment of the present application can be applied to an e-commerce platform or a device related to the e-commerce platform. Here, the device can be a computer, a smart phone, a tablet computer, a laptop computer, a PDA, a navigation device, etc., and the embodiment of the present application does not specifically limit this.

[0046] Specifically, the user recommendation method can be applied to related scenarios such as recommending users to target stores, recommending stores to target users, and recommending products to target users. Here, a store can refer to a specific e-commerce store, or a mobile application (Application, App), a media channel, a physical store, etc. In other words, based on the inventive concept provided by the embodiments of the present application, those skilled in the art can implement the user recommendation method in a variety of business scenarios after adaptively modifying it, which are all within the protection scope of the embodiments of the present application.

[0047] For the sake of convenience, the following is a specific description using the business scenario of recommending potential new users to a target store as an example.

[0048] At this time, it is necessary to make user recommendations to the target store based on the user product attribute value data and the store product attribute value data. Here, the user product attribute value data contains the attribute information of the products that the recommended user has interacted with in the recent period of time (such as browsed products, clicked products, etc.), and the store product attribute value data refers to the attribute information of the products sold in the target store. In addition, in the business scenario of attracting new customers to the store, the users to be recommended generally refer to several users who have not had any interaction with the target store, or several users who have not placed orders in the target store, and can be screened and determined according to actual needs. For a clearer explanation, the users to be recommended will also be referred to as several users below.

[0049] S102: Perform feature analysis on the user product attribute value data and the store product attribute value data respectively using a preset product feature model to obtain user feature information and store feature information.

[0050] It should be noted that the embodiment of the present application involves a preset product feature model, and the preset product feature model is used to indicate the feature vectors of different product attribute values. Here, the user product attribute value data and the store product attribute value data are analyzed by the preset product feature model to obtain user feature information and store feature information, and further, based on the similarity between the user feature information and the store feature information, potential new users can be recommended to the target store.

[0051] In some embodiments, the use of a preset product feature model to perform feature analysis on the user product attribute value data and the store product attribute value data to obtain user feature information and store feature information may include:

[0052] Processing the user's commodity attribute value data based on the preset behavior category and the preset commodity attribute category to obtain the behavior-commodity category attribute value data of the user to be recommended;

[0053] Using a preset product feature model to perform feature analysis on the behavior of the recommended user-product category attribute value data to obtain user feature information; and

[0054] Processing the store product attribute value data based on the preset product attribute categories to obtain the product category attribute value data of the target store;

[0055] The preset product feature model is used to perform feature analysis on the product category attribute value data of the target store to obtain store feature information.

[0056] In the embodiments of the present application, generally, store product attribute value data refers to the attribute-related information of the products sold in the store, and user product attribute value data refers to the product interaction behavior sequence generated by the interaction between different users and the e-commerce platform. It should be understood that through the historical operation data of the e-commerce platform, it is possible to easily obtain the product interaction behavior sequence of each user, which also constitutes the user product attribute value data.

[0057] In other words, for a store, the store product attribute value data includes the attribute values ​​of the products for sale, however, for the user to be recommended, the user product attribute value data includes the product attribute values ​​of the user to be recommended under different interactive behaviors.

[0058] Therefore, for stores, since the same product has different attribute values ​​for different product attribute categories, for example, the product attribute category can be product brand, product category, or (product brand × product category), it is necessary to process the store product attribute value data according to the preset product attribute category to obtain the product category attribute value data, and then use the preset product feature model to analyze it to determine the store feature information.

[0059] For the users to be recommended, in addition to considering the product attribute categories, different interactive behaviors of the users to be recommended on the same product also contribute different information. Therefore, it is necessary to process the user product attribute value data according to the preset behavior categories and preset product attribute categories at the same time, obtain the behavior-product category attribute value data of the users to be recommended, and then analyze it using the preset product feature model to determine the user feature information.

[0060] Here, the preset behavior category includes at least one of the following: browsing behavior, clicking behavior, adding to shopping cart (or called add-to-cart) behavior, searching behavior and purchasing behavior; the preset product attribute category includes at least one of the following: product brand and product category.

[0061] For example, taking the user to be calculated (the user to be calculated can be understood as a specific user among the users to be recommended) as an example, the specific processing method is as follows:

[0062] First, based on the system log of the e-commerce platform, the interaction behavior records of the user to be calculated, that is, the user's product attribute value data, can be obtained. Here, the interaction behavior records include the product sequence with which the user to be calculated has interacted, such as the product sequence clicked by the user to be calculated, the product sequence collected by the user to be calculated, and so on.

[0063] Second, the degree of interest of users in products varies according to different interactive behaviors. Therefore, the user product attribute value data is classified according to the preset behavior category to obtain the behavior-product category attribute value data of the user to be calculated, which can be a browse-product attribute value set and a favorite-product attribute value set.

[0064] Third, taking the browsed-product attribute value set as an example, it is classified again according to different product attributes to obtain the browsed-product brand subset, the browsed-product category subset, and the browsed-(product brand×product category) subset.

[0065] That is to say, the user product attribute value data can be regarded as a plurality of behavior-product category attribute value sets of a plurality of users, and each behavior-product category attribute value set can be further divided into a plurality of subsets.

[0066] Furthermore, in some embodiments, the use of a preset product feature model to perform feature analysis on the behavior-product category attribute value data of the recommended user to obtain user feature information may include:

[0067] Inputting the behavior-product category attribute value data of the user to be recommended into the preset product feature model to obtain the behavior-feature vector data of the user to be recommended;

[0068] Performing mean processing on the behavior-feature vector data of the user to be recommended to obtain the behavior feature data of the user to be recommended;

[0069] The behavioral feature data of the user to be recommended are spliced ​​to obtain user feature information.

[0070] It should be noted that, for the user to be recommended, first, the behavior-product category attribute value data is input into the preset product feature model in sequence, and the behavior-feature vector data output by the preset product feature model can be obtained accordingly; then, the behavior-feature vector data is averaged to obtain the behavior feature data; finally, the behavior feature data is spliced ​​to obtain the feature information, thereby obtaining the user feature information.

[0071] Exemplarily, assuming that the behavior-product category attribute value data includes a set of browsed-product category attribute values ​​of the user to be calculated, and the set of browsed-product category attribute values ​​includes a browsed-product brand subset, a browsed-product category subset, and a browsed-(product brand×product category) subset, the processing method is as follows:

[0072] First, the browse-product brand subset, browse-product category subset, and browse-(product brand × product category) subset are sequentially input into the preset product feature model, and the browse-brand feature vector subset, browse-category feature vector subset, and browse-brand category feature vector subset (equivalent to behavior-feature vector data) are obtained;

[0073] Secondly, taking the subset as the unit, the browsing-brand feature vector subset, the browsing-category feature vector subset and the browsing-brand category feature vector subset are averaged once to obtain the browsing-brand mean vector, the browsing-category mean vector and the browsing-brand category mean vector; then, the browsing-brand mean vector, the browsing-category mean vector and the browsing-brand category mean vector are averaged twice to obtain the browsing feature vector of the user to be calculated relative to the browsing behavior;

[0074] Finally, by performing similar processing according to the above method, the collection feature vector, purchase feature vector, etc. (equivalent to behavioral feature data) of the user to be calculated can be obtained. By splicing these vectors, the feature information of the user to be calculated can be obtained.

[0075] It should be understood that the preset product feature model may include multiple preset product sub-models to flexibly process browsing-product attribute value subsets under different preset product attribute categories and different preset behavior categories.

[0076] In this way, user feature information integrates multi-dimensional user behaviors and different product attribute granularities (brand / category / (product brand × product category)), which can better present the behavioral characteristics of the users to be recommended, so as to improve the prediction accuracy.

[0077] Furthermore, in some embodiments, the step of performing feature analysis on the commodity category attribute value data of the target store using a preset commodity feature model to obtain store feature information may include:

[0078] Input the commodity category attribute value data of the target store into the preset commodity feature model to obtain the store feature vector data;

[0079] The store feature vector data is averaged to obtain the store feature information.

[0080] It should be noted that, for commodities, the commodity category attribute value data is first input into the preset commodity feature model in sequence, and the store feature vector data output by the preset commodity feature model can be obtained accordingly; then the store feature vector data is averaged to obtain the store feature information.

[0081] Exemplarily, the commodity category attribute value data may include multiple store commodity attribute value subsets, such as store commodity category attribute value subsets, store commodity brand attribute value subsets, store (commodity brand × commodity category) attribute value subsets, etc.; these store commodity attribute value subsets are sequentially input into the preset commodity feature model to obtain the corresponding multiple store feature vector subsets (equivalent to store feature vector data), and each store feature vector subset is averaged once to obtain multiple store feature vectors, and then the multiple store feature vectors are averaged twice to obtain store feature information.

[0082] In this way, through the above processing, the store characteristics and multi-behavior user characteristics (i.e., user characteristics under different behavior categories) are dimensionally aligned, and the generated expanded store feature information also integrates different product attribute granularities. When using user feature information and expanded store feature information for similarity calculation, the results are more accurate, further improving the prediction effect.

[0083] S103: Calculate the similarity between the user feature information and the store feature information, determine a target user set from the users to be recommended according to the calculation result, and recommend the target user set to the target store.

[0084] It should be noted that the similarity calculation is performed on the user feature information and the store feature information, and the target user set (equivalent to the aforementioned potential new users) is determined from the users to be recommended (the users to be recommended refer to a user group, which may include multiple specific users), and the target user set is recommended to the target store.

[0085] Furthermore, in some embodiments, the performing similarity calculation on the user characteristic information and the store characteristic information and determining the target user set from the users to be recommended according to the calculation result may include:

[0086] Performing dimension expansion processing on the store feature information to obtain expanded store feature information; wherein the expanded store feature information has the same dimension as the user feature information;

[0087] Calculate the similarity between the user feature information and the expanded store feature information to obtain the relevance value between the user to be recommended and the target store;

[0088] The recommended users are sorted according to their relevance values, and the target user set is determined based on the sorting results.

[0089] It should be noted that, according to the above content, for users, due to the distinction of preset behavior categories, and the user behavior-feature vector data under different preset behavior categories need to be averaged and spliced ​​to obtain user feature information; however, for stores, there is no distinction of preset behavior categories, and the calculated store feature vector data can be directly used as store feature information after averaging.

[0090] In other words, the dimensions of user feature information and store feature information may be different, so it is necessary to perform dimension expansion processing on the store feature information according to the user feature information to obtain the expanded store feature information so that the dimension of the expanded store feature information is the same as that of the user feature information.

[0091] After the dimension expansion, the vector distance calculation is performed on the user feature information and the expanded store feature information to obtain the relevance values ​​of several users. Here, the relevance value can indicate the user's interest in the products sold in the target store.

[0092] Exemplarily, the preset product feature model may include multiple preset product sub-models to flexibly process browsing-product attribute value subsets under different preset behavior categories. Then, dimensional expansion can be understood as: using multiple preset product sub-models to process the product attribute value data of the store to be processed separately, and splicing the processing results.

[0093] It should be understood that the user feature information actually includes the feature vectors of several users. The feature vector of each user is calculated with the store feature information (the essence of the store feature information is also a feature vector), and the calculated vector distance value is the relevance value of the user. Here, the vector distance value can be a cosine distance value, Jaccard distance, Euclidean distance or other similar indicators.

[0094] After obtaining the relevance values ​​of several users, since the relevance value refers to the vector distance value, the smaller the relevance value, the higher the relevance between the user and the store product, the several users are sorted from small to large in terms of relevance value, and the first M users are selected from the sorting results as the target user set. Here, M is a positive integer, and the value of M needs to be determined according to the business scenario.

[0095] Through the above steps, potential new users of the target store can be identified from a number of users, thereby helping the target store to complete the store's new user operation. The store can then design corresponding sales strategies for potential new users, improve the store's product conversion rate, and thus improve the store's operating efficiency.

[0096] The embodiment of the present application provides a user recommendation method, which determines the user product attribute value data of the user to be recommended and the store product attribute value data of the target store; uses a preset product feature model to perform feature analysis on the user product attribute value data and the store product attribute value data, respectively, to obtain user feature information and store feature information; performs similarity calculation on the user feature information and the store feature information, determines the target user set from the users to be recommended according to the calculation results, and recommends the target user set to the target store. In this way, by determining the user feature information and the store feature information through the preset product feature model, the correlation between the user behavior and the store product can be accurately calculated, so as to determine the potential new users recommended to the target store; at the same time, the technical solution of the present application can realize the update of the user feature information and the store feature information only by updating the preset product feature model, and then can update the potential new users recommended to the target store in real time, thereby not only improving the accuracy of user recommendation, but also ensuring the real-time nature of user recommendation, thereby improving the operation efficiency of the store.

[0097] In another embodiment of the present application, a training method for a preset commodity feature model is described in detail. Figure 2 , which shows a flow chart of another user recommendation method provided in an embodiment of the present application.

[0098] like Figure 2 As shown, the method may include:

[0099] S201: Obtain sample data of user product attribute values ​​of sample users.

[0100] It should be noted that the embodiment of the present application mainly provides a training method for a preset product feature model.

[0101] Specifically, the product interaction behavior sequence of the sample user within a period of time (the specific duration needs to be determined according to the actual application scenario, such as 30 minutes, 1 day, etc.) is obtained from the e-commerce platform to obtain the sample data of the user's product attribute value. Here, the product interaction behavior sequence may include the products browsed, clicked, purchased, etc. by the sample user. The sample user may actually include multiple specific users.

[0102] S202: Preprocess the sample data of user product attribute values ​​to obtain training sample data.

[0103] It should be noted that the original user product attribute value sample data is preprocessed to obtain training sample data for model training. Here, the training sample data includes at least one product attribute value and at least one training label value of each product attribute value. Then, the neural network model can be used to learn the characteristics of each product attribute value, and then obtain a preset product feature model. In other words, the preset product feature model is used to indicate the feature vectors of different product attribute values.

[0104] Furthermore, in some embodiments, preprocessing the user product attribute value sample data to obtain training sample data may include:

[0105] Processing the user's commodity attribute value sample data according to the preset behavior category and the preset commodity attribute category to obtain a plurality of commodity attribute value combination pairs;

[0106] Performing statistical analysis on multiple commodity attribute value combination pairs, and determining respective weight values ​​of the multiple commodity attribute value combination pairs according to the statistical analysis results;

[0107] Based on the plurality of commodity attribute value combination pairs and the weight values ​​of the plurality of commodity attribute value combination pairs, training sample data is determined.

[0108] It should be noted that the user product attribute value sample data is processed according to the preset behavior category and the preset product attribute category to obtain multiple product attribute value combination pairs, including:

[0109] (1) Classifying the product interaction data according to preset behavior categories, and obtaining behavior-product category attribute value sample data, such as browse-product attribute value set, click-product attribute value set, purchase-product attribute value set, etc.;

[0110] (2) Splitting and combining the behavior-product category attribute value sample data according to the preset product attribute category to obtain multiple product attribute value combination pairs. When splitting and combining the behavior-product category attribute value sample data, it is necessary to combine the product attribute values ​​of the same user under the same preset behavior category, and the two combined product attribute values ​​are from different products.

[0111] After obtaining a plurality of commodity attribute value combination pairs, statistical analysis is performed on each commodity attribute value combination pair, and the number of sample users having the commodity attribute value combination pair is determined as the weight value of the commodity attribute value combination pair.

[0112] Exemplarily, it is assumed that the behavior-product category attribute value sample data includes a browsed product set A{(brand 1×category 2), (brand 2×category 3)} of a certain sample user; wherein the product attribute value is expressed as product brand×product category.

[0113] At this time, the browsed product set A is split and combined to obtain the following product attribute value combination pairs: [Brand 1, Brand 2], [Brand 1, Category 3], [Category 2×Brand 2], [Category 2, Category 3], [Brand 1×Brand 2, Brand 2], [Brand 2, Brand 2×Category 3].

[0114] After performing the above processing on all sample users, the number of sample users with the product attribute value combination pair [brand 1, brand 2] is counted to obtain the weight value of [brand 1, brand 2].

[0115] Through the above steps, a plurality of commodity attribute value combination pairs and respective weight values ​​of the plurality of commodity attribute value combination pairs can be obtained, and then the training sample data can be determined.

[0116] Furthermore, in some embodiments, determining the training sample data based on the multiple commodity attribute value combination pairs and the weight values ​​of the multiple commodity attribute value combination pairs may include:

[0117] Constructing a product attribute association graph according to a plurality of product attribute value combination pairs and respective weight values ​​of the plurality of product attribute value combination pairs;

[0118] Constructing a product attribute association graph according to a plurality of product attribute value combination pairs and respective weight values ​​of the plurality of product attribute value combination pairs;

[0119] Calculate the product attribute association graph using a preset random walk model to obtain at least one nearest neighbor sequence;

[0120] Performing window partitioning processing on at least one neighbor sequence to obtain at least one training input value and at least one training label value;

[0121] At least one training input value and at least one training label value are determined as training sample data.

[0122] It should be noted that the commodity attribute association graph is constructed according to a plurality of commodity attribute value combination pairs and respective weight values ​​of the plurality of commodity attribute value combination pairs.

[0123] Specifically, each product attribute value is regarded as a vertex, each product attribute value combination pair indicates that there is an edge between two vertices (the edge is a directed edge, pointing from the first product attribute value to the second product attribute value), and the weight value of each product attribute value combination pair is the weight value of the edge.

[0124] For example, assuming that the training sample data includes: [brand 1×category 2, category 3](6), [brand 1×category 2, brand 2](4), [brand 1×category 2, brand 2×category 3](2), [brand 2×category 2, brand 1×category 2](1), [brand 2×category 2, brand 3×category 3](1); where the numbers in brackets are the weight values ​​of the product attribute value combination pairs, the product attribute association graph formed by the above data is as follows: Figure 3 shown.

[0125] In constructing the product attribute association graph, it is necessary to distinguish according to different preset behavior categories. That is, for click behavior, there is a corresponding product attribute association graph, and for browsing behavior, there is another corresponding product attribute association graph.

[0126] After obtaining the commodity attribute association graph, the commodity attribute association graph is calculated using a random walk sub-model, thereby obtaining at least one nearest neighbor sequence, and at least one nearest neighbor sequence includes multiple commodity attribute values.

[0127] The obtained neighbor sequence is subjected to window partitioning processing to determine at least one product attribute value association pair (specifically, each product attribute value association pair includes a training input value and a training label value), and then at least one training input value and at least one training label value are obtained, and thus training sample data is obtained. Here, assuming that the product attribute value association pair includes product attribute value 1 and product attribute value 2, it can be understood that product attribute value 2 is the product object that the user may interact with next after interacting with the product corresponding to product attribute value 1.

[0128] Further, in some embodiments, window partitioning is performed on at least one neighbor sequence to obtain at least one training input value and at least one training label value, which can be implemented as follows: a preset window threshold N is determined; in each neighbor sequence in at least one neighbor sequence, the i-th product attribute value is determined as the training input value, and the (i±n)-th product attribute value is determined as the training label value to obtain at least one training input value and at least one training label value; wherein i, n and N are all positive integers, and n is less than or equal to N.

[0129] It should be noted that, in order to determine the associated pairs of commodity attribute values, first, a preset window threshold N is set (set according to the actual application scenario). Exemplarily, N is 2. At this time, in each neighbor sequence, the i-th commodity attribute value is associated with the (i-2)-th, (i-1)-th, (i+1)-th or (i±2)-th commodity attribute value to obtain the aforementioned commodity attribute value associated pairs. In other words, the i-th commodity attribute value is a training input value, and the (i±n)-th commodity attribute value is the training label value of the i-th commodity attribute value.

[0130] In this way, through the above processing, training sample data containing multiple commodity attribute value association pairs can be obtained, so as to obtain a preset commodity feature model through subsequent training.

[0131] S203: Using the training sample data to train the preset neural network model to obtain a preset product feature model.

[0132] It should be noted that the preset neural network model is trained using the training sample data, and when the loss function reaches a preset threshold, a preset product feature model is obtained.

[0133] Specifically, the preset neural network model may include an input layer, a hidden layer, and a classification layer. The preset neural network model is trained using the first commodity attribute value in the multiple commodity attribute value association pairs as input and the second commodity attribute value as a label. When the loss function reaches a preset threshold, the feature vectors of different commodity attribute values ​​can be determined according to the trained hidden layer vector, thereby obtaining a preset commodity feature model.

[0134] Through the above processing, a preset product feature model of a feature vector indicating the attribute value of each product can be obtained to facilitate the store to attract new customers.

[0135] Based on the aforementioned user recommendation method, the application scenario of the embodiment of the present application is a store new customer attraction scenario, and the problem to be solved is to find potential new users (or new customers) for unpopular stores. New customers for unpopular stores can be defined as users who have not made any purchase records for the store in the past year. However, the difficulty of the unpopular store new customer attraction scenario lies in the fact that there are too few users who have interacted with the unpopular store. Therefore, whether it is collaborative filtering or the statistical feature modeling method of user-store / user / store, it is difficult to accurately predict whether an unknown user will make a purchase in the future. In other words, in order to improve the accuracy of the model, it is necessary to connect users and stores in terms of features. Since there is no direct interaction feature between unpopular stores and new customers, the interaction features between users and stores can be indirectly established through the feature vector of the product attribute value.

[0136] The specific implementation method includes: using the product attribute value sequence (category / (brand×category) / brand) of the user's interactive behavior (browsing / clicking / adding to cart / searching / purchasing) within a period of time as input, and calculating the feature vector used to describe the product attribute value (category / (brand×category) / brand) through an algorithm, and then obtaining the user's feature vector (equivalent to user feature information) and the store's feature vector (equivalent to store feature information) through the calculated feature vector of the product attribute value (category / (brand×category) / brand), and finally calculating the cosine similarity of the user's feature vector and the store's feature vector, and sorting them to obtain the most similar target user set, which is the store's potential new customers.

[0137] Since the feature vectors of the store and the user in the embodiment of the present application are calculated from the user behavior, compared to the collaborative filtering method in the related art, the store feature vector and the user feature vector are determined by the user's behavior, and the possible interactive relationship between the store and the user is indirectly obtained, which can more accurately represent the characteristics of the store from the perspective of user behavior. In the final prediction results, the store is often a store that includes categories / (brand × category) / brand products that are predicted to be strongly correlated with the user's behavior. Compared with the method of using user-store behavior samples for modeling in the related art, there is no disadvantage of only recalling users who have behavior in the store and the recall sample size is too small, thereby avoiding the disadvantages of sparse interactive features and low model accuracy.

[0138] In short, the user recommendation method provided in the embodiment of the present application may include the following contents:

[0139] (1) The behavior sequence of sample users (equivalent to the sample data of user product attribute values) is processed through an algorithm to generate features describing the products (brand, category, (brand × category)) (equivalent to product attribute values), and then generate features describing the store (store feature information) and features describing the user (user feature information). The store features generated by the above method have a strong correlation with the store products, and the user features also have a strong correlation with the user behavior. It is suitable for finding potential new customers for the store and has a high prediction accuracy.

[0140] (2) The user features generated by this method integrate multi-dimensional user behaviors and different product granularities (brand, category, (brand × category)), and the generated store features integrate different product granularities. The store features are dimensionally aligned with the multi-behavior user features;

[0141] (3) New customer recommendations can be made through store characteristics and user characteristics. The similarity between store characteristics and user characteristics is the behavioral correlation between the user and the store.

[0142] The embodiment of the present application provides a user recommendation method. The specific implementation method of the aforementioned embodiment is elaborated in detail through this embodiment. It can be seen that the interaction characteristics between users and stores can be indirectly established through the preset product feature model, and there is no need to focus only on user data that has interacted with the store. Therefore, the preset product feature model can be fully trained and the model accuracy is high; moreover, by generating store feature information and user feature information through the preset product feature model, the correlation between user behavior and store products can be characterized, so that there is a strong correlation between the target user set and the products in the target store, and the prediction accuracy is high; in addition, only the new user product attribute value data needs to be retrained to update the preset product feature model, further update the user feature information and store feature information, and finally update the target user set, which also improves the versatility of user recommendation.

[0143] In another embodiment of the present application, see Figure 4 , which shows a flow chart of another user recommendation method provided by an embodiment of the present application. Figure 4 As shown, the method may include:

[0144] S301: Obtain sample data of user product attribute values.

[0145] It should be noted that the user recommendation method of the embodiment of the present application can be divided into two parts. The first part is to establish a preset product feature model based on user product attribute value sample data, and the second part is to perform new store acquisition operations based on the preset product feature model.

[0146] For the first part, we need to collect user behavior data, which can be understood as the sequence of products that users browse / click / add to cart / search / purchase in a session, and thus obtain the user's product attribute value data. The purchase session lasts for 1 day, and the remaining sessions last for 30 minutes. In addition, in the product sequence, each product is represented in the form of (brand × category).

[0147] S302: Processing the sample data of user commodity attribute values.

[0148] It should be noted that after obtaining the sample data of the user's product attribute value, the (brand × category) in the Session is classified and deduplicated. For example, the data after classification and deduplication (equivalent to the aforementioned behavior-product category attribute value data) is in the following form:

[0149] User 1 browses: [(brand 1 × category 2), (brand 2 × category 3), (brand 3 × category 4)…]

[0150] User 1 clicks: [(brand 1 × category 2)…]

[0151] User 1 adds to cart: [(brand 2 × category 3)…]

[0152] User 1 searches: [(brand 1×category 3)…]

[0153] User 1 purchases: [(brand 1 × category 3)…]

[0154] User 2 browses: [(brand 2 × category 2), (brand 3 × category 3), (brand 4 × category 4)…]

[0155] User 2 clicks: [(Brand 2 × Category 2)…]

[0156] User 2 adds to cart: [(Brand 3 × Category 3)…]

[0157] User 2 searches: [(brand 1×category 3)…]

[0158] User 2 purchases: [(brand 3 × category 3)…]

[0159] For the above data, outliers are removed. Here, the outlier removal rule can be determined according to the application scenario, for example, the data with a browsing quantity (brand × category) greater than 100 or only 1 in each Session is deleted.

[0160] After removing outliers, the (brand × number of categories) in each user's behavior sequence is expanded in pairs by brand and category (the specific combination rules can be determined according to business needs) to generate and count product attribute value combination pairs (referred to as Pair pairs). The data format of Pair pairs and their counts is as follows (the above takes the browsing sequence as an example, and the same applies to other behavior sequences):

[0161] [Brand 1 × Category 2, Brand 2 × Category 3]2

[0162] [Brand 1×Category 2, Brand 2]4

[0163] [Brand 1×Category 2, Category 3] 6

[0164] [Brand 2 × Category 2, Brand 3 × Category 3] 1

[0165] [Brand 2 × Category 2, Brand 1 × Category 2] 1

[0166] Among them, each Pair pair includes two product attribute values, that is, the product attribute value can be the brand information of a certain product, the category information of a certain product, or the brand category information of a certain product. For the convenience of explanation, the product attribute value can also be called the product identification (Identity Document, ID), and the ID is used to refer to the product attribute value below. In addition, the number after each Pair pair refers to the number of sample users who have the Pair pair, that is, the count of the Pair pair. In addition, the above is only an example of the data format of the Pair pair, and there is no connection with the previous example.

[0167] S303: Construct a product attribute association graph based on the data processing results.

[0168] It should be noted that the product attribute association graph (Graph) is constructed based on the Pair pairs and their weight values ​​obtained above. The ID in the Pair pairs is the vertex (Node) in the product attribute association graph, and the Pair pair count is the weight value of the edge (Edge) between the two vertices. Based on the sample data of the Pair pairs and their counts above, the obtained product attribute association graph is as follows Figure 3 shown.

[0169] S304: Using the product attribute association graph to train a preset node embedding model.

[0170] It should be noted that the preset node embedding model is trained using the commodity attribute association graph, the commodity attribute association graph is input into the preset node embedding model, and the output is the trained feature vector of each ID. Here, the preset node embedding model can be understood as the whole of the aforementioned preset random walk model and the preset neural network model.

[0171] The preset node embedding model is the Node2Vec algorithm as an example for specific explanation.

[0172] In the first step, for the product attribute association graph, a biased random walk method (i.e., a preset random walk model) is used to obtain the neighbor sequence of each vertex.

[0173] Given the current vertex v, the probability of the next vertex x in the neighbor sequence is as shown in formula (1):

[0174]

[0175] Among them, Z is the normalization constant (needs to be determined according to the actual application scenario), The unnormalized transition probability from vertex v to vertex x, E indicates that there is an edge between v and x.

[0176] Here, assuming that the current random walk passes through the edge (t, v) to reach the current vertex v, then let And ωvx is the weight of the edge between vertex v and x, α pq The value of (t,x) is shown in formula (2):

[0177]

[0178] Among them, d tx is the distance of the shortest path from vertex t to vertex x. p and q are parameters (the values ​​need to be determined according to the actual application environment). p is used to control the probability of repeatedly visiting the vertex just visited. q controls whether the walk is outward or inward. If q>1, the random walk tends to visit vertices close to t, tending to the breadth-first search algorithm (Breath First Search, BFS). If q<1, it tends to visit vertices far away from t, tending to the depth-first search algorithm (Depth First Search, DFS). See Figure 5 , which shows a schematic diagram of the working process of another user recommendation method provided by an embodiment of the present application. When accessing from t to v, Figure 5 Describes the α corresponding to each vertex when deciding the next vertex to visit pq (t,x).

[0179] The second step is to generate training samples of a preset neural network model (such as Word2Vec model) based on the results of the above algorithm. Set p = 100, q = 0.5, for Figure 3 Generate 100 neighbor sequences of length 6 for each vertex in . Here, the number and length of the neighbor sequence can be determined according to the actual application scenario. The example results of the neighbor sequence are as follows:

[0180] [Brand 13333, Category 5, Brand 22, Category 339, Brand 121, Category 23…]

[0181] [Brand 13, Category 5, Brand 321, Brand 22, Category 339, Brand 121, Category 23…]

[0182] [Brand 13, Category 22, Brand 22, Category 339, Category 23…]

[0183] [Brand 13 Category 22, Brand 22 Category 339, Category 121…]

[0184] …

[0185] The above neighbor sequence is divided according to the window threshold to construct the training input of the Word2Vec model. For example, the neighbor sequence is [brand 13333, category 5, brand 22, category 339, brand 121 category 23], and the window threshold (equivalent to the aforementioned preset threshold) is 2. Then the product attribute value association pair is generated for the i-th store and the (i±n)-th store, and then the training sample data is obtained. Since the window threshold is 2, n is 1 or 2. At this time, for [brand 13333, category 5, brand 22, category 339, brand 121 category 23...], the obtained product attribute value association pairs are as follows:

[0186] [Brand 13333, category 5], [Brand 13333, brand 22], [Category 5, brand 13333], [Category 5, brand 22], [Category 5, category 339], [Brand 22, brand 13333], [Brand 22, category 5], [Brand 22, category 339], [Brand 22, brand 121 category 23], [Category 339, category 5], [Category 339, brand 22], [Category 339, brand 121 category 23], [Brand 121 category 23, brand 22], [Brand 121 category 23, category 339].

[0187] At this time, what is generated is a product attribute value association pair in the form of [target ID, context neighbor ID], where the target store is the training input value and the context neighbor store is the training input value.

[0188] The third step is to use the generated training samples to train the Word2Vec model (equivalent to the aforementioned preset neural network model). The Word2Vec model is a neural network algorithm. In the embodiment of the present application, the Word2Vec model is specifically a Skip-Gram model, which includes an input layer, a hidden layer, and an output layer.

[0189] Assume that the input layer vector is a, the size is (n'×1), and the hidden layer vector h=W T ·a+b. Where W is a matrix of size (n'×m), b is a matrix of size m, W and b are randomly generated at the beginning, and then the training sample (X, Y) is input, X is the input vector (i.e., the training input value), Y is the true label (i.e., the training output label value), and the hidden layer is obtained after calculation for the input vector, and then the output layer is obtained. The Softmax function is used to score the output layer result. The Softmax function is shown in formula (3):

[0190]

[0191] Among them, x' is the output vector of the output layer, y j is the jth element of y.

[0192] The output of the Softmax layer is the predicted probability value According to the predicted probability The loss is calculated by the training label value, and the loss function is set to cross entropy. The specific calculation process is shown in formula (4).

[0193]

[0194] After calculating the loss, reverse the weight derivation through gradient descent Let the derivative be 0, which is the direction of the fastest gradient descent, update the weights W and b, and then perform the forward hidden layer output layer Softmax calculation to iteratively update the weights W and b.

[0195] Assume that there are 100,000 IDs in total, the feature vector (hidden layer) of each ID is set to 128, cross entropy is used as the loss function, and the window size is set to 2. Figure 6 , which shows a schematic diagram of the working process of another user recommendation method provided by an embodiment of the present application. Figure 6 As shown, the preset neural network model includes an input layer, a hidden layer (Hidden Layer) and an output layer (Output Layer); the input vector (Input Vector) of the input layer is 10000 dimensions (Positions), the hidden layer (Hidden Layer) uses 300 linear neurons (Linear Neurons), the output layer uses the Softmax function for scoring (or called Softmax Classifier), and the output layer has 10000 neurons.

[0196] based on Figure 6 The training process of the preset neural network model (i.e., the preset product feature model) is as follows: the training samples generated by the nearest neighbor sequence are sequentially input into the word vector sub-model Skip-Gram model, the network weights are randomly initialized, the weight matrix size between the randomly initialized input layer and the hidden layer is (100000×128), the weight matrix size between the hidden layer and the output layer is (100000×128), each input ID corresponds to a randomly initialized hidden layer vector, and then the forward network calculates the loss, and then the gradient descent is used to reversely update the weight, and finally the weight obtained after multiple iterations of loss convergence is the trained weight. In other words, the trained hidden layer vector is the feature vector of the ID.

[0197] S305: Generate a preset product feature model based on the training results.

[0198] It should be noted that the preset commodity feature model is generated according to the trained hidden layer vector. Here, the preset commodity feature model is used to indicate the feature vector of each ID, that is, the feature vector of each commodity attribute value.

[0199] In summary, the constructed product attribute association graph is taken as the input of the preset node embedding model, and the output is the required feature vector representing the ID.

[0200] S306: Generate user characteristic information.

[0201] It should be noted that after obtaining the preset product feature model, the technical content of the first part is completed. The following is an explanation of the technical content of the second part.

[0202] According to the preset product feature model, user feature information is generated. The essence of user feature information is a vector. In other words, the feature vector (Embedding) of user behavior is generated according to each user's recent different behavior sequence ID (equivalent to the user's product attribute value data).

[0203] Specifically, based on the generated ID and the feature vector of each ID (i.e., the preset product feature model), the user behavior sequence ID (brand / category / brand×category) is vectorized. Here, the user behavior sequence ID needs to be classified by brand / category / (brand×category), vectorized (embedded) under each attribute category, and then the vectors of all IDs in each category are averaged (average pooling), which generates the brand vector, category vector, and (brand×category) vector of the user behavior respectively, and then the three vectors are average pooled again to generate the user behavior vector.

[0204] See also Figure 7 , which shows a schematic diagram of the working process of another user recommendation method provided in an embodiment of the present application. Figure 7 As shown in the figure, the user's behavior sequence ID includes brand 3, brand 8, brand 1 and brand 2. It is necessary to use the preset product feature model to determine the feature vectors of brand 3, brand 8, brand 1 and brand 2, and average the obtained feature vectors to obtain the brand vector (or brand Embedding); similarly, continue to obtain the category vector (or category Embedding) and (brand × category) vector (or (brand × category) Embedding), the brand vector, category vector and (brand × category) vector can also be called ID vector (or ID Embedding); finally, average the brand vector, category vector and (brand × category) vector to obtain the user's behavior vector.

[0205] Further, see Figure 8 , which shows a schematic diagram of the working process of another user recommendation method provided in an embodiment of the present application. Figure 8 As shown, according to different behavior categories (browsing / searching / clicking / adding to cart / purchasing), different behavior vectors of the user can be obtained, namely, user browsing vector (or called user browsing Embedding), user search vector (or called user search Embedding), user purchase vector (or called user purchase Embedding), user click vector (or called user click Embedding) and user adding to cart vector (or called user adding to cart Embedding). These vectors (Embeddings) of different behaviors are connected (Soncatenation) together to obtain the user vector (or called user Embedding) of the user. The user vectors of different users constitute the user feature information.

[0206] S307: Generate store feature information.

[0207] Here, step S306 and step S307 can be executed in parallel, and there is no particular order in which they are executed.

[0208] It should be noted that, similar to the above process, store feature information is generated according to the preset product feature model, and the essence of store feature information is a vector. It should be understood that if the dimensions of store feature information and user feature information are different, the dimensions of store feature information need to be expanded.

[0209] See also Fig. 9 , which shows a schematic diagram of the working process of another user recommendation method provided in an embodiment of the present application. Fig. 9As shown, according to the store's main product ID (brand / category / (brand×category)), as well as the generated ID and feature vectors (Embedding) of different IDs, the vectors of brand / category / (brand×category) are average pooled to generate the store's brand vector (or brand embedding), category vector (or category embedding) and (brand×category) vector (or (brand×category) embedding) respectively. The brand vector, category vector and (brand×category) vector can also be called ID vector (or ID embedding). The three vectors are average pooled again to generate the store's embedding. When the store embedding and the user embedding have different dimensions, the store embedding needs to be expanded to the same dimension as the store's final feature vector (or store unified embedding), which is equivalent to the aforementioned expanded store feature information.

[0210] S308: Calculate the correlation value between the user and the store based on the user feature information and the store feature information.

[0211] It should be noted that the cosine distance between the store feature information and the user feature information of all other users (except those who have been determined as new customers) is calculated to obtain the correlation value between the user and the store. It should be understood that if the store feature information has been dimensionally expanded, the correlation value needs to be calculated for the expanded store feature information and user feature information.

[0212] Specifically, the vector With vector The cosine distance is calculated by formula (5):

[0213]

[0214] S309: Sort multiple users according to the relevance values ​​between the multiple users and the stores to generate a target user set.

[0215] It should be noted that multiple users are sorted according to their relevance values ​​to the store, and the first 1 million users are sorted in descending order to be the target user set, that is, the final set of potential new customers.

[0216] It should also be noted that, in the above embodiments, the user's behavior sequence may be browsing / clicking / exposure / searching / adding to cart / purchasing and other behaviors, and the time length of the behavior sequence varies with the type of behavior; the algorithm for generating feature vectors based on the user behavior sequence can use a variety of existing algorithms, such as the DeepWalk algorithm or similar algorithms (such as the LINE algorithm, the Word2Vec algorithm), and is not limited to the Node2Vec algorithm; the similarity calculation between the store vector and the user vector can use cosine distance, Jaccard distance, Euclidean distance and other similar indicators.

[0217] In summary, the overall process of the user recommendation method provided in the embodiment of the present application is as follows: first, obtain the user's interactive behavior (browsing / clicking / adding to cart / searching / purchasing) data on the commodity attribute value (brand × category) from the underlying data of the e-commerce platform; second, divide each user behavior into sessions, combine and count the IDs in each session, and use the obtained ID-ID Pair pairs and counts are used to construct a product attribute association graph; secondly, the constructed product attribute association graph is subjected to biased random walk through the Node2Vec algorithm to generate an ID adjacent sequence, and the obtained ID adjacent sequence is used to generate training samples of the Skip-Gram model in the form of (target ID, context adjacent ID), i.e., (input, label) according to the window; thirdly, the Skip-Gram model in the Node2Vec algorithm is trained with the training samples, and the trained hidden layer is used as the feature vector of ID (brand / category / brand×category); finally, store vectors and user vectors are generated based on the feature vector of ID, and the similarity between the feature vectors of users and stores is calculated. For each store, the top 1 million (Top 1 million) with the highest similarity according to the cosine distance are selected as the final potential new customers.

[0218] The embodiment of the present application provides a user recommendation method. The specific implementation method of the above embodiment is described in detail through this embodiment. It can be seen that the training sample data is processed by the Node2Vec algorithm to obtain a preset product feature model, and then the user feature information and store feature information of multiple users are determined, so as to determine the target user set from multiple users. Therefore, the interaction characteristics between users and stores can be indirectly established through the preset product feature model, without only focusing on the user data that has interacted with the store, and then the preset product feature model can be fully trained, and the model accuracy is high; in addition, the store feature information and user feature information are generated by the preset product feature model, which can characterize the correlation between user behavior and store products, so that the target user set and the products in the target store have a strong correlation, and the prediction accuracy is high; in addition, the method provided by the embodiment of the present application is universal and real-time, and only needs to regularly update the input user behavior data to retrain the preset node embedding model, so as to update the feature vector (preset product feature model) of each ID, and the user vector and store vector can be updated every day according to the feature vector of each ID, so as to generate potential new customers for each store, so that the store can formulate corresponding sales strategies for potential new customers, thereby improving the operation efficiency of the store.

[0219] In yet another embodiment of the present application, see Fig.10 , which shows a schematic diagram of the composition structure of a user recommendation device 40 provided in an embodiment of the present application. Fig.10 As shown, the user recommendation device 40 includes a determination unit 401, an analysis unit 402 and a recommendation unit 403, wherein:

[0220] A determination unit 401 is configured to determine the user product attribute value data of the user to be recommended and the store product attribute value data of the target store;

[0221] The analysis unit 402 is configured to perform feature analysis on the user product attribute value data and the store product attribute value data respectively using a preset product feature model to obtain user feature information and store feature information;

[0222] The recommendation unit 403 is configured to calculate the similarity between the user feature information and the store feature information, determine a target user set from the users to be recommended according to the calculation result, and recommend the target user set to the target store.

[0223] In some embodiments, the analysis unit 402 is specifically configured to process the user's product attribute value data based on preset behavior categories and preset product attribute categories to obtain the behavior-product category attribute value data of the user to be recommended; perform feature analysis on the behavior-product category attribute value data of the user to be recommended using a preset product feature model to obtain user feature information; and process the store product attribute value data based on the preset product attribute categories to obtain the product category attribute value data of the target store; perform feature analysis on the product category attribute value data of the target store using a preset product feature model to obtain store feature information.

[0224] In some embodiments, the analysis unit 402 is further configured to input the behavior-product category attribute value data of the user to be recommended into a preset product feature model to obtain the behavior-feature vector data of the user to be recommended; perform mean processing on the behavior-feature vector data of the user to be recommended to obtain the behavior feature data of the user to be recommended; and perform splicing processing on the behavior feature data of the user to be recommended to obtain user feature information.

[0225] In some embodiments, the analysis unit 402 is further configured to input the commodity category attribute value data of the target store into a preset commodity feature model to obtain store feature vector data; and perform mean processing on the store feature vector data to obtain store feature information.

[0226] In some embodiments, the analysis unit 402 is further configured to perform dimensional expansion processing on the store feature information to obtain expanded store feature information; wherein the expanded store feature information has the same dimension as the user feature information; perform similarity calculation on the user feature information and the expanded store feature information to obtain a correlation value between the user to be recommended and the target store; sort the users to be recommended according to the correlation value, and determine the target user set based on the sorting result.

[0227] In some embodiments, Fig.11 As shown, the user recommendation device 40 also includes a training unit 404, which is configured to obtain sample data of user product attribute values ​​of sample users; preprocess the sample data of user product attribute values ​​to obtain training sample data; and use the training sample data to train a preset neural network model to obtain a preset product feature model.

[0228] In some embodiments, the training unit 404 is further configured to process user product attribute value sample data according to preset behavior categories and preset product attribute categories to obtain multiple product attribute value combination pairs; perform statistical analysis on the number of sample users on the multiple product attribute value combination pairs to determine the respective weight values ​​of the multiple product attribute value combination pairs; and determine the training sample data based on the multiple product attribute value combination pairs and the respective weight values ​​of the multiple product attribute value combination pairs.

[0229] In some embodiments, the training unit 404 is further configured to construct a product attribute association graph based on multiple product attribute value combination pairs and respective weight values ​​of the multiple product attribute value combination pairs; calculate the product attribute association graph using a preset random walk model to obtain at least one nearest neighbor sequence; perform window partitioning processing on at least one nearest neighbor sequence to obtain at least one training input value and at least one training label value; and determine at least one training input value and at least one training label value as training sample data.

[0230] In some embodiments, the preset behavior categories include at least one of the following: browsing behavior, clicking behavior, adding to shopping cart behavior, searching behavior, and purchasing behavior; the preset product attribute categories include at least one of the following: product brand and product category.

[0231] It can be understood that in this embodiment, a "unit" can be a part of a circuit, a part of a processor, a part of a program or software, etc., and of course it can also be a module, or it can be non-modular. Moreover, the components in this embodiment can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional module.

[0232] 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 prior art 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: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc., various media that can store program codes.

[0233] Therefore, this embodiment provides a computer storage medium, which stores a computer program. When the computer program is executed by multiple processors, the steps of the method in any one of the above embodiments are implemented.

[0234] Based on the above-mentioned composition of a user recommendation device 40 and computer storage medium, see Fig.12 , which shows a hardware structure diagram of an electronic device 50 provided in an embodiment of the present application. Fig.12 As shown, the electronic device 50 may include: a communication interface 501, a memory 502 and a processor 503; each component is coupled together via a bus device 504. It is understood that the bus device 504 is used to realize the connection and communication between these components. In addition to the data bus, the bus device 504 also includes a power bus, a control bus and a status signal bus. However, for the sake of clarity, Fig.12 Various buses are marked as bus devices 504. Among them, the communication interface 501 is used to receive and send signals during the process of sending and receiving information between other external network elements;

[0235] A memory 502, used to store a computer program that can be run on the processor 503;

[0236] The processor 503 is used to execute, when running the computer program:

[0237] Determine the user product attribute value data of the user to be recommended and the store product attribute value data of the target store;

[0238] Using the preset product feature model, feature analysis is performed on the user product attribute value data and the store product attribute value data to obtain user feature information and store feature information;

[0239] The similarity between the user feature information and the store feature information is calculated, and the target user set is determined from the users to be recommended based on the calculation results, and the target user set is recommended to the target store.

[0240] It can be understood that the memory 502 in the embodiment of the present application can be a volatile memory or a non-volatile memory, or can include both volatile and non-volatile memories. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory can be a random access memory (RAM), which is used as an external cache. By way of example but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM) and direct memory bus random access memory (DRRAM). The memory 502 of the apparatus and method described in the present application is intended to include, but is not limited to, these and any other suitable types of memory.

[0241] The processor 503 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the hardware integrated logic circuit or software instructions in the processor 503. The above processor 503 may be a general processor, a digital signal processor (Digital Signal Processor, DSP), an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), a field programmable gate array (Field Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components. The methods, steps and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general processor can be a microprocessor or the processor can also be any conventional processor, etc. The steps of the method disclosed in the embodiments of the present application can be directly embodied as a hardware decoding processor to execute, or the hardware and software modules in the decoding processor are combined and executed. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory 502, and the processor 503 reads the information in the memory 502 and completes the steps of the above method in combination with its hardware.

[0242] It is understood that the embodiments described in the present application can be implemented by hardware, software, firmware, middleware, microcode or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions of the present application or a combination thereof.

[0243] For software implementation, the technology of the present application can be implemented by a module (e.g., a process, a function, etc.) that performs the functions of the present application. The software code can be stored in a memory and executed by a processor. The memory can be implemented in the processor or outside the processor.

[0244] Optionally, as another embodiment, the processor 503 is further configured to execute the steps of the method in any one of the aforementioned embodiments when running the computer program.

[0245] In yet another embodiment of the present application, based on the composition diagram of the above-mentioned user recommendation device 40, see Fig.13 , which shows a schematic diagram of the composition structure of another electronic device 50 provided in an embodiment of the present application. Fig.13 As shown, the electronic device 50 at least includes the user recommendation device 40 of any one of the aforementioned embodiments.

[0246] For the electronic device 50, since it includes the user recommendation device 40, and the user recommendation device 40 can determine the user characteristic information and the store characteristic information according to the preset product characteristic model, it can accurately calculate the correlation between the user behavior and the store product, so as to determine the potential new users recommended to the target store; at the same time, the technical solution of the present application can realize the update of the user characteristic information and the store characteristic information only by updating the preset product characteristic model, and then can update the potential new users recommended to the target store in real time, thereby improving the accuracy of user recommendations.

[0247] The above are only preferred embodiments of the present application and are not intended to limit the protection scope of the present application.

[0248] It should be noted that, in this application, the terms "include", "comprises" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0249] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0250] The methods disclosed in several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0251] The features disclosed in several product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0252] The features disclosed in several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0253] The above are only specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A user recommendation method, characterized in that: The method comprises: Determine the user product attribute value data of the user to be recommended and the store product attribute value data of the target store; Using a preset product feature model, feature analysis is performed on the user product attribute value data and the store product attribute value data to obtain user feature information and store feature information; Calculating similarity between the user characteristic information and the store characteristic information, determining a target user set from the users to be recommended according to the calculation result, and recommending the target user set to the target store; The using of a preset commodity feature model to perform feature analysis on the user commodity attribute value data and the store commodity attribute value data to obtain user feature information and store feature information includes: Processing the user commodity attribute value data based on preset behavior categories and preset commodity attribute categories to obtain the behavior-commodity category attribute value data of the to-be-recommended user; Using the preset product feature model to perform feature analysis on the behavior-product category attribute value data of the user to be recommended to obtain the user feature information; and Processing the store product attribute value data based on the preset product attribute category to obtain product category attribute value data of the target store; The preset commodity feature model is used to perform feature analysis on the commodity category attribute value data of the target store to obtain the store feature information.

2. The user recommendation method according to claim 1, characterized in that: The using the preset commodity feature model to perform feature analysis on the behavior-commodity category attribute value data of the to-be-recommended user to obtain the user feature information includes: Inputting the behavior-product category attribute value data of the to-be-recommended user into the preset product feature model to obtain the behavior-feature vector data of the to-be-recommended user; Performing mean processing on the behavior-feature vector data of the user to be recommended to obtain the behavior feature data of the user to be recommended; The behavior characteristic data of the user to be recommended are concatenated to obtain the user characteristic information.

3. The user recommendation method according to claim 1, characterized in that: The using the preset commodity feature model to perform feature analysis on the commodity category attribute value data of the target store to obtain the store feature information includes: Inputting the commodity category attribute value data of the target store into the preset commodity feature model to obtain store feature vector data; The store feature vector data is averaged to obtain the store feature information.

4. The user recommendation method according to claim 1, characterized in that: The performing similarity calculation on the user characteristic information and the store characteristic information, and determining a target user set from the to-be-recommended users according to the calculation result, comprises: Performing dimension expansion processing on the store characteristic information to obtain expanded store characteristic information; wherein the expanded store characteristic information has the same dimension as the user characteristic information; Calculating the similarity between the user characteristic information and the expanded store characteristic information to obtain a correlation value between the user to be recommended and the target store; The users to be recommended are sorted according to the relevance values, and the target user set is determined according to the sorting result.

5. The user recommendation method according to any one of claims 1 to 4, characterized in that: The method further comprises: Obtain sample data of user product attribute values ​​of sample users; Preprocessing the user commodity attribute value sample data to obtain training sample data; The preset neural network model is trained using the training sample data to obtain the preset commodity feature model.

6. The user recommendation method according to claim 5, characterized in that: The preprocessing of the user commodity attribute value sample data to obtain training sample data includes: Processing the user commodity attribute value sample data according to preset behavior categories and preset commodity attribute categories to obtain a plurality of commodity attribute value combination pairs; Performing statistical analysis on the number of sample users for the multiple commodity attribute value combination pairs, and determining weight values ​​of the multiple commodity attribute value combination pairs; The training sample data is determined based on the multiple commodity attribute value combination pairs and the weight values ​​of the multiple commodity attribute value combination pairs.

7. The user recommendation method according to claim 6, characterized in that: The determining the training sample data based on the multiple commodity attribute value combination pairs and the weight values ​​of the multiple commodity attribute value combination pairs includes: constructing a commodity attribute association graph according to the plurality of commodity attribute value combination pairs and respective weight values ​​of the plurality of commodity attribute value combination pairs; Calculating the commodity attribute association graph using a preset random walk model to obtain at least one nearest neighbor sequence; Performing window partitioning processing on the at least one neighbor sequence to obtain at least one training input value and at least one training label value; The at least one training input value and the at least one training label value are determined as the training sample data.

8. The user recommendation method according to claim 1 or 6, characterized in that: The preset behavior category includes at least one of the following: browsing behavior, clicking behavior, adding to shopping cart behavior, searching behavior and purchasing behavior; The preset commodity attribute category includes at least one of the following: commodity brand and commodity category.

9. A user recommendation device, characterized in that: The user recommendation device comprises a determination unit, an analysis unit and a recommendation unit; wherein, The determining unit is configured to determine the user product attribute value data of the user to be recommended and the store product attribute value data of the target store; The analysis unit is configured to perform feature analysis on the user product attribute value data and the store product attribute value data respectively using a preset product feature model to obtain user feature information and store feature information; The recommendation unit is configured to perform similarity calculation on the user characteristic information and the store characteristic information, determine a target user set from the users to be recommended according to the calculation result, and recommend the target user set to the target store; The analysis unit is specifically configured to process the user's product attribute value data based on preset behavior categories and preset product attribute categories to obtain the behavior-product category attribute value data of the user to be recommended; perform feature analysis on the behavior-product category attribute value data of the user to be recommended using the preset product feature model to obtain the user feature information; and process the store's product attribute value data based on the preset product attribute category to obtain the product category attribute value data of the target store; perform feature analysis on the product category attribute value data of the target store using the preset product feature model to obtain the store feature information.

10. An electronic device, characterized in that: The electronic device comprises a memory and a processor; wherein, The memory is used to store a computer program that can be run on the processor; The processor is configured to execute the steps of the method according to any one of claims 1 to 8 when running the computer program.

11. A computer storage medium, characterized in that: The computer storage medium stores a computer program, and when the computer program is executed, the steps of the method according to any one of claims 1 to 8 are implemented.

Images