Product recommended processing methods, devices, equipment and storage media

Abstract

The embodiment of the present invention provides a method, device, equipment and storage medium for processing product recommendations, the method comprising: receiving a product search request from a first user; determining, according to the product search request, a target theme scene to which a first scene word included in the product search request belongs; obtaining products in a product pool corresponding to the target theme scene as products to be recommended; and recommending the products to be recommended to the first user. Since the product pool corresponding to each theme scene is obtained based on the user's historical behavior data, a large number of theme scenes and their corresponding scene words can be effectively obtained, so as to better and more accurately recommend products to users, solving the problem that the existing technology has few scene words and cannot accurately recommend products to users.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method, device, equipment and storage medium for processing product recommendations. Background Art

[0002] With the rapid development of natural language processing technology, its application in the field of e-commerce is becoming more and more extensive. When users buy products on e-commerce platforms, they usually search for specific products. E-commerce platforms recommend corresponding products to users based on their search requests for them to choose and buy.

[0003] As the quality of life continues to improve, users' demands for purchasing products are also becoming more diversified. Simple search services may not be able to meet users' needs. For example, when users are in a certain scenario and are not sure what product to buy, product recommendations based on thematic scenarios can provide users with more accurate and high-quality product lists.

[0004] In the prior art, scene words corresponding to theme scenes are manually summarized, which is inefficient and results in relatively few scene words. It is therefore impossible to accurately recommend corresponding products to users in response to their search needs for various scenes. Summary of the invention

[0005] The embodiments of the present invention provide a method, device, equipment and storage medium for processing product recommendations to solve the defects of low recommendation accuracy in the prior art.

[0006] In a first aspect, an embodiment of the present invention provides a method for processing product recommendations, including:

[0007] receiving a product search request from a first user;

[0008] Determining, according to the product search request, a target theme scene to which the first scene word included in the product search request belongs;

[0009] Obtain products in a product pool corresponding to the target theme scene as products to be recommended, where the product pool corresponding to each theme scene is obtained based on user historical behavior data;

[0010] Recommend the product to be recommended to the first user.

[0011] Optionally, a product pool corresponding to each theme scenario is obtained based on the user's historical behavior data, including:

[0012] Based on the user review content data, determine the first group of scene words and the products associated with each scene word;

[0013] Taking the first group of scene words as target scene words, merging the target scene words by semantic similarity to obtain at least one theme scene, each theme scene corresponding to at least one target scene word;

[0014] According to the target scene words corresponding to each theme scene and the products associated with each target scene word, the product pool corresponding to each theme scene is determined.

[0015] Optionally, obtaining a product pool corresponding to each theme scenario based on user historical behavior data also includes:

[0016] Based on the user's historical search data, determine the second group of scene words;

[0017] removing duplicates from the first group of scene words and the second group of scene words as the target scene words;

[0018] The method further comprises:

[0019] Based on the user's historical click behavior data and historical order data, determine the target products involved and build a binary relationship between the target products;

[0020] Determining an adjacency matrix of the target products according to the binary relationships between the target products;

[0021] Determining product features of the target product based on user review content of the target product and attributes of the target product;

[0022] Based on the adjacency matrix and product features of the target products, a trained scene classification model is used to determine the theme scenes associated with each of the target products;

[0023] Determining product information corresponding to each theme scene according to the scene words corresponding to each theme scene and the products associated with each scene word includes:

[0024] According to the scene words corresponding to each theme scene and the products associated with each scene word, as well as the theme scenes associated with each target product, a product pool corresponding to each theme scene is determined.

[0025] Optionally, determining the product features of the target product according to the user review content of the target product and the product attributes of the target product includes:

[0026] Performing word segmentation processing on the user review content of the target product to obtain a word segmentation result;

[0027] According to the word segmentation results, the TF-IDF algorithm is used to determine the keyword features of the target product;

[0028] Determining the attribute characteristics of the target product according to the product attributes of the target product;

[0029] The keyword features and the attribute features are used as product features of the target product.

[0030] Optionally, the method further comprises:

[0031] One-hot encoding is performed on the product features of the target product to obtain the product feature code corresponding to the target product.

[0032] Optionally, the determining of the theme scenes associated with each of the target products using a trained scene classification model based on the adjacency matrix and product features of the target products includes:

[0033] Input the adjacency matrix and product feature codes into the trained GCN classification model to obtain classification results of each target product, wherein the classification results include the probability that each target product belongs to a different theme scene;

[0034] Based on the classification results of each of the target products, a theme scene associated with each of the target products is determined.

[0035] Optionally, the method further comprises:

[0036] Get product attribute information;

[0037] Based on the product attribute information, determining a third group of scene words and products associated with each scene word;

[0038] The first group of scene words, the second group of scene words, and the third group of scene words are removed after being duplicated and used as the target scene words.

[0039] Optionally, the step of performing semantically similar merging on the target scene words to obtain at least one theme scene includes:

[0040] The trained word2vector model is used to construct the spatial vector corresponding to each target scene word;

[0041] Based on the spatial vectors corresponding to the target scene words, a spatial cosine similarity algorithm is used to determine the semantic similarity between the target scene words;

[0042] Semantic similarity merging is performed based on the semantic similarity between the target scene words to obtain at least one theme scene.

[0043] Optionally, the method further comprises:

[0044] Structural processing is performed on each theme scene to determine the superordinate scene to which each theme scene belongs;

[0045] Product recommendations are made to users based on the superordinate scenarios and thematic scenarios included in each superordinate scenario.

[0046] Optionally, after obtaining the product pool corresponding to each theme scene, the method further includes:

[0047] For the product pool corresponding to each theme scene, the products in the product pool are sorted according to preset rules to obtain a sorting result;

[0048] According to the ranking results, a preset number of products ranked high are used as products to be recommended.

[0049] Optionally, sorting the products in the product pool according to a preset rule to obtain a sorting result includes:

[0050] Obtain at least one factor of the sales volume, number of unique visitors, number of visits, favorable comment rate, and unit price of each product in the product pool;

[0051] Determine the score of each product based on the preset weight of each factor;

[0052] Sort the products according to the scores to obtain the sorting results.

[0053] An embodiment of the above invention has the following advantages or beneficial effects:

[0054] By combining user comment content data and user historical search data to mine more thematic scenarios, and based on user historical click behavior data and historical order data, we can effectively expand the products in the product pool corresponding to each thematic scenario and improve product coverage.

[0055] Another embodiment of the above invention has the following advantages or beneficial effects:

[0056] We also use the scenario classification model to classify more products into scenarios, further improving the coverage of hot-selling products in the product pool.

[0057] Another embodiment of the above invention has the following advantages or beneficial effects:

[0058] You can also sort the products in the product pool by word-of-mouth, and recommend a certain number of top-ranked products to users as products to be recommended, further improving the accuracy of recommendations and thus improving user experience.

[0059] In a second aspect, an embodiment of the present invention provides a product recommendation processing device, including:

[0060] A receiving module, configured to receive a product search request from a first user;

[0061] A determination module, configured to determine, according to the product search request, a target theme scene to which the first scene word included in the product search request belongs;

[0062] An acquisition module, used to acquire products in a product pool corresponding to the target theme scene as products to be recommended, wherein the product pool corresponding to each theme scene is obtained based on user historical behavior data;

[0063] A recommendation module is used to recommend the product to be recommended to the first user.

[0064] Optionally, the device further includes a processing module, configured to:

[0065] Based on the user review content data, determine the first group of scene words and the products associated with each scene word;

[0066] Taking the first group of scene words as target scene words, merging the target scene words by semantic similarity to obtain at least one theme scene, each theme scene corresponding to at least one target scene word;

[0067] According to the target scene words corresponding to each theme scene and the products associated with each target scene word, the product pool corresponding to each theme scene is determined.

[0068] Optionally, the processing module is further used to:

[0069] Based on the user's historical search data, determine a second group of scene words; remove duplicates from the first group of scene words and the second group of scene words as the target scene words;

[0070] The processing module is further used for:

[0071] Based on the user's historical click behavior data and historical order data, determine the target products involved and build a binary relationship between the target products;

[0072] Determining an adjacency matrix of the target products according to the binary relationships between the target products;

[0073] Determining product features of the target product based on user review content of the target product and attributes of the target product;

[0074] Based on the adjacency matrix and product features of the target products, a trained scene classification model is used to determine the theme scenes associated with each of the target products;

[0075] The processing module is specifically used for:

[0076] According to the scene words corresponding to each theme scene and the products associated with each scene word, as well as the theme scenes associated with each target product, a product pool corresponding to each theme scene is determined.

[0077] Optionally, the processing module is specifically used to:

[0078] Performing word segmentation processing on the user review content of the target product to obtain a word segmentation result;

[0079] According to the word segmentation results, the TF-IDF algorithm is used to determine the keyword features of the target product;

[0080] Determining the attribute characteristics of the target product according to the product attributes of the target product;

[0081] The keyword features and the attribute features are used as product features of the target product.

[0082] Optionally, the processing module is further used to:

[0083] One-hot encoding is performed on the product features of the target product to obtain the product feature code corresponding to the target product.

[0084] Optionally, the processing module is specifically used to:

[0085] Input the adjacency matrix and product feature codes into the trained GCN classification model to obtain classification results of each target product, wherein the classification results include the probability that each target product belongs to a different theme scene;

[0086] Based on the classification results of each of the target products, a theme scene associated with each of the target products is determined.

[0087] Optionally, the processing module is specifically used to:

[0088] Get product attribute information;

[0089] Based on the product attribute information, determining a third group of scene words and products associated with each scene word;

[0090] The first group of scene words, the second group of scene words, and the third group of scene words are removed after being duplicated and used as the target scene words.

[0091] Optionally, the processing module is specifically used to:

[0092] The trained word2vector model is used to construct the spatial vector corresponding to each target scene word;

[0093] Based on the spatial vectors corresponding to the target scene words, a spatial cosine similarity algorithm is used to determine the semantic similarity between the target scene words;

[0094] Semantic similarity merging is performed based on the semantic similarity between the target scene words to obtain at least one theme scene.

[0095] Optionally, the processing module is further configured to:

[0096] Structural processing is performed on each theme scene to determine the superordinate scene to which each theme scene belongs;

[0097] Product recommendations are made to users based on the superordinate scenarios and thematic scenarios included in each superordinate scenario.

[0098] Optionally, the processing module is further configured to:

[0099] For the product pool corresponding to each theme scene, the products in the product pool are sorted according to preset rules to obtain a sorting result;

[0100] According to the ranking results, a preset number of products ranked high are used as products to be recommended.

[0101] Optionally, the processing module is specifically used to:

[0102] Obtain at least one factor of the sales volume, number of unique visitors, number of visits, favorable comment rate, and unit price of each product in the product pool;

[0103] Determine the score of each product based on the preset weight of each factor;

[0104] Sort the products according to the scores to obtain the sorting results.

[0105] In a third aspect, an embodiment of the present invention provides an electronic device, including: at least one processor and a memory;

[0106] The memory stores computer-executable instructions;

[0107] The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs the method described in the first aspect and various possible designs of the first aspect.

[0108] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer execution instructions. When a processor executes the computer execution instructions, the method described in the first aspect and various possible designs of the first aspect are implemented.

[0109] The product recommendation processing method, device, equipment and storage medium provided in the embodiments of the present invention determine the target theme scene to which the first scene word included in the product search request belongs based on the product search request of the first user, obtain the products in the product pool corresponding to the target theme scene as the products to be recommended, and recommend the products to be recommended to the user. Since the product pool corresponding to each theme scene is obtained based on the user's historical behavior data, a large number of theme scenes and their corresponding scene words can be effectively obtained, so as to recommend products to users better and more accurately, thereby solving the problem that the prior art has few scene words and cannot accurately recommend products to users. BRIEF DESCRIPTION OF THE DRAWINGS

[0110] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0111] Figure 1 A schematic diagram of the architecture of a processing system based on an embodiment of the present invention;

[0112] Figure 2 A flowchart of a product recommendation processing method provided in an embodiment of the present invention;

[0113] Figure 3 A flowchart of a product recommendation processing method provided by another embodiment of the present invention;

[0114] Figure 4 A schematic flow chart of another product recommended processing method provided by another embodiment of the present invention;

[0115] Figure 5 A schematic diagram of a binary relationship network provided by an embodiment of the present invention;

[0116] Figure 6 A schematic diagram of one-hot encoding provided by an embodiment of the present invention;

[0117] Figure 7 A schematic diagram of a network structure of a scene classification model provided by an embodiment of the present invention;

[0118] Figure 8 A schematic diagram of the word2vector model application process provided by an embodiment of the present invention;

[0119] Fig. 9 A schematic diagram of the structure of a processing device recommended for a product provided in an embodiment of the present invention;

[0120] Fig.10 A schematic diagram of the structure of a processing device recommended for a product provided by another embodiment of the present invention;

[0121] Fig.11 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention.

[0122] The above drawings have shown clear embodiments of the present invention, which will be described in more detail below. These drawings and text descriptions are not intended to limit the scope of the disclosed concept in any way, but to illustrate the concept of the present invention for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0123] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution 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 of 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.

[0124] First, the terms involved in the present invention are explained:

[0125] word2vector: is a tool that converts words into vectors. It is essentially a word clustering method and a means to achieve word semantic inference, sentence sentiment analysis, etc.

[0126] One-hot encoding: It is the process of converting categorical variables into a form that is easy to use for machine learning algorithms. It usually converts feature variables into vectors containing only 0 and 1.

[0127] Jieba word segmentation: It can also be called jieba word segmentation, which refers to the process of dividing a sequence of Chinese characters into individual words. Word segmentation is the process of recombining a continuous sequence of characters into a word sequence according to certain specifications;

[0128] TF-IDF: is a weighted technology used for information retrieval and data mining. TF is term frequency, IDF is inverse document frequency, and the formula is: TF*IDF. The main idea is: if a word or phrase appears in an article with a high frequency TF and rarely appears in other articles, it is considered that this word or phrase has good category differentiation ability.

[0129] GCN: Graph Convolutional Network, graph convolutional network, is a feature extractor whose object is graph network. GCN cleverly designed a method to extract features from graph network, so that these features can be used to classify nodes, graphs, and edges of graph networks, and can also obtain graph embedding representation. GCN is divided into spectral convolution and spatial domain convolution. In terms of efficiency, the amount of spectral convolution calculation will increase sharply with the size of the graph, because the model needs to calculate feature vectors or process large graphs at the same time, making it difficult for the model to process or scale large graphs in parallel. Spatial domain convolution directly aggregates neighbor nodes in the graph domain, so it has the potential to process large graphs by calculating a batch of data instead of calculating on the entire graph. If the number of neighbor nodes increases, the efficiency can be improved through sampling technology. In terms of versatility, spectral convolution assumes that the graph is fixed, so the generalization performance for new or different graphs is poor. Spatial domain convolution performs local graph convolution on each node, and weights can be easily shared between different positions and structures. Flexibility: spectral convolution is only applicable to undirected graphs. The only way to use spectral methods for directed graphs is to convert the directed graphs into undirected graphs, because there is no clear definition of the Laplacian matrix of the directed graph. Spatial domain convolution can merge the input into the aggregation function, so it is more flexible in processing multi-source inputs such as edge features and edge directions.

[0130] With the rapid development of natural language processing technology, its application in the field of e-commerce is becoming more and more extensive. When users buy products on e-commerce platforms, they usually search for specific products. E-commerce platforms recommend corresponding products to users based on user search requests for users to choose to buy. With the continuous improvement of the quality of life, users' needs for purchasing products have also become diversified. Simple search services may not meet the needs of users. For example, when users are in a certain scenario and are not sure what products they should buy, product recommendations based on theme scenarios can provide users with more accurate and higher-quality product lists. In the prior art, the scene words corresponding to the theme scenes are manually summarized, which is inefficient and the summarized scene words are relatively few. For users' search needs for various scenarios, it is impossible to accurately recommend corresponding products to users.

[0131] In view of the above problems existing in the prior art, the inventor conducted creative research and found that the prior art is a manual summary of theme scenes, which has low work efficiency on the one hand and few scene words summarized on the other hand, and cannot accurately recommend products for users' searches in various scenes. In order to solve the above problems, the inventor creatively discovered that there are a large number of relevant scene words in the user's historical behavior data. For example, the corresponding scene words are often easy to exist in the user's comments on the purchased products, which provides a good data support for the mining of scene words. Therefore, the embodiment of the present invention provides a product recommendation processing method, which mines scene words based on the user's historical behavior data, determines the product pool corresponding to different theme scenes, responds to user search requests based on the product pool corresponding to different theme scenes, and recommends products to users, which effectively improves the accuracy of product recommendations, thereby improving user experience.

[0132] The product recommendation processing method provided by the embodiment of the present invention is applicable to application scenarios where it is necessary to recommend products to users based on scenarios required by the users. Figure 1 As shown, it is a schematic diagram of the architecture of the processing system based on the embodiment of the present invention. The processing system may include an electronic device and at least one terminal. The terminal may be a mobile phone, a laptop computer, a desktop computer or other devices. The electronic device may be a server. The user accesses the purchase platform through the terminal, searches for products in the product search interface, and the terminal sends a product search request to the server. The server receives the product search request, and can determine the target theme scene to which the first scene word included in the product search request belongs, obtain the products in the product pool corresponding to the target theme scene as the recommended products, and recommend the recommended products to the user, wherein the product pool corresponding to each theme scene is obtained based on the user's historical behavior data, and the user's historical behavior data may include user's historical search data, user's comment content data (or high-quality comment content data), user's historical click behavior data, user's historical order data, etc., which can effectively increase the amount of scene words corresponding to the theme scene, and recommend products to users better and more accurately. Optionally, the processing system may also include other servers, such as servers for maintaining historical data of e-commerce platforms, and the electronic device obtains relevant data from other servers, such as user's historical behavior data. User's historical behavior data may also be pre-stored in the electronic device.

[0133] In addition, the terms "first", "second", etc. are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. In the description of the following embodiments, the meaning of "multiple" is more than two, unless otherwise clearly and specifically defined.

[0134] The following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments. The embodiments of the present invention will be described below in conjunction with the accompanying drawings.

[0135] An embodiment of the present invention provides a method for processing product recommendations, which is used to recommend products to users based on scenarios. The execution subject of this embodiment is a device for processing product recommendations, which can be set in an electronic device, specifically, in a server.

[0136] like Figure 2 FIG. 1 is a flow chart of a method for processing product recommendations provided in this embodiment, the method comprising:

[0137] Step 101: Receive a product search request from a first user.

[0138] Specifically, the first user can be any user. The user can log in to the e-commerce platform through the terminal to browse or purchase products. When the user does not know what product to buy in certain scenarios or wants to know what products are available in the scenario, the user can search through the product search function of the e-commerce platform, such as entering the corresponding scenario keyword (referred to as scenario word) in the search box, clicking search, and the terminal can obtain the scenario word entered by the user, send a product search request to the server, and the server can receive the product search request of the first user. The product search request includes the first scenario word entered by the first user, such as "travel", "game", "home" and other scenario words.

[0139] Step 102: Determine, according to the product search request, the target theme scene to which the first scene word included in the product search request belongs.

[0140] Specifically, after receiving a product search request, the scene word in the product search request can be obtained, which is called the first scene word for distinguishing the description. The first scene word is matched with the scene words corresponding to each pre-acquired theme scene to determine the theme scene to which the first scene word belongs, which is called the target theme scene for distinguishing the description. The pre-acquired theme scenes are obtained based on user historical behavior data mining. User historical behavior data may include user historical search data, user comment content data, and other related behavior data.

[0141] Exemplarily, some scene words can be mined from the user comment content data, and a batch of theme scenes can be obtained by semantic similarity merging. For example, the comments on a keyboard product in the user comment content are that playing games is very smooth and the operation is sensitive, etc., and the scene word "game" can be obtained, and it can be obtained that the keyboard product is associated with the game. For the mined scene words, there may be multiple scene words expressing the same semantics, so the mined scene words can be merged by semantic similarity, and multiple scene words representing the same semantics can be classified into a theme scene. Each theme scene can be represented by one of the scene words it represents, or a scene identifier can be reset for it, which can be set according to actual needs. For example, scene words such as "home", "home", and "home daily life" need to be merged by semantic similarity, and the merged scene corresponds to a theme scene. For example, the theme scene is represented by "home", or it can be represented by an identifying symbol. The corresponding relationship between each theme scene and the scene word it represents can be stored, or it can be called a scene word library corresponding to the theme scene. For example: Theme scene 1 = [scene word 1, scene word 2, ..., scene word N]. The specific storage method can be set according to actual needs.

[0142] Optionally, the user's historical behavior data and product attributes may be combined to jointly mine scene words and merge to obtain theme scenes.

[0143] Step 103, obtaining products in a product pool corresponding to the target theme scene as products to be recommended.

[0144] Among them, each theme scenario and the corresponding product pool are obtained based on user historical behavior data.

[0145] Specifically, after determining the target theme scene to which the first scene word belongs, the products in the product pool corresponding to the target theme scene can be obtained as the recommended products. Specifically, each theme scene corresponds to a corresponding product pool, and the product pool includes product information or a product list belonging to the theme scene. The product pool corresponding to each theme scene can also be obtained based on user historical behavior data.

[0146] For example, when mining scene words based on user review content data, the association relationship between the scene words and the products reviewed by the users can be obtained, so that the products associated with these scene words can be added to the product pool of the theme scene to which the scene words belong;

[0147] For example, when mining scene words based on user historical search data, a scene classification model can be trained based on the theme scenes to which each scene word belongs and the pre-labeled products, and based on the scene classification model, more products can be classified into corresponding theme scenes and added to the product pool corresponding to the corresponding theme scenes.

[0148] Optionally, some scene words can be mined based on product attributes to expand the existing scene word library.

[0149] Optionally, various other methods may be used to mine scene words, and new theme scenes and scene word libraries corresponding to each theme scene may be continuously supplemented.

[0150] Optionally, all products in the product pool corresponding to the target theme scene may be used as products to be recommended, or part of the products may be used as products to be recommended, which may be set specifically according to actual needs.

[0151] For example, the products in the product pool can be sorted according to certain requirements (such as one or more of sales volume, praise rate, price, number of independent visitors, number of visits, etc.), and the sorted product list can be used as products to be recommended, or a preset number of products with the highest ranking can be used as products to be recommended.

[0152] Step 104: recommend the product to be recommended to the first user.

[0153] Specifically, after the products to be recommended are determined, the products to be recommended may be recommended to the first user. For example, the list of products to be recommended may be sent to the terminal of the first user so that the terminal displays the list of products to be recommended to the first user.

[0154] The product recommendation processing method provided in this embodiment determines the target theme scene to which the first scene word included in the product search request belongs according to the product search request of the first user, obtains the products in the product pool corresponding to the target theme scene as the products to be recommended, and recommends the products to the user. Since the product pool corresponding to each theme scene is obtained based on the user's historical behavior data, a large number of theme scenes and their corresponding scene words can be effectively obtained, so as to recommend products to users better and more accurately, and solves the problem that the prior art has few scene words and cannot accurately recommend products to users.

[0155] Another embodiment of the present invention further supplements the method provided by the above embodiment.

[0156] like Figure 3 , which is a flow chart of a method for processing product recommendations provided in this embodiment.

[0157] As an implementable manner, based on the above embodiment, optionally, obtaining a product pool corresponding to each theme scene based on the user's historical behavior data specifically includes:

[0158] Step 2011: Determine a first group of scene words and products associated with each scene word based on the user review content data.

[0159] Specifically, some scene words can be mined from the user comment content data, which are called the first group of scene words in order to distinguish the description. The specific mining method can be to obtain high-quality comment content, and obtain the first group of scene words through word segmentation, filtering and screening. Since the user comment content data is the comments of each user on each product, the product corresponding to each comment content can be obtained, that is, the scene words contained in the comment content and the product association corresponding to the comment content can be obtained. It is also possible to match the obtained scene words with the comment content of more other products to mine more related products. It is also possible to match the obtained scene words with the product names, product attributes, etc. of more other products to mine more related products. The specific method of determining related products can be set according to actual needs.

[0160] For example, if a user's review of a keyboard product says that it is very easy to play games and has sensitive operation, etc., we can get the scene word "game" and find that the keyboard product is associated with games. We can match the scene word "game" with more other products. It is possible that some products have the scene word "game" in their product attributes, such as mouse, mobile phone, notebook, etc. Therefore, we can mine more products associated with the scene word "game".

[0161] Step 2012: taking the first group of scene words as target scene words.

[0162] Step 2013: merge the target scene words based on semantic similarity to obtain at least one theme scene, and each theme scene corresponds to at least one target scene word.

[0163] Specifically, since there may be multiple scene words that express the same semantics for the mined scene words, after obtaining the first group of scene words, the first group of scene words can be used as the target scene words, and the target scene words can be merged by semantic similarity, and multiple scene words representing the same semantics can be classified into a theme scene. Each theme scene can be represented by one of the scene words it represents, or a scene identifier can be reset for it, which can be set according to actual needs. For example, scene words such as "home", "home", and "daily life at home" need to be merged by semantic similarity, and the merged scene corresponds to a theme scene. For example, the theme scene is represented by "home", or it can be represented by an identifying symbol. The correspondence between each theme scene and the scene word it represents can be stored, or it can be called a scene word library corresponding to the theme scene. For example: Theme scene 1 = [scene word 1, scene word 2, ..., scene word N]. The specific storage method can be set according to actual needs.

[0164] Step 2014, determining a product pool corresponding to each theme scene according to the target scene words corresponding to each theme scene and the products associated with each target scene word.

[0165] Specifically, after obtaining each theme scene, the target scene words corresponding to each theme scene, and the products associated with each target scene word, the product pool corresponding to each theme scene can be determined. Specifically, for a theme scene, it may correspond to one or more target scene words, which can be called a scene word library. Each target scene word may be associated with one or more products. Then, the products associated with each scene word corresponding to the theme scene are added to the product pool corresponding to the theme scene.

[0166] like Figure 4 FIG. 1 is a flow chart of another method for processing product recommendations provided in this embodiment.

[0167] Optionally, obtaining a product pool corresponding to each theme scenario based on user historical behavior data also includes:

[0168] Step 2021, determine a second group of scene words based on the user's historical search data.

[0169] Specifically, when a user searches for a product, the user's search behavior data can be recorded as subsequent user history search data. For example, which keywords the user searched for. The user's history search data can be filtered according to preset filtering rules to obtain a second set of scene words. The filtering rules can, for example, remove obviously invalid keywords. Some keywords that are obviously not scene words can also be preset for filtering. Specific filtering rules can be set according to actual needs.

[0170] Step 2022: Remove duplicates from the first group of scene words and the second group of scene words and use them as target scene words.

[0171] There may be duplication of keywords mined in different ways, so the first set of scene words and the second set of scene words after deduplication can be used as target scene words for subsequent semantic similarity merging, and the theme scene obtained based on the merging can be used for training to obtain a scene classification model.

[0172] Accordingly, the method further includes:

[0173] Step 2031, based on the user's historical click behavior data and historical order data, determine the target products involved and build a binary relationship between the target products.

[0174] Specifically, in order to expand the product pool of each theme scene, products associated with scene words related to each theme scene can be mined based on user historical behavior data. Specifically, the products involved in the click behavior and orders can be determined based on user historical click behavior data and historical order data (called target products for distinguishing descriptions), and a binary relationship between the target products can be constructed.

[0175] For example, for the user's historical click behavior data, the products clicked twice within a preset time can form a binary group. For example, if the user clicks on a notebook product A at 10:00 and clicks on a mouse product B at 10:02, the notebook product A and the mouse product B form a binary group.<A,B> .

[0176] For example, for historical order data, multiple products included in an order can form multiple tuples. For example, if an order includes products A, B, C, and D, it can form a tuple<A,B> ,<A,C> ,<A,D> ,<B,C> ,<B,D> ,<C,D> The binary relationships between target products may form a network, with products as vertices, such as product identifiers (such as product IDs) as vertices, and the relationships where two products appear in the same binary as edges.

[0177] For example, Figure 5 , which is a schematic diagram of a binary relationship network provided in this embodiment.

[0178] Step 2032, determining the adjacency matrix of the target products according to the binary relationships between the target products.

[0179] Specifically, the proximity matrix of the target products can be determined according to the binary relationships between the target products, which can be used as an input of the scene classification model.

[0180] For example, Figure 5 The adjacency matrix corresponding to the binary relationship network shown is as follows:

[0181]

[0182] Step 2033, determining the product features of the target product according to the user review content of the target product and the attributes of the target product.

[0183] Steps 2032 and 2033 are performed in no particular order.

[0184] Specifically, the review content can be segmented, and the keywords related to the target product features can be calculated to extract keyword features. Based on the target product attributes, product attribute features can be extracted, and product features can be constructed by combining keyword features and product attribute features. The product features can also be one-hot encoded as another input to the scene classification model.

[0185] Step 2034, based on the adjacency matrix and product features of the target products, the trained scene classification model is used to determine the theme scenes associated with each target product.

[0186] Specifically, after the input of the scene classification model is determined, the adjacency matrix and product feature coding obtained above can be input into the scene classification model to obtain the classification result, and the theme scene to which each target product belongs can be determined according to the classification result.

[0187] Accordingly, according to the scene words corresponding to each theme scene and the products associated with each scene word, the product information corresponding to each theme scene is determined, specifically including:

[0188] Step 2035, determining a product pool corresponding to each theme scene based on the scene words corresponding to each theme scene and the products associated with each scene word, as well as the theme scenes associated with each target product.

[0189] Specifically, according to the scene words corresponding to each theme scene and the products associated with each scene word, these products are added to the product pool corresponding to each theme scene, and each target product is added to the product pool corresponding to its associated theme scene.

[0190] By combining user comment content data and user historical search data to mine more thematic scenarios, and based on user historical click behavior data and historical order data, the products in the product pool corresponding to each thematic scenario can be effectively expanded to improve product coverage.

[0191] Optionally, the product features of the target product are determined based on the user review content of the target product and the product attributes of the target product, specifically including:

[0192] Step 2041, perform word segmentation processing on the user review content of the target product to obtain a word segmentation result.

[0193] Step 2042, based on the word segmentation results, the TF-IDF algorithm is used to determine the keyword features of the target product.

[0194] Step 2043, determining the attribute characteristics of the target product according to the product attributes of the target product.

[0195] Step 2044, using the keyword features and attribute features as product features of the target product.

[0196] Specifically, Jieba word segmentation can be used to segment the user comments of the target product to obtain the word segmentation results. The specific word segmentation operation is an existing technology and can be set according to actual needs. After word segmentation, the TF-IDF algorithm can be used to calculate and obtain keyword features related to product features. And according to the product attributes of the target product, the attribute features of the target product are determined. Keyword features are combined with product attribute features to construct product features (i.e., product feature set). For example, the product features of a certain product are [ultra-high definition, multi-function, large size].

[0197] Optionally, the method may further include:

[0198] Step 2051, one-hot encoding is performed on the product features of the target product to obtain the product feature code corresponding to the target product.

[0199] Specifically, one-hot encoding is used to encode the product features of the target product into 0 and 1. Each feature element in the product feature is encoded separately.

[0200] For example, Figure 6 FIG. 1 is a schematic diagram of one-hot encoding provided in this embodiment. Here, the feature element “ultra-high definition” is taken as an example.

[0201] Accordingly, based on the adjacency matrix and product features of the target products, the trained scenario classification model is used to determine the theme scenarios associated with each target product, including:

[0202] Step 2052, input the adjacency matrix and product feature encoding into the trained GCN classification model to obtain the classification results of each target product, and the classification results include the probability that each target product belongs to different theme scenes.

[0203] Specifically, the GCN classification model is pre-trained, and some user historical click behavior data and historical order data can be extracted to determine the adjacency matrix of the training product according to the above process, and the user comment content data of the training product and the attributes of the training product can be extracted to construct the training product features according to the above steps. The GCN network is used to train the scene classification model, and the classification label is the above-determined theme scene, so the theme scene needs to be one-hot encoded.

[0204] After training the scene classification model, the adjacency matrix and product feature encoding of the target product are input into the scene classification model to obtain the classification results of the target product. The classification results include the probability that each target product belongs to different theme scenes.

[0205] For example, Figure 7 As shown, it is a schematic diagram of the network structure of the scene classification model provided in this embodiment. The network structure of the scene classification model includes a first convolution layer, a RELU activation function, a second convolution layer, and a softmax function, and the output is the probability that the target product belongs to each theme scene. Among them, the corresponding network parameters and number of layers for the first convolution layer and the second convolution layer can be set according to actual needs.

[0206] Step 2053: Determine the theme scene associated with each target product based on the classification results of each target product.

[0207] After the classification results of each target product are obtained, the theme scene associated with each target product can be determined according to the classification results, that is, which theme scene each target product belongs to. Specifically, for a target product, the theme scene with the highest probability can be selected as the theme scene associated with the target product.

[0208] For example, the probabilities that target product 1 belongs to theme scene 1, theme scene 2 and theme scene 3 are 0.1, 0.2 and 0.7 respectively. Then it is determined that target product 1 is associated with theme scene 3, and the target product needs to be added to the product pool corresponding to theme scene 3.

[0209] Optionally, the method may further include:

[0210] Step 2061, obtaining product attribute information.

[0211] Step 2062: Determine a third group of scene words and products associated with each scene word based on the product attribute information.

[0212] Step 2063, removing duplicates from the first group of scene words, the second group of scene words, and the third group of scene words as target scene words.

[0213] Specifically, for some products, some scene words may be included in their attribute information. The corresponding scene words can be extracted from their attribute information, which can be called the third group of scene words. Based on these scene words, they can be matched with more other product attribute information to dig out more related products and expand the corresponding theme scenes and corresponding product pools. In the actual processing process, the first group of scene words, the second group of scene words, and the third group of scene words can be removed and used as target scene words, and subsequent semantic similarity merging is performed to obtain each theme scene. Other subsequent processing is consistent with the above and will not be repeated here.

[0214] Optionally, semantically similar words are merged on the target scene words to obtain at least one theme scene, specifically including:

[0215] Step 2071, use the trained word2vector model to construct the spatial vector corresponding to each target scene word.

[0216] Step 2072: Based on the space vectors corresponding to the target scene words, a spatial cosine similarity algorithm is used to determine the semantic similarity between the target scene words.

[0217] Step 2073: perform semantic similarity merging based on the semantic similarity between the target scene words to obtain at least one theme scene.

[0218] Specifically, it is necessary to pre-train a word2vector model for the construction of a spatial vector. The training process of the word2vector model is consistent with the prior art and will not be repeated here. After constructing the spatial vector corresponding to the target scene word, the spatial cosine similarity algorithm can be used to determine the semantic similarity between the target scene words. The target scene words whose semantic similarity is greater than a preset threshold can be merged.

[0219] As another practicable manner, based on the above embodiment, optionally, the method may further include:

[0220] Step 2081, perform structural processing on each theme scene to determine the superordinate scene to which each theme scene belongs.

[0221] Step 2082: Recommend products to users based on the superordinate scenarios and thematic scenarios included in each superordinate scenario.

[0222] Specifically, after obtaining a theme scene, each theme scene can be structured to abstract a higher-level scene for the theme scene.

[0223] For example, scene words such as "beautiful legs", "beautiful back", and "slim waist" can abstract the superordinate scene of "shaping", that is, the output results are "shaping_beautiful legs", "shaping_beautiful back", and "shaping_slim waist".

[0224] It may be possible to pre-set structured processing rules and process each subject scenario according to the structured processing rules.

[0225] As another practicable manner, based on the above embodiment, optionally, after obtaining the product pool corresponding to each theme scene, the method further includes:

[0226] Step 2091, for the product pool corresponding to each theme scene, sort the products in the product pool according to preset rules to obtain a sorting result.

[0227] Step 2092: Based on the sorting results, a preset number of products with the highest rankings are selected as products to be recommended.

[0228] Accordingly, products in the product pool corresponding to the target theme scene are obtained as products to be recommended, including:

[0229] Step 1031, obtaining the products to be recommended in the product pool corresponding to the target theme scene.

[0230] Specifically, the product pool obtained according to the above process may include too many products, and it is inconvenient to recommend them all to users. Therefore, they can be sorted according to certain rules, such as sorting according to one or more factors such as sales volume, praise rate, price, etc., to obtain a word-of-mouth ranking list, and a preset number of products with high rankings (i.e., good reputation) can be recommended to users as products to be recommended.

[0231] Optionally, the products in the product pool are sorted according to preset rules to obtain sorting results, which may specifically include:

[0232] Step 20911, obtaining at least one factor of the sales volume, number of independent visitors, number of visits, rate of favorable comments, and unit price of each product in the product pool.

[0233] Step 20912, determine the score of each product based on the preset weight of each factor.

[0234] Step 20913, sort the products according to the scores to obtain the sorting results.

[0235] Specifically, the score of each product can be calculated in a linear weighted manner for each factor, and the products can be ranked according to the score. A high score can be considered to have a good reputation for the product.

[0236] For example, taking the above five factors into consideration, the calculation formula of the product score is as follows:

[0237] score=w1(sales volume)+w2(UV)+w3(PV)+w4(favorable review rate)+w5(unit price)

[0238] Among them, w1-w5 represent the weight of each factor respectively. UV refers to the number of unique visitors. PV refers to the number of visits. It can be understood that the situation for other numbers of factors is similar to this, and will not be repeated here one by one.

[0239] It should be noted that each factor needs to be normalized and the score is calculated based on the normalized value. Taking a product as an example, the specific normalization method is as follows:

[0240] Sales1 = log10(Sale+1) / log10(maxSale+1.0001)

[0241] UV1=log10(UV+1) / log10(maxUV+1.00001)

[0242] PV1=log10(PV+1) / log10(maxPV+1.00001)

[0243] Good review rate 1 = log10(GoodCmtNum+1) / log10(MaxGoodCmtNum+1.00001)

[0244] Unit Price 1 = log10(Price+1) / log10(MaxPrice+1.00001)

[0245] Among them, Sales1 represents the normalized sales value of the product, Sale is the original sales value of the product, and maxSale is the maximum sales value of all products in the product pool where the product is located. Other factors are similar to this and will not be described here one by one. GoodCmtNum represents the original favorable rating of the product, and Price represents the original unit price of the product.

[0246] The score of the product is:

[0247] score=w1(sales volume 1)+w2(UV1)+w3(PV1)+w4(favorable review rate 1)+w5(unit price 1)

[0248] As another practicable manner, based on the above embodiment, optionally, the method may further include:

[0249] The target scene words are filtered according to a preset first rule to obtain popular scene words as the target scene words.

[0250] Specifically, it can be based on the number of hot-selling products associated with each target scene word (X i1 ), the popularity of users searching for each target scene word (X i2 ) and the heat of each target scene word in the user's high-quality comment content (X i3 ) to conduct statistical analysis, perform weighted summation based on these three statistical indicators, obtain the score of each target scene word, sort according to the score, thereby filtering out non-hot scene words, retaining hot scene words, and constructing a hot scene word list as the final target scene word. For hot-selling products, there is a corresponding hot-selling product list. The number of hot-selling products associated with the target scene word is determined by matching the products associated with each target scene word with the hot-selling product list. The popularity of user searches for each target scene word can be determined by matching each target scene word with the user's historical search data. The popularity of each target scene word in the user's high-quality review content is determined by matching the target scene word with the user's high-quality review content.

[0251] For example, taking the popularity of user searches for each target scene word as an example, if there are 100 target scene words, and the total number of times users search for these target scene words is 10,000 times, among which, the number of times each target scene word is searched is Y. i, i = 1, 2, ..., 100, i represents the i-th target scene word. Then the user's search heat X for the i-th target scene word i2 =Y i / 10000*100%.

[0252] For the i-th target scene word, its score Si = a1*X i1 +a2*X i2 +a3*X i3

[0253] As an exemplary implementation, the method specifically includes the following steps:

[0254] 1. Extract product attribute information online and obtain relevant scene words in the product attribute information. Since the scene words obtained are relatively messy, some cleaning rules can be set to clean and filter them to obtain a relatively clean scene word list. Cleaning rules can be used to remove some invalid keywords.

[0255] 2. Perform statistical analysis based on the number of hot-selling products associated with each scene word in the scene word list, the popularity of each scene word in user searches, and the popularity of each scene word in user high-quality comment content. Calculate the score of each scene word by weighted summation based on these three statistical indicators, sort each scene word by score to filter out non-popular scene words, retain popular scene words, and build a popular scene word list.

[0256] 3. Since different scene words in the popular scene vocabulary may express the same meaning, for example, scene words such as "home", "home life", and "daily life at home", they need to be merged based on semantic similarity. Therefore, the word2vector model can be pre-trained to construct the spatial vector of the scene words, and the spatial cosine similarity calculation method can be used to merge semantically similar scene words to obtain a batch of theme scenes and the scene vocabulary corresponding to each theme scene.

[0257] For example, Figure 8 As shown, it is a schematic diagram of the word2vector model application process provided in this embodiment.

[0258] 4. Structural processing is performed on the obtained theme scenes. For example, theme scenes such as "beautiful legs", "beautiful back", and "slim waist" can abstract the large scene of "body shaping", that is, the result output format is "body shaping_beautiful legs", "body shaping_beautiful back", "body shaping_waist shaping", etc.

[0259] 5. Match the scene words in each theme scene with the product name, product attributes, and product review content, build products associated with the scene words, and then build a product pool corresponding to each theme scene. However, since the product pool at this time is built based on matching, and it is difficult to mine the implicit information of the product based on the matching method, the overall coverage of the products in the product pool for hot-selling products is insufficient. For example, the name and attributes of a product do not contain scene words, but its characteristics can be classified under a certain scene word and then classified into a certain theme scene. In order to further improve the coverage of hot-selling products, the GCN graph network algorithm can be used to mine hidden information, classify more products into the correct theme scene, and add them to the product pool corresponding to the corresponding theme scene.

[0260] 6. The GCN graph network can simultaneously learn the node feature information and structural information on the graph end-to-end, extracting features layer by layer, and GCN has a wide range of applicability. Based on the GCN graph network, a scene classification network is constructed and trained to obtain a scene classification model. The specific training process is as follows:

[0261] 1) The user's historical click behavior data and historical order data for a certain period can be extracted. Based on the user's click behavior and the purchase behavior of multiple products in one order, a binary relationship between products is constructed, namely: <Product 1, Product 2>, thereby generating a network with product IDs as vertices and the relationship between two products appearing in the same binary group as edges. Finally, an input of the GCN scene classification network (also called GCN classification network) is generated: the adjacency matrix, which can be called the training adjacency matrix to distinguish it from the above description.

[0262] 2) Extract the products obtained in step 1), extract the user comments of each product, perform Jieba word segmentation on the user comments, use TF-IDF to calculate the keywords related to the product features, and extract the keyword features. Combine the product attribute features of each product to build a product feature set, and use one-hot encoding to encode the product feature set to 0,1 to obtain the product feature code, which can be called training product feature code to distinguish it from the aforementioned application process.

[0263] 3) Based on the training adjacency matrix and training product feature encoding obtained above as the training set, the GCN scene classification network is trained. The classification label can be the theme scene. It can be understood that the classification label can also be the scene label under each theme scene. It can be set according to actual needs as long as the type used in training and subsequent applications is consistent. Taking the classification label as the theme scene as an example, the theme scene needs to be one-hot encoded.

[0264] 4) The output of the scene classification network is the probability that the product belongs to different theme scenes. The final classification result takes the label with the highest probability as the theme scene to which the product belongs. The classification result is compared with the actual label, and the preset loss function is used to monitor the end of the training process. After the training is completed, the GCN scene classification model (i.e., GCN classification model) is obtained.

[0265] 7. Classify more products based on the obtained GCN classification model and add more products to the product pool corresponding to the corresponding theme scene.

[0266] 8. Sort the products in the product pool under different theme scenarios by word-of-mouth ranking, and use the aforementioned linear weighted method to calculate the score of each product to determine the product word-of-mouth ranking list.

[0267] 9. Provide users with scenario-based product search services based on the product reputation ranking list of each product pool.

[0268] It should be noted that the various practicable methods in this embodiment may be implemented separately, or may be combined in any combination without conflict to implement the present invention without limitation.

[0269] The product recommendation processing method provided in this embodiment can effectively expand the products in the product pool corresponding to each theme scene by combining user comment content data and user historical search data to improve product coverage based on user historical click behavior data and historical order data. It can also classify more products by scene through the scene classification model to further improve the coverage of hot-selling products in the product pool. It can also sort the products in the product pool by word of mouth, and recommend a certain number of products with the highest ranking to users as products to be recommended, further improving the accuracy of recommendations and thus improving user experience.

[0270] Still another embodiment of the present invention provides a product recommendation processing device for executing the method of the above embodiment.

[0271] like Fig. 9 , which is a schematic diagram of the structure of the product recommendation processing device provided in this embodiment. The product recommendation processing device 30 includes a receiving module 31, a determining module 32, an acquiring module 33 and a recommending module 34.

[0272] Among them, the receiving module is used to receive a product search request from a first user; the determining module is used to determine, based on the product search request, a target theme scene to which the first scene word included in the product search request belongs; the acquiring module is used to acquire products in a product pool corresponding to the target theme scene as products to be recommended, and the product pool corresponding to each theme scene is obtained based on user historical behavior data; the recommending module is used to recommend the products to be recommended to the first user.

[0273] Regarding the device in this embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0274] According to the product recommendation processing device provided in this embodiment, based on the product search request of the first user, the target theme scene to which the first scene word included in the product search request belongs is determined, the products in the product pool corresponding to the target theme scene are obtained as the products to be recommended, and the products to be recommended are recommended to the user. Since the product pool corresponding to each theme scene is obtained based on the user's historical behavior data, a large number of theme scenes and their corresponding scene words can be effectively obtained, so as to recommend products to users better and more accurately, thereby solving the problem that the prior art has few scene words and cannot accurately recommend products to users.

[0275] Another embodiment of the present invention further supplements the device provided in the above embodiment.

[0276] like Fig.10 , which is a schematic diagram of the structure of a processing device recommended for the product provided in this embodiment.

[0277] As an implementable manner, based on the above embodiment, optionally, the device further includes a processing module 35. The processing module is used to:

[0278] Based on the user review content data, determine the first group of scene words and the products associated with each scene word;

[0279] The first group of scene words are used as target scene words, and the target scene words are merged by semantic similarity to obtain at least one theme scene, and each theme scene corresponds to at least one target scene word;

[0280] According to the target scene words corresponding to each theme scene and the products associated with each target scene word, the product pool corresponding to each theme scene is determined.

[0281] Optionally, the processing module is further configured to:

[0282] Based on the user's historical search data, determine the second group of scene words; remove the duplication of the first group of scene words and the second group of scene words as target scene words;

[0283] The processing module is also used to:

[0284] Based on the user's historical click behavior data and historical order data, determine the target products involved and build a binary relationship between the target products;

[0285] Determine the adjacency matrix of the target products according to the binary relationships between the target products;

[0286] Determine the product features of the target product based on the user reviews of the target product and the attributes of the target product;

[0287] Based on the adjacency matrix and product features of the target products, the trained scenario classification model is used to determine the theme scenarios associated with each target product;

[0288] Processing module, specifically used for:

[0289] According to the scene words corresponding to each theme scene, the products associated with each scene word, and the theme scenes associated with each target product, the product pool corresponding to each theme scene is determined.

[0290] Optionally, the processing module is specifically configured to:

[0291] Perform word segmentation on the user reviews of the target product to obtain word segmentation results;

[0292] According to the word segmentation results, the TF-IDF algorithm is used to determine the keyword features of the target product;

[0293] Determine the attribute characteristics of the target product according to the product attributes of the target product;

[0294] Use keyword features and attribute features as product features of the target product.

[0295] Optionally, the processing module is further configured to:

[0296] One-hot encoding is performed on the product features of the target product to obtain the product feature code corresponding to the target product.

[0297] Optionally, the processing module is specifically configured to:

[0298] Input the adjacency matrix and product feature encoding into the trained GCN classification model to obtain the classification results of each target product, including the probability that each target product belongs to different theme scenarios;

[0299] Based on the classification results of each target product, determine the theme scenarios associated with each target product.

[0300] Optionally, the processing module is specifically configured to:

[0301] Get product attribute information;

[0302] Based on the product attribute information, determine a third group of scene words and products associated with each scene word;

[0303] The first group of scene words, the second group of scene words, and the third group of scene words are removed and used as target scene words.

[0304] Optionally, the processing module is specifically configured to:

[0305] Use the trained word2vector model to construct the spatial vector corresponding to each target scene word;

[0306] Based on the spatial vectors corresponding to each target scene word, the spatial cosine similarity algorithm is used to determine the semantic similarity between each target scene word;

[0307] Semantic similarity merging is performed based on the semantic similarity between the target scene words to obtain at least one theme scene.

[0308] As another practicable manner, based on the above embodiment, optionally, the processing module is further configured to:

[0309] Structural processing is performed on each theme scene to determine the superordinate scene to which each theme scene belongs;

[0310] Recommend products to users based on the superordinate scenarios and thematic scenarios included in each superordinate scenario.

[0311] As another practicable manner, based on the above embodiment, optionally, the processing module is further configured to:

[0312] For the product pool corresponding to each theme scenario, the products in the product pool are sorted according to preset rules to obtain the sorting results;

[0313] According to the sorting results, a preset number of top-ranked products are used as products to be recommended.

[0314] Optionally, the processing module is specifically configured to:

[0315] Obtain at least one factor of the sales volume, number of unique visitors, number of visits, favorable comment rate, and unit price of each product in the product pool;

[0316] Determine the score of each product based on the preset weight of each factor;

[0317] Sort the products by score to obtain the sorting results.

[0318] Regarding the device in this embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0319] It should be noted that the various practicable methods in this embodiment may be implemented separately, or may be combined in any combination without conflict to implement the present invention without limitation.

[0320] According to the product recommendation processing device of this embodiment, by combining user comment content data and user historical search data to mine more theme scenes, and based on user historical click behavior data and historical order data, the products in the product pool corresponding to each theme scene can be effectively expanded to improve product coverage. In addition, through the scene classification model, more products can be classified into scenes to further improve the coverage of hot-selling products in the product pool. It is also possible to sort the products in the product pool by word of mouth, and recommend a certain number of products with the highest ranking to users as products to be recommended, further improving the accuracy of recommendations and thus improving user experience.

[0321] Yet another embodiment of the present invention provides an electronic device for executing the method provided in the above embodiment.

[0322] like Fig.11 FIG. 5 is a schematic diagram of the structure of an electronic device provided in this embodiment. The electronic device 50 includes: at least one processor 51 and a memory 52;

[0323] The memory stores computer-executable instructions; at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method provided in any of the above embodiments.

[0324] According to the electronic device of this embodiment, based on the product search request of the first user, the target theme scene to which the first scene word included in the product search request belongs is determined, the products in the product pool corresponding to the target theme scene are obtained as products to be recommended, and the products to be recommended are recommended to the user. Since the product pool corresponding to each theme scene is obtained based on the user's historical behavior data, a large number of theme scenes and their corresponding scene words can be effectively obtained, so as to recommend products to users better and more accurately, thereby solving the problem of the prior art that there are too few scene words and products cannot be accurately recommended to users.

[0325] Yet another embodiment of the present invention provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the method provided in any of the above embodiments is implemented.

[0326] According to the computer-readable storage medium of this embodiment, based on the product search request of the first user, the target theme scene to which the first scene word included in the product search request belongs is determined, the products in the product pool corresponding to the target theme scene are obtained as the products to be recommended, and the products to be recommended are recommended to the user. Since the product pool corresponding to each theme scene is obtained based on the user's historical behavior data, a large number of theme scenes and their corresponding scene words can be effectively obtained, so as to recommend products to users better and more accurately, thereby solving the problem that the prior art has too few scene words and cannot accurately recommend products to users.

[0327] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0328] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be regarded as exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0329] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A method for processing product recommendations, characterized in that: include: Receiving a product search request from a first user; the first user is any user who has a product recommendation requirement; Determining, according to the product search request, a target theme scene to which the first scene word included in the product search request belongs; Obtain products from a product pool corresponding to the target theme scene as products to be recommended; recommending the product to be recommended to the first user; Among them, the product pool corresponding to each theme scenario is constructed based on the following methods: Determine multiple scene words and products associated with each scene word according to the user's historical behavior data; the user indicated by the user's historical behavior data is the user who has generated the historical behavior data; Merge multiple scene words by semantic similarity to obtain at least one theme scene, each theme scene corresponds to at least one scene word; According to the scene words corresponding to each theme scene and the products associated with each scene word, an initial product pool corresponding to each theme scene is obtained; Determine the target products involved based on the user's historical click behavior data and historical order data; Using the trained scene classification model, determine the theme scenes associated with each of the target products; Each of the target products is added to the theme scene associated with the target product to supplement the initial product pool corresponding to the corresponding theme scene, so as to obtain a product pool corresponding to each theme scene.

2. The method according to claim 1, characterized in that The determining of multiple scene words and products associated with each scene word according to the user's historical behavior data includes: Based on the user review content data, a first group of scene words and products associated with each scene word are determined.

3. The method according to claim 1, characterized in that The determining of multiple scene words and products associated with each scene word according to the user's historical behavior data includes: Based on the user review content data, determine the first group of scene words and the products associated with each scene word; Based on the user's historical search data, determine the second group of scene words; The first group of scene words and the second group of scene words are deduplicated to obtain the plurality of scene words.

4. The method according to any one of claims 1 to 3, characterized in that: The trained scene classification model is used to determine the theme scenes associated with each target product, including: constructing a binary relationship between the target products; Determining an adjacency matrix of the target products according to the binary relationships between the target products; Determine the product features of the target product according to the user review content of the target product and the product attributes of the target product; Based on the adjacency matrix and product features of the target products, a trained scene classification model is used to determine the theme scenes associated with each of the target products.

5. The method according to claim 4, characterized in that Determining the product features of the target product according to the user review content of the target product and the product attributes of the target product includes: Performing word segmentation processing on the user review content of the target product to obtain a word segmentation result; According to the word segmentation results, the TF-IDF algorithm is used to determine the keyword features of the target product; Determining attribute characteristics of the target product according to the product attributes of the target product; The keyword features and the attribute features are used as product features of the target product.

6. The method according to claim 4, characterized in that The method further comprises: One-hot encoding is performed on the product features of the target product to obtain the product feature code corresponding to the target product.

7. The method according to claim 6, characterized in that The method of determining the theme scene associated with each target product by using the trained scene classification model based on the adjacency matrix and product features of the target product includes: Input the adjacency matrix and product feature codes into the trained GCN classification model to obtain classification results of each target product, wherein the classification results include the probability that each target product belongs to a different theme scene; Based on the classification results of each of the target products, a theme scene associated with each of the target products is determined.

8. The method according to claim 1, characterized in that The determining of multiple scene words and products associated with each scene word according to the user's historical behavior data includes: Based on the user review content data, determine the first group of scene words and the products associated with each scene word; Based on the user's historical search data, determine the second group of scene words; Based on the product attribute information, determine a third group of scene words and products associated with each scene word; The plurality of scene words are obtained by removing duplicates from the first group of scene words, the second group of scene words, and the third group of scene words.

9. The method according to claim 1, characterized in that: The step of combining multiple scene words with similar semantics to obtain at least one theme scene includes: Use the trained word2vector model to construct the spatial vector corresponding to each scene word; Based on the spatial vectors corresponding to each scene word, the spatial cosine similarity algorithm is used to determine the semantic similarity between the scene words; Semantic similarity merging is performed based on the semantic similarity between the scene words to obtain at least one theme scene.

10. The method according to claim 1, characterized in that The method further comprises: Structural processing is performed on each theme scene to determine the superordinate scene to which each theme scene belongs; Product recommendations are made to users based on the superordinate scenarios and thematic scenarios included in each superordinate scenario.

11. The method according to any one of claims 1 to 3, 5 to 10, characterized in that: After obtaining the product pool corresponding to each theme scene, the method further includes: For the product pool corresponding to each theme scene, the products in the product pool are sorted according to preset rules to obtain a sorting result; According to the ranking results, a preset number of products ranked high are used as products to be recommended.

12. The method according to claim 11, characterized in that Sorting the products in the product pool according to a preset rule to obtain a sorting result includes: Obtain at least one factor of the sales volume, number of unique visitors, number of visits, favorable comment rate, and unit price of each product in the product pool; Determine the score of each product based on the preset weight of each factor; Sort the products according to the scores to obtain the sorting results.

13. A product recommendation processing device, characterized in that: include: A receiving module, configured to receive a product search request from a first user; the first user is any user who has a product recommendation requirement; A determination module, configured to determine, according to the product search request, a target theme scene to which the first scene word included in the product search request belongs; An acquisition module, used to acquire products in a product pool corresponding to the target theme scene as products to be recommended, wherein the product pool corresponding to each theme scene is obtained based on user historical behavior data; A recommendation module, used for recommending the product to be recommended to the first user; A processing module is used to determine multiple scene words and products associated with each scene word according to the user's historical behavior data; the user indicated by the user's historical behavior data is the user who has generated historical behavior data; multiple scene words are semantically similarly merged to obtain at least one theme scene, each theme scene corresponds to at least one scene word; according to the scene words corresponding to each theme scene and the products associated with each scene word, an initial product pool corresponding to each theme scene is obtained; Based on the user's historical click behavior data and historical order data, the target products involved are determined; the trained scenario classification model is used to determine the theme scenarios associated with each of the target products; each of the target products is added to the theme scenarios associated with the target products to supplement the initial product pool corresponding to the corresponding theme scenarios, thereby obtaining a product pool corresponding to each theme scenario.

14. An electronic device, characterized in that: include: at least one processor and memory; The memory stores computer-executable instructions; The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs the method according to any one of claims 1 to 12.

15. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions, and when the processor executes the computer-executable instructions, the method according to any one of claims 1 to 12 is implemented.

Images