A user multi-feature intelligent recommendation method and system

Abstract

The present invention provides a method and system for intelligent recommendation of multiple features of users, the method comprising: obtaining historical behavior feature data of users, constructing a multi-interest user portrait based on a MIND model; calculating the comprehensive weight of each category of candidate products based on the user's behavior interaction frequency and behavior interaction time for each category of candidate products; and selecting candidate products based on the comprehensive weight to recommend to users. The present invention constructs a multi-interest user portrait based on the MIND model, and by combining a weighted model based on behavior frequency, a time decay model, and a weighted voting mechanism, it accurately selects products of appropriate categories to recommend to users, and can more flexibly handle the user's diverse behaviors and demand changes. The combination of these models and mechanisms not only improves the accuracy and relevance of the recommended content, but also further improves the user's satisfaction and participation, ensuring that the recommendation system can still maintain efficient and accurate recommendation capabilities when facing complex user behaviors.

Description

Technical Field

[0001] The present invention relates to the field of intelligent recommendation technology, and more specifically, to a method and system for intelligent recommendation of multiple features of users. Background Art

[0002] The product recommendation distribution system can recommend appropriate products to users in appropriate scenarios. A common example is the e-commerce recommendation system, which recommends various products to users on the Internet, such as new products on the shelves, discounted products, and hot-selling products.

[0003] Currently, e-commerce recommendation systems on the Internet generally recommend products based on product sales rankings, user evaluation scores of products, and user browsing frequency of products. Due to the lack of intelligent analysis factors required for personalized recommendations to users, in many cases the products recommended to users are not what the users like; or users may be interested in multiple different categories of products, but the recommendation system only recommends products in one category, which is not comprehensive and accurate. Summary of the invention

[0004] The present invention aims at the technical problems existing in the prior art and provides a method and system for intelligent recommendation of multiple features of users, thereby improving the accuracy and personalization of recommended contents.

[0005] According to a first aspect of the present invention, there is provided a method for intelligent recommendation of multiple features of a user, comprising:

[0006] Acquire user historical behavior feature data, wherein the user historical behavior feature data includes user online browsing, clicking and purchasing behavior data;

[0007] According to the user's historical behavior feature data, a multi-interest user portrait is constructed based on the MIND model, wherein the multi-interest user portrait expresses multiple categories of candidate commodities that the user is interested in;

[0008] Calculate the first weight of each category of candidate products based on the user's behavioral interaction frequency with the candidate products of each category;

[0009] According to the user's behavioral interaction time with each category of candidate products, a second weight of each category of candidate products is calculated based on a time decay model;

[0010] Calculating a comprehensive weight of the candidate commodities in each category according to the first weight and the second weight of the candidate commodities in each category;

[0011] Based on the comprehensive weight of the candidate products in each category, one or more candidate products in all categories are selected from the candidate products in all categories and recommended to the user.

[0012] According to a second aspect of the present invention, there is provided a user multi-feature intelligent recommendation system, comprising:

[0013] An acquisition module is used to acquire user historical behavior feature data, wherein the user historical behavior feature data includes user behavior data of browsing, clicking and purchasing goods online;

[0014] A construction module is used to construct a multi-interest user portrait based on the MIND model according to the user's historical behavior feature data, wherein the multi-interest user portrait expresses multiple categories of candidate commodities that the user is interested in;

[0015] a calculation module, configured to calculate a first weight of each category of candidate commodities based on the user's behavioral interaction frequency with respect to each category of candidate commodities; and to calculate a second weight of each category of candidate commodities based on a time decay model according to the user's behavioral interaction time with respect to each category of candidate commodities; and to calculate a comprehensive weight of each category of candidate commodities based on the first weight and the second weight of each category of candidate commodities;

[0016] The recommendation module is used to select one or more categories of candidate products from all categories of candidate products based on the comprehensive weight of each category of candidate products and recommend them to users.

[0017] The present invention provides a user multi-feature intelligent recommendation method and system, which builds multi-interest user portraits based on the MIND model, and accurately selects suitable categories of goods to recommend to users by combining a weighted model based on behavior frequency, a time decay model, and a weighted voting mechanism, and can more flexibly handle users' diverse behaviors and demand changes. The combination of these models and mechanisms not only improves the accuracy and relevance of recommended content, but also further improves user satisfaction and participation, ensuring that the recommendation system can still maintain efficient and accurate recommendation capabilities when facing complex user behaviors. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flow chart of a method for intelligent recommendation of multiple features of a user provided by the present invention;

[0019] Figure 2 A schematic diagram of the structure of a user multi-feature intelligent recommendation system provided by the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In addition, the technical features in the various embodiments or single embodiments provided by the present invention can be arbitrarily combined with each other to form a feasible technical solution. This combination is not subject to the constraints of the sequence of steps and / or the structural composition mode, but must be based on the ability of ordinary technicians in this field to achieve. When the combination of technical solutions is contradictory or cannot be achieved, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0021] Figure 1 A flowchart of a user multi-feature intelligent recommendation method provided by the present invention, such as Figure 1 As shown, the method includes:

[0022] Step 1: Obtain user historical behavior feature data, which includes user's online browsing, clicking and purchasing behavior data.

[0023] It is understandable that the collection of user behavior characteristics of browsing, clicking and purchasing goods online is called user historical behavior characteristic data, which includes basic information of users, such as gender, age, etc., as well as information of browsed, clicked and purchased goods, such as price and category of goods.

[0024] Step 2: construct a multi-interest user portrait based on the MIND model according to the user's historical behavior feature data, wherein the multi-interest user portrait expresses multiple categories of candidate commodities that the user is interested in.

[0025] It is understandable that the MIND model is used to construct multi-interest user portraits based on the collected user historical behavior feature data.

[0026] Exemplarily, the MIND model includes an embedding pooling layer, a multi-interest extraction layer, and a tag-aware attention layer; the multi-interest user portrait is constructed based on the MIND model according to the user historical behavior feature data, including:

[0027] The embedding pooling layer is used to embed the user historical behavior feature data into a high-dimensional vector space;

[0028] The multi-interest extraction layer processes the user historical behavior feature data based on the dynamic routing mechanism of the capsule network, dynamically clusters different interest categories of the user, and extracts multiple interest embedding vectors from the user historical behavior feature data, wherein each interest embedding vector corresponds to an interest category of the user;

[0029] The label-aware attention layer is used to calculate the similarity between multiple interest embedding vectors and each product vector through an attention mechanism; and to take products with similarity greater than a preset similarity threshold as candidate products of interest to the user.

[0030] Specifically, in order to effectively capture the diverse needs of users, a multi-interest user portrait strategy can be adopted. This method decomposes users' interests into multiple sub-portraits by identifying different behavioral categories. For example, female users may have both their own shopping needs (such as women's clothing and skin care products) and purchasing needs (such as buying clothes or baby products for men). By classifying these needs, corresponding products can be dynamically recommended according to specific situations without affecting the quality of recommendations due to mixed needs.

[0031] To further implement this strategy, the MIND model (Multi-Interest Network with Dynamic Routing) can be introduced. The MIND model can dynamically extract multiple points of interest of users and weight these interests through the attention mechanism to ensure that the recommendation results accurately match the current needs of users. The MIND model mainly includes the Embedding & Pooling Layer, the Multi-Interest Extractor Layer, and the Label-aware Attention Layer.

[0032] The embedding pooling layer embeds the user's behavioral features into a high-dimensional vector space;

[0033] The multi-interest extraction layer processes the user's behavioral feature data based on the dynamic routing mechanism of the capsule network. Through this mechanism, the user's different interest categories can be dynamically clustered and multiple interest embedding vectors of the user can be extracted, including an interest embedding vector. These vectors will be used to represent the user's multi-interest features to ensure that the system can capture the correlation between different user behaviors. The multi-interest extraction layer extracts N interest embedding vectors .

[0034] Among them, the label-aware attention layer uses the multi-head self-attention mechanism to calculate the similarity between multiple interest embedding vectors and each product vector. By mapping the interest embedding vectors to different subspaces, the label-aware attention layer can capture the different dimensions of user interests and assign weights to each subspace. Only when the similarity between the product vector and the interest embedding vector is greater than the preset similarity threshold, the product will be considered as a candidate product related to the user's current interest, ensuring a high match between the recommendation results and the user's interests.

[0035] Specifically, the label-aware attention layer calculates the similarity between multiple interest embedding vectors and each product vector through the attention mechanism, including:

[0036] ;

[0037] in, represents the i-th interest embedding vector, express The attention weights ensure that the most relevant points of interest are prioritized in the recommendation. Indicates similarity.

[0038] The similarity calculation formula is used to calculate the similarity between the user's N interest embedding vectors and the products of each category that the user has browsed, and the products whose similarity with the user's N interest embedding vectors is greater than the preset similarity threshold are regarded as candidate products that the user is interested in. Then, candidate products of multiple categories that a user is interested in are obtained, and a multi-interest user portrait is constructed.

[0039] Step 3: Calculate the first weight of each category of candidate products based on the user's behavioral interaction frequency with respect to each category of candidate products.

[0040] It is understandable that the above step 2 constructs a multi-interest user portrait. The user's behavior frequency is crucial in the recommendation, and the recommendation content can be dynamically adjusted according to the user's interaction frequency with different categories of goods.

[0041] Exemplarily, based on the user's behavioral interaction frequency with respect to each category of candidate commodities, calculating the first weight of each category of candidate commodities includes:

[0042] ;

[0043] in, Indicates the user's category The behavioral interaction frequency of candidate products, For users to classify The behavioral interaction frequency of candidate products, M represents the total number of categories of candidate products, Indicates category The first weight of the candidate product.

[0044] In this way, different recommendation weights can be assigned to different categories of goods based on the user's long-term behavior frequency. For example, if a user frequently purchases men's clothing over a long period of time, it means that he has a strong demand for purchasing on behalf of others, and the recommendation weight of this category of goods will be increased accordingly; while for occasional purchases, their weight can be reduced to avoid excessive influence of short-term behavior on recommendations.

[0045] Among them, user feedback behaviors (such as clicks, purchases, etc.) provide valuable data for optimizing user portraits. The portrait can be dynamically adjusted according to the user's actual operation. For example, when a user clicks on a certain type of product but does not buy it, the recommendation weight of such products can be reduced accordingly; and when the user buys a specific product, the portrait can be updated again to increase the recommendation weight of the relevant products. Specifically, assuming that the user has a certain type of product, the recommendation weight of the relevant product can be increased. The click rate is The purchase rate is , the first weight can be adjusted by the following formula, and the adjusted first weight is:

[0046] ;

[0047] in, For users to classify The purchase rate of candidate products, For users to classify The click-through rate of candidate products, and are the weight coefficients of purchase rate and click-through rate respectively.

[0048] Step 4: Calculate the second weight of each category of candidate products based on a time decay model according to the user's behavioral interaction time with each category of candidate products.

[0049] It is understandable that in addition to the frequency of behavior, the time factor also plays a key role in accurately identifying user needs. The time decay model can be used to deal with the timeliness of user behavior. The core idea of ​​the time decay model is that as time goes by, the impact of the user's past behavior on the recommendation gradually weakens. According to the user's behavioral interaction time for each category of candidate products, the expression for calculating the second weight of each category of candidate products based on the time decay model is:

[0050] ;

[0051] in, is the current time, For users to classify The time when the behavior of the candidate product occurs, is the time decay rate constant, For Category The second weight of the candidate product.

[0052] Smaller Values ​​indicate that the behavior influences longer retention, while larger A smaller value indicates that the influence of the behavior decays faster. Through this model, the user profile can be dynamically adjusted to ensure that recent behaviors have a greater weight in the recommendation, thereby better capturing the user's immediate needs.

[0053] Step 5: Calculate the comprehensive weight of the candidate commodities in each category according to the first weight and the second weight of the candidate commodities in each category.

[0054] Exemplarily, calculating the comprehensive weight of the candidate commodities in each category according to the first weight and the second weight of the candidate commodities in each category includes:

[0055] ;

[0056] in, For Category The comprehensive weight of candidate products, For Category The first weight of the candidate product, For Category The second weight of the candidate product, and is the coefficient.

[0057] Step 6: Based on the comprehensive weight of the candidate products in each category, select one or more categories of candidate products from all categories of candidate products and recommend them to the user.

[0058] It is understandable that the comprehensive weight of each category of candidate products is calculated according to step 5, and suitable products are screened out from the candidate products according to the size of the comprehensive weight and recommended to the user. Specifically, the candidate products of all categories are sorted from large to small according to the comprehensive weight, and the candidate products of a preset number of categories with the highest ranking are recommended to the user.

[0059] When faced with contradictory or diverse behaviors of users, such as users buying both men's and women's clothing, the system needs to effectively handle these conflicts. Weighted voting mechanism is an effective method that can label each behavior of users and assign different weights to different categories of behaviors based on factors such as frequency and time of behavior.

[0060] For example, when making recommendations based on the comprehensive weights of candidate products, if there are multiple categories of candidate products whose comprehensive weights are all relatively large and equal or similar, there will be a choice contradiction.

[0061] At this time, there are multiple recommendation strategies. For example, when the comprehensive weights of candidate products in m categories are equal or similar, multiple recommendation strategies can be combined during recommendation. For example, all the candidate products in these m categories can be recommended to the user at the same time, or one or more candidate products can be selected and recommended to the user.

[0062] So, which recommendation strategy is the best? The present invention adopts a weighted voting mechanism to select the recommendation category.

[0063] When there are m candidate products of different categories with equal or similar comprehensive weights, k recommendation strategies are generated based on the m candidate products of different categories;

[0064] The scores of candidate products in each category are calculated based on a weighted voting mechanism:

[0065] ;

[0066] in, is the number of different recommended strategies, is the k-th strategy pair category The comprehensive weight of candidate products, is the k-th strategy pair category Voting for candidate products, For category The ratings of candidate products;

[0067] One or more candidate products with high scores among m candidate products of different categories are recommended to the user.

[0068] See also Figure 2 , provides a user multi-feature intelligent recommendation system of the present invention, the system comprising:

[0069] The acquisition module 201 is used to acquire user historical behavior feature data, wherein the user historical behavior feature data includes user behavior data of browsing, clicking and purchasing goods online.

[0070] The construction module 202 is used to construct a multi-interest user portrait based on the MIND model according to the user's historical behavior feature data, and the multi-interest user portrait expresses multiple categories of candidate commodities that the user is interested in.

[0071] Among them, the construction module 202 is mainly used to: based on the user's historical behavioral feature data, use the MIND model to dynamically generate a multi-interest user portrait, wherein the MIND model includes an embedding pooling layer, a multi-interest extraction layer, and a label-aware attention layer. First, the system converts the user's behavioral characteristics into a high-dimensional vector representation through the embedding pooling layer; then, the multi-interest extraction layer extracts the user's multi-interest embedding vector through a dynamic clustering algorithm; finally, the label-aware attention layer generates a list of multi-interest candidate products for the user based on the feature weights of different products and the matching degree of the user's interests. This module can effectively capture the complex interest changes of users in multiple scenarios, ensuring the diversity and accuracy of the system's recommendation results.

[0072] The calculation module 203 is used to calculate the first weight of the candidate products in each category based on the user's behavioral interaction frequency with the candidate products in each category; and to calculate the second weight of the candidate products in each category based on the time attenuation model according to the user's behavioral interaction time with the candidate products in each category; and is also used to calculate the comprehensive weight of the candidate products in each category based on the first weight and the second weight of the candidate products in each category.

[0073] The recommendation module 204 is used to select one or more categories of candidate commodities from all categories of candidate commodities based on the comprehensive weight of each category of candidate commodities and recommend them to the user.

[0074] The recommendation module is responsible for intelligently screening all candidate products based on their comprehensive weights. The specific steps include:

[0075] Sort the candidate products in each category from high to low according to their comprehensive weight;

[0076] Through the weighted voting mechanism, multiple categories with similar weights are processed to generate multiple recommendation strategies;

[0077] Comprehensively consider the user's current interest hotspots, behavior frequency, and time decay.

[0078] It can be understood that the user multi-feature intelligent recommendation system provided by the present invention corresponds to the user multi-feature intelligent recommendation method provided by the aforementioned embodiments. The relevant technical features of the user multi-feature intelligent recommendation system can refer to the relevant technical features of the user multi-feature intelligent recommendation method, which will not be repeated here.

[0079] The embodiment of the present invention provides a user multi-feature intelligent recommendation method and system, which builds a multi-interest user portrait based on the MIND model, and accurately selects suitable categories of goods to recommend to users by combining a weighted model based on behavior frequency, a time decay model, and a weighted voting mechanism, and can more flexibly handle users' diverse behaviors and demand changes. The combination of these models and mechanisms not only improves the accuracy and relevance of recommended content, but also further improves user satisfaction and participation, ensuring that the recommendation system can still maintain efficient and accurate recommendation capabilities when facing complex user behaviors.

[0080] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0081] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0082] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0083] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0084] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0085] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0086] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A user multi-feature intelligent recommendation method, characterized in that: include: Acquire user historical behavior feature data, wherein the user historical behavior feature data includes user online browsing, clicking and purchasing behavior data; According to the user's historical behavior feature data, a multi-interest user portrait is constructed based on the MIND model, wherein the multi-interest user portrait expresses multiple categories of candidate commodities that the user is interested in; Calculate the first weight of each category of candidate products based on the user's behavioral interaction frequency with the candidate products of each category; According to the user's behavioral interaction time with each category of candidate products, a second weight of each category of candidate products is calculated based on a time decay model; Calculating a comprehensive weight of the candidate commodities in each category according to the first weight and the second weight of the candidate commodities in each category; Based on the comprehensive weight of candidate products in each category, select one or more candidate products from all categories and recommend them to users; The MIND model includes an embedding pooling layer, a multi-interest extraction layer, and a tag-aware attention layer; and constructing a multi-interest user portrait based on the MIND model according to the user's historical behavior feature data includes: The embedding pooling layer is used to embed the user historical behavior feature data into a high-dimensional vector space; The multi-interest extraction layer processes the user historical behavior feature data based on the dynamic routing mechanism of the capsule network, dynamically clusters different interest categories of the user, and extracts multiple interest embedding vectors from the user historical behavior feature data, wherein each interest embedding vector corresponds to an interest category of the user; The tag-aware attention layer is used to calculate the similarity between the multiple interest embedding vectors and each product vector through an attention mechanism; and to take products with similarity greater than a preset similarity threshold as candidate products of interest to the user; Calculating the comprehensive weight of the candidate commodities in each category according to the first weight and the second weight of the candidate commodities in each category includes: ； in, For Category The comprehensive weight of candidate products, For Category The first weight of the candidate product, For Category The second weight of the candidate product, and is the coefficient.

2. The user multi-feature intelligent recommendation method according to claim 1, characterized in that: The multiple interest embedding vectors extracted by the multi-interest extraction layer are expressed as , N is the number of interest embedding vectors, and the similarity between multiple interest embedding vectors and each product vector is calculated through the attention mechanism, including: ； in, represents the i-th interest embedding vector, express The attention weight, Indicates similarity.

3. The user multi-feature intelligent recommendation method according to claim 1, characterized in that: The calculating the first weight of each category of candidate commodities based on the user's behavioral interaction frequency with respect to each category of candidate commodities comprises: ； in, Indicates the user's category The behavioral interaction frequency of candidate products, For users to classify The behavioral interaction frequency of candidate products, M represents the total number of categories of candidate products, Indicates category The first weight of the candidate product.

4. The user multi-feature intelligent recommendation method according to claim 3, characterized in that: The method also includes the following steps: The click-through rate and purchase rate of candidate products for categories The first weight of the candidate product After adjustment, the first weight is: ； in, For users to classify The purchase rate of candidate products, For users to classify The click-through rate of candidate products, and are the weight coefficients of purchase rate and click-through rate respectively.

5. The user multi-feature intelligent recommendation method according to claim 1, characterized in that: The calculating the second weight of each category of candidate products based on the time decay model according to the user's behavioral interaction time with each category of candidate products includes: ； in, is the current time, For users to classify The time when the behavior of the candidate product occurs, is the time decay rate constant, For Category The second weight of the candidate product.

6. The user multi-feature intelligent recommendation method according to claim 1, characterized in that: The step of selecting one or more categories of candidate products from all categories of candidate products based on the comprehensive weight of each category of candidate products and recommending them to the user includes: The candidate products of all categories are sorted from large to small according to their comprehensive weights, and the candidate products of a preset number of categories with the highest rankings are recommended to the user.

7. The user multi-feature intelligent recommendation method according to claim 6, characterized in that: The step of selecting one or more categories of candidate products from all categories of candidate products based on the comprehensive weight of each category of candidate products and recommending them to the user includes: When there are m candidate products of different categories with equal or similar comprehensive weights, k recommendation strategies are generated based on the m candidate products of different categories; The scores of candidate products in each category are calculated based on a weighted voting mechanism: ； in, is the number of different recommended strategies, is the k-th strategy pair category The comprehensive weight of candidate products, is the k-th strategy pair category Voting for candidate products, For category The ratings of candidate products; One or more candidate products with high scores among m candidate products of different categories are recommended to the user.

8. A user multi-feature intelligent recommendation system, characterized in that: include: An acquisition module is used to acquire user historical behavior feature data, wherein the user historical behavior feature data includes user behavior data of browsing, clicking and purchasing goods online; A construction module is used to construct a multi-interest user portrait based on the MIND model according to the user's historical behavior feature data, wherein the multi-interest user portrait expresses multiple categories of candidate commodities that the user is interested in; A calculation module, used to calculate a first weight of each category of candidate commodities based on the user's behavioral interaction frequency with respect to each category of candidate commodities; and calculating the second weight of each category of candidate products based on a time decay model according to the user's behavioral interaction time with each category of candidate products; Also used for calculating the comprehensive weight of the candidate commodities in each category according to the first weight and the second weight of the candidate commodities in each category; A recommendation module is used to select one or more candidate products from all categories of candidate products based on the comprehensive weight of the candidate products in each category and recommend them to the user; The MIND model includes an embedding pooling layer, a multi-interest extraction layer, and a tag-aware attention layer; and constructing a multi-interest user portrait based on the MIND model according to the user's historical behavior feature data includes: The embedding pooling layer is used to embed the user historical behavior feature data into a high-dimensional vector space; The multi-interest extraction layer processes the user historical behavior feature data based on the dynamic routing mechanism of the capsule network, dynamically clusters different interest categories of the user, and extracts multiple interest embedding vectors from the user historical behavior feature data, wherein each interest embedding vector corresponds to an interest category of the user; The tag-aware attention layer is used to calculate the similarity between the multiple interest embedding vectors and each product vector through an attention mechanism; and to take products with similarity greater than a preset similarity threshold as candidate products of interest to the user; Calculating the comprehensive weight of the candidate commodities in each category according to the first weight and the second weight of the candidate commodities in each category includes: ； in, For Category The comprehensive weight of candidate products, For Category The first weight of the candidate product, For Category The second weight of the candidate product, and is the coefficient.

Images