Product recommendation methods, devices, electronic devices and storage media

Abstract

This invention provides a product recommendation method, apparatus, electronic device, and storage medium. The method includes: determining the target recommended product for the target user based on a first feature vector of the target user and a second feature vector of each product in the product library; wherein the first feature vector of the target user and the second feature vector of each product are determined based on the historical purchase information of at least one user, the basic user information of each user, and the basic product information of each product; the first feature vector of the target user represents the target user's purchase tendency, and the second feature vector of the product represents the product's purchase tendency. This invention comprehensively considers the user's historical purchase information and the product's historical purchase information, and integrates the user's basic information and the product's basic information, and can recommend more suitable products to the target user based on the target user's basic information, effectively solving the defect of not being able to accurately recommend suitable products in low-frequency consumption scenarios.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a product recommendation method, apparatus, electronic device, and storage medium. Background Technology

[0002] With the rapid development of the e-commerce industry, users hope to find products that interest them from a large amount of product information and find products that are related to their target purchase.

[0003] Current technologies often employ recall models to filter out hundreds or thousands of candidate products that users might be interested in, and then use ranking models to sort these candidates before finally presenting them to consumers. However, in cold start scenarios, they cannot accurately recommend the products that users currently wish to purchase. Summary of the Invention

[0004] This invention provides a product recommendation method, apparatus, electronic device, and storage medium to address the shortcomings of existing technologies in accurately recommending suitable products in low-frequency consumption scenarios, thereby enabling the recommendation of more suitable products to target users.

[0005] In a first aspect, the present invention provides a product recommendation method, comprising:

[0006] Based on the first feature vector of the target user and the second feature vector of each product in the product library, the target recommended product for the target user is determined.

[0007] The first feature vector of the target user and the second feature vector of each product in the product library are determined based on the historical purchase information of the at least one user, the user basic information of each user, and the product basic information of each product; the first feature vector of the target user is used to represent the target user's purchase tendency, and the second feature vector of the product is used to represent the purchase tendency of the product.

[0008] Optionally, according to the product recommendation method provided by the present invention, the method further includes:

[0009] Perform matrix decomposition on the user purchase matrix to obtain the user purchase feature matrix and the product purchase feature matrix;

[0010] Based on the user purchase feature matrix and the product purchase feature matrix, the user feature tower and product feature tower are obtained by training a dual-tower model.

[0011] Based on the user basic information and the user feature pyramid, a first feature vector is determined for each user;

[0012] Based on the basic information of the goods and the feature pyramid of the goods, a second feature vector is determined for each of the goods;

[0013] The user purchase matrix is ​​determined based on information about at least one user's historical purchases.

[0014] Optionally, according to the product recommendation method provided by the present invention, determining the first feature vector of each user based on the user basic information and the user feature pyramid includes:

[0015] For the first user who has made a historical purchase, the user purchase feature matrix of the first user and the user basic information of the first user are used as feature inputs of the user feature tower to obtain the first feature vector of the first user.

[0016] For a second user who has never made a purchase in the past, the first feature vector of the second user is obtained based on the mean of the first feature vectors of all the first users.

[0017] Optionally, according to the product recommendation method provided by the present invention, obtaining the second feature vector of each product based on basic product information and product feature pyramid includes:

[0018] For the first product that has been purchased in the past, the product purchase feature matrix of the first product and the basic product information of the first product are used as feature inputs to the product feature tower to obtain the second feature vector of the first product.

[0019] For a second product that has never been purchased in the past, the second feature vector of the second product is obtained based on the mean of the second feature vectors of all the first products.

[0020] Optionally, according to the product recommendation method provided by the present invention, determining the target recommended product for the target user includes:

[0021] Based on the locality-sensitive hashing multi-bucket strategy, a second feature vector in the adjacent bucket is determined that is related to the first feature vector of the target user in the adjacent bucket.

[0022] The product corresponding to the second feature vector is determined to be the target recommended product for the target user;

[0023] Wherein, the inner product of the second feature vector of the target recommended product and the first feature vector is less than the first threshold.

[0024] Optionally, according to the product recommendation method provided by the present invention, the method includes:

[0025] Based on the purchase conversion rate model, determine the purchase conversion rate of the target recommended products for the target users;

[0026] Based on the purchase conversion rate of the target user's target recommended products, the target user's target recommended products are ranked.

[0027] The purchase conversion rate of the target recommended product is used to represent the probability that the target recommended product will be purchased after being clicked.

[0028] Optionally, according to the product recommendation method provided by the present invention, the method further includes:

[0029] The purchase conversion rate model is trained based on the target user's historical clicked products and the target user's historical purchased products.

[0030] The information on the target user's historical clicks on products includes:

[0031] Products that the target user has historically clicked;

[0032] Products searched in the target user's historical search history;

[0033] The time of the target user's most recent click on a product;

[0034] The target user's historical purchase information includes:

[0035] Products purchased in the target user's history;

[0036] The time of the target user's most recent purchase;

[0037] The price of goods purchased by the target user in the past and the target user's balance.

[0038] Secondly, the present invention also provides a product recommendation device, comprising:

[0039] The determination module is used to determine the target recommended products for the target user based on the first feature vector of the target user and the second feature vector of each product in the product library.

[0040] The first feature vector of the target user and the second feature vector of each product in the product library are determined based on the historical purchase information of the at least one user, the user basic information of each user, and the product basic information of each product; the first feature vector of the target user is used to represent the target user's purchase tendency, and the second feature vector of the product is used to represent the purchase tendency of the product.

[0041] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the product recommendation method provided in the first aspect.

[0042] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the product recommendation method as provided in the first aspect.

[0043] The product recommendation method, apparatus, electronic device, and storage medium provided by this invention, by comprehensively considering the user's historical purchase information and the product's historical purchase information, and by integrating the user's basic information and the product's basic information, can accurately recommend more suitable product categories to the target user based on the target user's basic information, effectively solving the defect of not being able to accurately recommend suitable products to users in low-frequency consumption scenarios. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating the product recommendation method provided in an embodiment of this application;

[0046] Figure 2 This is a schematic diagram of the improved dual-tower model recall strategy provided by the present invention;

[0047] Figure 3 This is a schematic diagram of the Wide-ESSM network architecture provided by the present invention;

[0048] Figure 4 This is a schematic diagram of the product recommendation architecture provided by the present invention;

[0049] Figure 5 This is a schematic diagram of the product recommendation device provided by the present invention;

[0050] Figure 6 A schematic diagram of the physical structure of an electronic device is provided. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0052] The following is combined with Figures 1-5The present invention describes a product recommendation method and apparatus.

[0053] Figure 1 This is a flowchart illustrating the product recommendation method provided in the embodiments of this application, such as... Figure 1 As shown, the method includes the following steps:

[0054] Step 100: Based on the first feature vector of the target user and the second feature vector of each product in the product library, determine the target recommended product for the target user;

[0055] The first feature vector of the target user and the second feature vector of each product in the product library are determined based on the historical purchase information of the at least one user, the user basic information of each user, and the product basic information of each product; the first feature vector of the target user is used to represent the target user's purchase tendency, and the second feature vector of the product is used to represent the purchase tendency of the product.

[0056] The rapid development of the e-commerce industry has driven the emergence of a new consumption scenario: users hope to discover products of interest from a vast amount of product information, and find products related to their target purchase. Currently, mature recommendation models generally include recall models and ranking models. Recall models quickly filter out hundreds or thousands of candidate products that users may be interested in from tens of thousands of products, while ranking models are often more refined and iterate faster. By combining more multi-dimensional features and contextual information, they rank the products in the candidate set and ultimately present them to the consumer.

[0057] If the basic attributes of users or products are not considered when recommending products to users, the recommendation behavior will not perform well in cold start scenarios (such as when users have no history of behavior, or when products are newly listed and no users have purchased them before).

[0058] In multi-path recall methods, each recall strategy is relatively independent, and it is impossible to determine which recall strategy should have a greater impact. Although the graph embedding-based recall method solves the cold start problem for products well, it is still powerless for the cold start scenario for users.

[0059] Optionally, the effectiveness of incorporating basic user information into the ranking model heavily relies on the candidate products obtained from the recall model. If a user's historical behavior is insufficient, re-ranking from inaccurate recall candidate products is meaningless. Therefore, basic user and product information can be incorporated into the recall strategy.

[0060] Optionally, this invention proposes incorporating basic user and product information into the recommended recall strategy. By combining the user's historical purchase information and basic user information, the user's embedding feature (i.e., the first feature vector) is obtained. Similarly, by combining the product's basic information and its historical purchase information, the product's embedding feature (i.e., the second feature vector) is obtained. Therefore, when recommending products to a target user, the target recommended products can be determined based on the target user's first feature vector and each product's second feature vector. This effectively solves the recommendation problem in low-frequency consumption scenarios and also effectively helps with user and product cold start issues.

[0061] Optionally, the user's basic information may include, but is not limited to, the user's age, gender, geographical location, the types of apps installed on the user's phone, and the user's remaining points.

[0062] Optionally, the basic product information may include, but is not limited to, the product category, such as daily necessities, baby products, school supplies, books, hardware, food, etc. The basic product information may also include the product brand, price, etc.

[0063] Taking an points mall as an example, suppose a user has only purchased toilet paper once in the points mall. However, based solely on this one purchase record, recommending household goods every time the user browses the points mall's product library might not be the category the user intended to buy at that time, and it might also make the user feel that the product selection in the points mall is monotonous. Therefore, it is possible to recommend household goods and baby products based on the user's basic information. For example, if the user is female and has a maternity and baby app installed on their phone, then the recommendations could be based on these categories. If the user is male and has a fitness app installed on their phone, then the recommendations could be based on fitness equipment. The correspondence between the user's basic information and the product's basic information can be pre-set or determined based on big data; this invention does not impose any limitations on this.

[0064] To address the cold start problem for users and products in low-frequency consumption scenarios of points malls, this invention proposes incorporating basic user and product information into the recommendation recall strategy. By combining user historical behavior and basic user information, a user embedding (i.e., the first feature vector) is obtained. Similarly, by combining product ID, basic product information, and product popularity, a product embedding (i.e., the second feature vector) is obtained. This effectively solves the recommendation challenge in low-frequency consumption scenarios and also significantly helps address the cold start problems for both users and products.

[0065] The product recommendation method, apparatus, electronic device, and storage medium provided by this invention, by comprehensively considering the user's historical purchase information and the product's historical purchase information, and by integrating the user's basic information and the product's basic information, can accurately recommend more suitable product categories to the target user based on the target user's basic information, effectively solving the defect of not being able to accurately recommend suitable products to users in low-frequency consumption scenarios.

[0066] Optionally, the method further includes:

[0067] Perform matrix decomposition on the user purchase matrix to obtain the user purchase feature matrix and the product purchase feature matrix;

[0068] Based on the user purchase feature matrix and the product purchase feature matrix, the user feature tower and product feature tower are obtained by training a dual-tower model.

[0069] Based on the user basic information and the user feature pyramid, a first feature vector is determined for each user;

[0070] Based on the basic information of the goods and the feature pyramid of the goods, a second feature vector is determined for each of the goods;

[0071] The user purchase matrix is ​​determined based on information about at least one user's historical purchases.

[0072] Optionally, a first feature vector for each user and a second feature vector for each product can be obtained firstly based on the historical purchase information of the at least one user, the user basic information of each user, and the product basic information of each product.

[0073] Optionally, when determining the first feature vector for each user and the second feature vector for each product, a user purchase matrix (i.e., a user-product preference matrix U1) can be constructed first. The values ​​in matrix U1 can reflect the number of times a user has purchased a product. For example, each row of the matrix corresponds to a user ID, and each column corresponds to a product ID. The value of each element can indicate whether the user in the row corresponding to that element has purchased the product in the column corresponding to that element, or the number of times they have purchased that product. The dimension of the matrix is ​​u×i, where u is the number of users and i is the number of product IDs.

[0074] Optionally, the U1 matrix can be decomposed using ALS to obtain a u×k dimensional matrix representing the user, i.e., the user purchase feature matrix, and an i×k dimensional matrix representing the product, i.e., the product purchase feature matrix, where k is the latent vector dimension.

[0075] Optionally, for tens of thousands of candidate products, a recall process can be performed first to filter out several hundred personalized products for recommendation. Since the recall network structure based on deep learning is complex, directly using one-hot encoding as input to train the embeddings of users and products is very slow. Therefore, this invention proposes to pre-train the embeddings of products and users using ALS matrix factorization.

[0076] This invention uses ALS to pre-train the dual-tower model, shortening the training time while retaining the model's end-to-end training capability. Simultaneously, this patent models infrequently changing user information within the recall strategy, while the ranking model focuses on scenario and contextual information, effectively reducing the training iteration difficulty of the ranking model.

[0077] Optionally, the dual-tower model uses a complex deep learning network to embed user-related features and advertising-related features separately, forming two independent towers. In the final cross layer, the user features and advertising features are multiplied. If the user has purchased the product before, the output label is 1; otherwise, it is 0.

[0078] Optionally, the u×k dimensional matrix representing users, i.e., the user purchase feature matrix, can be used as the initial parameter for user embedding in the dual-tower model;

[0079] Optionally, the i×k dimensional matrix representing the goods, i.e., the product purchase feature matrix, can be used as the initial parameter for the product embedding in the dual-tower model.

[0080] Optionally, user feature towers and product feature towers can be obtained by training a dual-tower model based on the user purchase feature matrix and the product purchase feature matrix.

[0081] Optionally, a dual-tower model is used as the recall network, where user purchase features and product purchase features are processed separately using embedding, forming two independent towers: the user feature tower and the product feature tower. The user features and advertising features are multiplied in the final cross layer. During the training process, the loss can be calculated to complete the network training. After training, the vectors generated by the intermediate layers of the network can be used as embedding features, resulting in user embedding features (i.e., the first feature vector) and product embedding features (i.e., the second feature vector) that integrate multi-dimensional information.

[0082] Optionally, the present invention can employ a dual-tower model as the recall network, performing embedding processing on user purchase characteristics and product purchase characteristics separately to form two independent towers, namely the user feature tower and the product feature tower. Then, based on the user's basic information and the user feature tower, a first feature vector (which can be called the user embedding feature) can be determined for each user, and based on the product's basic information and the product feature tower, a second feature vector (which can be called the product embedding feature) can be determined for each product.

[0083] This invention incorporates user basic information and product basic information into the recommended recall strategy. Through a dual-tower model, the user feature tower integrates the user's historical purchase behavior and user basic information to obtain the user embedding features of each user, and the product feature tower integrates the product ID, product basic information, and product popularity to obtain the product embedding features.

[0084] Optionally, since users have limited historical purchasing behavior, the latent vectors obtained by ALS cannot fully express users' purchasing biases. Optimization through end-to-end training with deep learning is more reasonable. Therefore, this invention uses ALS matrix factorization to pre-train the embeddings of products and users, obtaining initial parameters for user embeddings and product embeddings in the dual-tower model, which more easily expresses users' purchasing biases.

[0085] To address the cold start problem for users and products in low-frequency consumption scenarios of points malls, this invention proposes incorporating basic user and product information into the recommendation recall strategy. Through a dual-tower model, the user tower integrates historical user behavior and basic user information to obtain the user's embedding, while the product tower integrates the product's ID, basic product information, and product popularity to obtain the product's embedding. This effectively solves the recommendation challenge in low-frequency consumption scenarios and significantly helps address both user and product cold start issues.

[0086] Optionally, determining the first feature vector for each user based on the user basic information and the user feature pyramid includes:

[0087] For the first user who has made a historical purchase, the user purchase feature matrix of the first user and the user basic information of the first user are used as feature inputs of the user feature tower to obtain the first feature vector of the first user.

[0088] For a second user who has never made a purchase in the past, the first feature vector of the second user is obtained based on the mean of the first feature vectors of all the first users.

[0089] Optionally, the first user can be a user with a history of purchasing behavior;

[0090] Optionally, the second user can be a user with no history of purchasing behavior;

[0091] Optionally, Figure 2 This is a schematic diagram of the improved dual-tower model recall strategy provided by the present invention, as shown below. Figure 2 As shown, the present invention uses the user purchase feature matrix and user basic information obtained from ALS, representing the user's historical behavior, such as the user's age, gender, geographical location, types of apps installed on the phone, and the user's remaining points, as the feature input of the user feature tower in the dual-tower model to obtain the first feature vector of the first user.

[0092] Optionally, when recalling a second user who has no historical behavior, the average of the first feature vectors of all first users can be used as the first feature vector of the second user.

[0093] Optionally, obtaining the second feature vector for each product based on its basic information and feature pyramid includes:

[0094] For the first product that has been purchased in the past, the product purchase feature matrix of the first product and the basic product information of the first product are used as feature inputs to the product feature tower to obtain the second feature vector of the first product.

[0095] For a second product that has never been purchased in the past, the second feature vector of the second product is obtained based on the mean of the second feature vectors of all the first products.

[0096] Optionally, the first item may be an item that has been purchased;

[0097] Optionally, the second product can be a product that has not been purchased before, such as a product that has just been listed.

[0098] Optionally, the product purchase feature matrix representing the product ID obtained from ALS, and basic product information such as product brand, price, and product category can be used as feature inputs to the product feature tower in the dual-tower model to obtain the second feature vector of the first product.

[0099] Optionally, when recalling a second product, the average of the second feature vectors of all first products can be used as the second feature vector of that first product.

[0100] Optionally, the present invention employs a dual-tower model, which can obtain user embedding features and product embedding features that integrate multi-dimensional information.

[0101] Optionally, determining the target recommended product for the target user includes:

[0102] Based on the locality-sensitive hashing multi-bucket strategy, a second feature vector in the adjacent bucket is determined that is related to the first feature vector of the target user in the adjacent bucket.

[0103] The product corresponding to the second feature vector is determined to be the target recommended product for the target user;

[0104] Wherein, the inner product of the second feature vector of the target recommended product and the first feature vector is less than the first threshold.

[0105] Traditional methods for selecting candidate items in a recall model involve calculating the inner product between the item and the user embedding, and then selecting all items whose product value is less than a certain threshold. This means that all items in the candidate set need to be traversed. In a k-dimensional embedding space, if the total number of items is n, the time complexity of traversing and calculating the similarity between user and item vectors is O(kn). This complexity is unacceptable in recommendation systems with tens of thousands of items.

[0106] Optionally, the present invention can use a locality-sensitive hashing multi-bucket strategy to obtain the products that are near the user's feature vector, i.e., the target recommended products.

[0107] Optionally, after obtaining the user embedding features and the product embedding features, i.e., the first feature vector of each user and the second feature vector of each product, when determining the target recommended product for the target user, in order to achieve online recall, a locality-sensitive hashing multi-bucket strategy can be used. This strategy involves placing adjacent points into a "bucket". In this way, when performing a neighbor search, it is only necessary to search within a few adjacent buckets. That is, the second feature vectors that are close to the high-order spatial distance of the first feature vector of the target user can be searched in adjacent buckets, thereby determining that the products corresponding to these second feature vectors are the target recommended products of the target user.

[0108] Optionally, when determining a second feature vector that is close to the high-order spatial distance of the first feature vector of the target user, the second feature vector whose inner product with the first feature vector is less than a first threshold can be selected; wherein, the first threshold can be a pre-set fixed value.

[0109] Optionally, this invention can utilize the property that low-dimensional space can preserve the proximity relationships in high-dimensional space. The locality-sensitive hashing multi-bucket strategy maps the high-dimensional embedding vector to a one-dimensional space by multiplying it with a randomly generated k-dimensional mapping vector, thus transforming it into a numerical value.

[0110] h(v) = v·x;

[0111] Bucketing can be performed using a hash function h(v).

[0112]

[0113] Where w is the width of the bucket, and b is a uniformly distributed random variable between 0 and w.

[0114] Optionally, to avoid misclassification of nearby points, multiple hash functions can be used for bucketing simultaneously. After finding a candidate set of adjacent points through bucketing, the K nearest neighbors of the target point can be found by traversing the finite candidate set.

[0115] Optionally, the present invention obtains user embedding features and product embedding features based on the recall model, and uses the local sensitive hash multi-bucket strategy to obtain the products near the user feature vector, which is the recall candidate product set, that is, the target recommended products of the target user.

[0116] Optionally, the method includes:

[0117] Based on the purchase conversion rate model, determine the purchase conversion rate of the target recommended products for the target users;

[0118] Based on the purchase conversion rate of the target user's target recommended products, the target user's target recommended products are ranked.

[0119] The purchase conversion rate of the target recommended product is used to represent the probability that the target recommended product will be purchased after being clicked.

[0120] Regarding the network architecture of ranking models, each algorithm has its own advantages and disadvantages, and it is worth considering which one is more suitable for the application scenario of points mall. Among them, the LR model is simple and has a short prediction time, but it requires feature engineering; the FM model can automatically complete feature crossing and learn the correlation between features, effectively avoiding the Simpson problem in statistics, but it can only learn second-order cross features; deep learning models can theoretically fit any form of high-order function and have strong generalization ability, but often have reduced memory ability; Wide&Deep inputs the features that need to be generalized into the deep learning network, and concatenates the features that have a direct impact on the recommendation results with the output features of deep learning before inputting them into the logistic regression layer, which preserves the model's generalization ability while improving the network's ability to remember important features.

[0121] A single model can only optimize one objective function. In the field of e-commerce recommendation, in order to simultaneously optimize user click-through rate and purchase rate, and to use the former to assist the latter in solving the training problem caused by insufficient sample size, ESSM can be used to solve this problem. However, ESSM treats all features equally, which is not entirely applicable to the points mall scenario in this invention.

[0122] In the points mall scenario, the large user base, low purchase frequency, and infrequent product updates mean that solutions for user-based cold start are just as important as those for product-based cold start. Many users have only one or two past purchases, making recall strategies based solely on user history ineffective. Users returning to the points mall six months later are far less likely to still prefer similar products from their past purchases compared to other leading domestic e-commerce apps.

[0123] Meanwhile, a significant difference between points-based marketplaces and other e-commerce platforms lies in their points system. The products users tend to purchase heavily depend on their points balance and the points value of the items. However, simply recommending items priced lower than a user's remaining points is not a good solution. Enhancing the user's browsing experience and engagement is also crucial. Even high-priced items that users are highly interested in should be highlighted, providing more opportunities to expose other products or advertisements to the user, rather than letting them leave after making a purchase. Therefore, optimizing click-through rates and conversion rates are equally important.

[0124] This invention utilizes a trained Wide-ESSM model (purchase conversion rate model) to predict the purchase conversion rate of users for the recall candidate products in sequence, sorts them by purchase conversion rate score, and finally pushes the sorted products to users in order.

[0125] Optionally, the present invention utilizes the ESSM model, which is divided into a CTR (click-through rate) optimization model and a CVR (click-to-purchase conversion rate) optimization model. The final purchase conversion rate can be obtained by multiplying the outputs of the two networks.

[0126] This invention proposes a Wide-ESSM ranking strategy, which draws on Alibaba's multi-objective optimization model ESSM. It also uses click and conversion data for global optimization. Furthermore, it incorporates two features that directly affect the results—the user's remaining points and the product price—by performing feature cross-feeding and inputting them into the output layers of CVR and CTR, respectively. This approach preserves the model's generalization ability while enhancing its memory capacity.

[0127] This invention utilizes a trained Wide-ESSM model to predict the purchase conversion rate of users for the recall candidate products in sequence, and sorts them according to the purchase conversion rate score. Finally, the sorted products are pushed to users in order, so that the products that users are more likely to buy are placed first, and the products that are less likely to be purchased are placed later, making it easier for users to see suitable products, effectively improving the user's purchase probability and user experience.

[0128] Optionally, the method further includes:

[0129] The purchase conversion rate model is trained based on the target user's historical clicked products and the target user's historical purchased products.

[0130] The information on the target user's historical clicks on products includes:

[0131] Products that the target user has historically clicked;

[0132] Products searched in the target user's historical search history;

[0133] The time of the target user's most recent click on a product;

[0134] The target user's historical purchase information includes:

[0135] Products purchased in the target user's history;

[0136] The time of the target user's most recent purchase;

[0137] The price of goods purchased by the target user in the past and the target user's balance.

[0138] Optionally, Figure 3 This is a schematic diagram of the Wide-ESSM network architecture provided by the present invention, as shown below. Figure 3 As shown, this invention can add information about the target user's historical clicked products and historical purchased products to the CTR and CVR networks, and can realize the display of the user's clicked, purchased, the time since the last clicked product, the time since the last purchased product, the search category name, and the number of times the same subcategory of products has been exposed in the CTR and CVR networks.

[0139] Optionally, the present invention includes the number of times that products in the same subcategory have been exposed, which can prevent the product categories pushed to users from being too limited.

[0140] Optionally, since the user's remaining points and the price of the goods have a more direct impact on the click-through rate and purchase rate, users will almost never consider goods with a value greater than their remaining points when purchasing goods. This invention can consider the price of goods purchased by the target user in the past and the balance of the target user in the CTR and CVR networks, and can effectively filter out goods that users are unlikely to purchase because the price is much higher than the user's balance.

[0141] This method directly inputs the product price and the user's remaining points into the logistic regression layer of the CTR and CVR models. Embeddings for all features in the CVR and CTR models are shared.

[0142] This invention improves upon the Wide-Deep network architecture by using a pre-trained Wide-ESSM model to predict the user's purchase conversion rate for the recall candidate products, sorting them by purchase conversion rate score, and finally pushing the sorted products to the user in order.

[0143] Figure 4 This is a schematic diagram of the product recommendation architecture provided by the present invention, such as... Figure 4 As shown, its specific implementation method is as follows:

[0144] (1) The present invention first pre-trains the product and user embeddings by ALS matrix decomposition to obtain the initial parameters of user embeddings and product embeddings in the dual-tower model.

[0145] (2) A dual-tower model is used as the recall network. The user’s historical purchase features and the product’s historical purchase features are embedded to form two independent towers. In the final cross layer, the user features and the advertising features are multiplied to obtain the user embedding features (first feature vector) and the product embedding features (second feature vector) that integrate multi-dimensional information.

[0146] (3) Based on the recall model, user embedding features and product embedding features are obtained. The products near the user feature vector are obtained by using the local sensitive hash multi-bucket strategy, which is the recall candidate product set.

[0147] (4) Using the trained Wide-ESSM model, predict the purchase conversion rate of users for the recall candidate products in turn, sort them by purchase conversion rate score, and finally push the sorted products to users in order.

[0148] This invention incorporates basic user and product information into the recommended recall strategy. A dual-tower model is used to obtain user embeddings by integrating historical user behavior and basic user information in the user tower, and product embeddings by integrating product IDs, basic product information, and product popularity in the product tower. The dual-tower model is pre-trained using ALS. Based on the dual-tower recall model, user embedding features and product embedding features are obtained. A Locality Sensitive Hashing (LSH) multi-bucket strategy is used to find products adjacent to the user's feature vector. A pre-trained Wide-ESSM model is then used to predict the user's purchase conversion rate for the recall candidate products, and the products are ranked according to their purchase conversion rate scores. Finally, the ranked products are pushed to users in sequence.

[0149] This invention proposes a recommendation algorithm for low-frequency consumption scenarios that considers the points mechanism of a points mall. The method includes a recall strategy and a ranking strategy. Recall is performed using a dual-tower model with basic user and product attributes, and ranking is performed using a Wide-ESSM network structure improved based on the ESSM network, thereby solving the cold start problem of users and products in low-frequency consumption scenarios of points malls.

[0150] The product recommendation method, apparatus, electronic device, and storage medium provided by this invention, by comprehensively considering the user's historical purchase information and the product's historical purchase information, and by integrating the user's basic information and the product's basic information, can accurately recommend more suitable product categories to the target user based on the target user's basic information, effectively solving the defect of not being able to accurately recommend suitable products to users in low-frequency consumption scenarios.

[0151] The product recommendation device provided by the present invention will be described below. The product recommendation device described below can be referred to in correspondence with the product recommendation method described above.

[0152] Figure 5 This is a schematic diagram of the product recommendation device provided by the present invention, as shown below. Figure 5 As shown, the device includes: a determining module 510; wherein:

[0153] The determination module 510 is used to determine the target recommended product for the target user based on the first feature vector of the target user and the second feature vector of each product in the product library;

[0154] The first feature vector of the target user and the second feature vector of each product in the product library are determined based on the historical purchase information of the at least one user, the user basic information of each user, and the product basic information of each product; the first feature vector of the target user is used to represent the target user's purchase tendency, and the second feature vector of the product is used to represent the purchase tendency of the product.

[0155] Optionally, the product recommendation device can determine the target recommended product for the target user based on the first feature vector of the target user and the second feature vector of each product in the product library through the determination module 510.

[0156] The product recommendation device provided by this invention, by comprehensively considering the user's historical purchase information and the product's historical purchase information, and integrating the user's basic information and the product's basic information, can accurately recommend more suitable product categories to the target user based on the target user's basic information, effectively solving the defect of not being able to accurately recommend suitable products to users in low-frequency consumption scenarios.

[0157] Figure 6An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 6 As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a product recommendation method, which includes:

[0158] Based on the first feature vector of the target user and the second feature vector of each product in the product library, the target recommended product for the target user is determined.

[0159] The first feature vector of the target user and the second feature vector of each product in the product library are determined based on the historical purchase information of the at least one user, the user basic information of each user, and the product basic information of each product; the first feature vector of the target user is used to represent the target user's purchase tendency, and the second feature vector of the product is used to represent the purchase tendency of the product.

[0160] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0161] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer is able to execute the product recommendation method provided by the above methods, the method comprising:

[0162] Based on the first feature vector of the target user and the second feature vector of each product in the product library, the target recommended product for the target user is determined.

[0163] The first feature vector of the target user and the second feature vector of each product in the product library are determined based on the historical purchase information of the at least one user, the user basic information of each user, and the product basic information of each product; the first feature vector of the target user is used to represent the target user's purchase tendency, and the second feature vector of the product is used to represent the purchase tendency of the product.

[0164] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the aforementioned product recommendation methods, the method comprising:

[0165] Based on the first feature vector of the target user and the second feature vector of each product in the product library, the target recommended product for the target user is determined.

[0166] The first feature vector of the target user and the second feature vector of each product in the product library are determined based on the historical purchase information of the at least one user, the user basic information of each user, and the product basic information of each product; the first feature vector of the target user is used to represent the target user's purchase tendency, and the second feature vector of the product is used to represent the purchase tendency of the product.

[0167] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0168] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A product recommendation method, characterized in that, include: Based on the first feature vector of the target user and the second feature vector of each product in the product library, the target recommended product for the target user is determined. The first feature vector of the target user and the second feature vector of each item in the product library are determined based on the historical purchase information of at least one user, the user basic information of each user, and the product basic information of each item; the first feature vector of the target user is used to represent the target user's purchase tendency, and the second feature vector of the item is used to represent the purchase tendency of the item. A purchase conversion rate model is trained based on information about products clicked and purchased by the target user in the past. The information about products purchased by the target user in the past includes the price of the products when the target user purchased them and the target user's balance. The purchase conversion rate model is divided into a CTR (Click-Through Rate) optimization model and a CVR (Conversion Rate) optimization model. The purchase conversion rate is obtained by multiplying the output results of the CTR optimization model and the CVR optimization model. When training the purchase conversion rate model, the target user's balance and the product price are cross-referenced and then input into the logistic regression layer of the CTR optimization model and the CVR optimization model, respectively. Based on the purchase conversion rate model, determine the purchase conversion rate of the target recommended product for the target user; Based on the purchase conversion rate of the target user's target recommended products, the target user's target recommended products are ranked. The purchase conversion rate of the target recommended product is used to represent the probability that the target recommended product will be purchased after being clicked.

2. The product recommendation method according to claim 1, characterized in that, The method further includes: Perform matrix decomposition on the user purchase matrix to obtain the user purchase feature matrix and the product purchase feature matrix; Based on the user purchase feature matrix and the product purchase feature matrix, the user feature tower and product feature tower are obtained by training a dual-tower model. Based on the user basic information and the user feature pyramid, a first feature vector is determined for each user; Based on the basic information of the goods and the feature pyramid of the goods, a second feature vector is determined for each of the goods; The user purchase matrix is ​​determined based on information about at least one user's historical purchases.

3. The product recommendation method according to claim 2, characterized in that, The step of determining the first feature vector for each user based on the user basic information and the user feature pyramid includes: For the first user who has made a historical purchase, the user purchase feature matrix of the first user and the user basic information of the first user are used as feature inputs of the user feature tower to obtain the first feature vector of the first user. For a second user who has never made a purchase in the past, the first feature vector of the second user is obtained based on the mean of the first feature vectors of all the first users.

4. The product recommendation method according to claim 2, characterized in that, The process of obtaining a second feature vector for each product based on its basic information and feature pyramid includes: For the first product that has been purchased in the past, the product purchase feature matrix of the first product and the basic product information of the first product are used as feature inputs to the product feature tower to obtain the second feature vector of the first product. For a second product that has never been purchased in the past, the second feature vector of the second product is obtained based on the mean of the second feature vectors of all the first products.

5. The product recommendation method according to any one of claims 1-4, characterized in that, The process of determining the target recommended products for the target user includes: Based on the locality-sensitive hashing multi-bucket strategy, a second feature vector in the adjacent bucket is determined that is related to the first feature vector of the target user in the adjacent bucket. The product corresponding to the second feature vector is determined to be the target recommended product for the target user; Wherein, the inner product of the second feature vector of the target recommended product and the first feature vector is less than the first threshold.

6. The product recommendation method according to claim 1, characterized in that, The method further includes: The information on the target user's historical clicks on products includes: Products that the target user has historically clicked; Products searched in the target user's historical search history; The time of the target user's most recent click on a product; The target user's historical purchase information also includes: Products purchased in the target user's history; The time when the target user last purchased a product.

7. A product recommendation device, characterized in that, The device includes: The determination module is used to determine the target recommended products for the target user based on the first feature vector of the target user and the second feature vector of each product in the product library. The first feature vector of the target user and the second feature vector of each item in the product library are determined based on the historical purchase information of at least one user, the user basic information of each user, and the product basic information of each item; the first feature vector of the target user is used to represent the target user's purchase tendency, and the second feature vector of the item is used to represent the purchase tendency of the item. The determining module is further configured to train a purchase conversion rate model based on information about products clicked in the target user's history and information about products purchased in the target user's history, wherein the information about products purchased in the target user's history includes the price of the products when the target user purchased them in the past and the target user's balance. The purchase conversion rate model is divided into a CTR (Click-Through Rate) optimization model and a CVR (Conversion Rate) optimization model. The purchase conversion rate is obtained by multiplying the output results of the CTR optimization model and the CVR optimization model. When training the purchase conversion rate model, the target user's balance and the product price are cross-referenced and then input into the logistic regression layer of the CTR optimization model and the CVR optimization model, respectively. Based on the purchase conversion rate model, determine the purchase conversion rate of the target recommended product for the target user; Based on the purchase conversion rate of the target user's target recommended products, the target user's target recommended products are ranked. The purchase conversion rate of the target recommended product is used to represent the probability that the target recommended product will be purchased after being clicked.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the product recommendation method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the product recommendation method as described in any one of claims 1 to 6.

Images