Product recommendation method, product recommendation model training method, device and equipment

Abstract

This disclosure provides a product recommendation method related to the field of artificial intelligence. The method comprises: obtaining at least one product metric for each of N first users; obtaining an evaluation score for each first user based on the at least one product metric for each first user; classifying M third users based on the first user features of K second users to obtain S fourth users; and inputting the second user features of each of the K second users and the S fourth users into a pre-trained product recommendation model to obtain recommended products for each user. This disclosure also provides a product recommendation model training method, apparatus, device, storage medium, and program product.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence, and more specifically to a product recommendation method, a training method, an apparatus, a device, a medium, and a program product for a product recommendation model. Background Art

[0002] With the advent of the big data era, product data (such as financial products) is increasing exponentially. Finding an effective product recommendation method to meet users' personalized needs has become a hot topic of current research. Taking the banking industry as an example, its user base can reach hundreds of millions. Lack of technological expertise and insufficient computing power will hinder the use of deep learning. Model training or prediction based on the entire user base results in a large amount of ineffective computation, and recommendations for non-valuable users also incur unnecessary marketing costs. Therefore, how to improve product recommendation efficiency while reducing ineffective computation during AI model training or prediction while meeting users' personalized needs is a pressing issue. Summary of the Invention

[0003] In view of the above problems, the present disclosure provides a product recommendation method, a training method, an apparatus, a device, a medium and a program product for a product recommendation model that reduce invalid operations and improve product recommendation efficiency.

[0004] One aspect of an embodiment of the present disclosure provides a product recommendation method, comprising: obtaining at least one product indicator for each of N first users, wherein the N first users include users who purchase products within a predetermined time period, and the at least one product indicator is obtained based on information about product purchases within the predetermined time period; obtaining an evaluation score for each first user based on the at least one product indicator for each first user; classifying M third users based on the first user features of K second users to obtain S fourth users, wherein the K second users include first users among the N first users whose evaluation scores are greater than or equal to a first preset value, the S fourth users are of the same category as the K second users, and the M third users include users who have not purchased products within the predetermined time period, N, K, S and M are respectively integers greater than or equal to 1, N is greater than or equal to K, and M is greater than or equal to S; inputting the second user features of each of the K second users and the S fourth users into a pre-trained product recommendation model to obtain recommended products for each user.

[0005] According to an embodiment of the present disclosure, obtaining the evaluation score of each first user based on at least one product indicator of each first user includes: obtaining the category to which each first user belongs in each product indicator, wherein each product indicator is divided into at least one category; and obtaining the evaluation score of each first user based on the score and weight corresponding to the category of each product indicator.

[0006] According to an embodiment of the present disclosure, the at least one product indicator includes at least one of a consumption time indicator, a consumption frequency indicator and a consumption amount indicator: the consumption time indicator includes the time of the most recent purchase of the product; the consumption frequency indicator includes the number of times consumption is made in a specific sub-time period within the predetermined time period; the consumption amount indicator includes the amount of money spent on purchasing the product in the specific sub-time period.

[0007] According to an embodiment of the present disclosure, inputting the second user features of each of the K second users and the S fourth users into a pre-trained product recommendation model includes: performing feature screening on the second user features of each user based on an importance analysis result, wherein the importance analysis result includes the importance of at least one sub-feature in the second user feature, and the importance analysis result is obtained by an importance analysis pre-performed by a decision tree model; and inputting the screened second user features of each user into the pre-trained product recommendation model.

[0008] According to an embodiment of the present disclosure, classifying the M third users based on the first user features of the K second users includes: inputting the first user features of each of the K second users and the M third users into a support vector machine model, and obtaining a classification result output by the support vector machine model.

[0009] Another aspect of an embodiment of the present disclosure provides a method for training a product recommendation model, comprising: obtaining at least one product indicator of each first training user among n first training users, wherein the n first training users include training users who purchase products within a predetermined time period, and the at least one product indicator is obtained based on information about product purchases within the predetermined time period; obtaining an evaluation score of each first training user based on the at least one product indicator of each first training user; classifying m third training users based on the first user features of k second training users to obtain s fourth training users, wherein the k second training users include first training users among the n first training users whose evaluation scores are greater than or equal to a first preset value, the s fourth training users are of the same category as the k second training users, and the m third training users include training users who have not purchased products within the predetermined time period, and n, k, s and m are respectively integers greater than or equal to 1, n is greater than or equal to k, and m is greater than or equal to s; and inputting the second user features of each of the k second training users and the s fourth training users into the product recommendation model for training.

[0010] According to an embodiment of the present disclosure, before inputting the second user features of each of the k second training users and the s fourth training users into the product recommendation model for training, it also includes: obtaining a user feature set based on the second user features of each training user; inputting the user feature set into a decision tree model for importance analysis, wherein the importance analysis includes obtaining the importance of each sub-feature in the second user feature, and / or the importance of at least two sub-features combined; screening the user feature set based on the importance analysis result to obtain a new user feature set, wherein the screening includes removing at least one sub-feature with low importance in the user feature set.

[0011] According to an embodiment of the present disclosure, the inputting the second user features of each of the k second training users and the s fourth training users into the product recommendation model for training includes: inputting the new user feature set into the product recommendation model for training, wherein the product recommendation model includes a deep factor decomposition machine model, and the deep factor decomposition machine model is constructed based on a factor decomposition machine and a deep neural network.

[0012] Another aspect of an embodiment of the present disclosure provides a product recommendation device, comprising: a first indicator module, configured to obtain at least one product indicator for each of N first users, wherein the N first users include users who purchase products within a predetermined time period, and the at least one product indicator is obtained based on information about product purchases within the predetermined time period; a first score module, configured to obtain an evaluation score for each first user based on the at least one product indicator for each first user; a first classification module, configured to classify M third users based on the first user features of K second users to obtain S fourth users, wherein the K second users include first users among the N first users whose evaluation scores are greater than or equal to a first preset value, the S fourth users are of the same category as the K second users, and the M third users include users who have not purchased products within the predetermined time period, N, K, S, and M are respectively integers greater than or equal to 1, N is greater than or equal to K, and M is greater than or equal to S; and a first input module, configured to input the second user features of each of the K second users and the S fourth users into a pre-trained product recommendation model to obtain recommended products for each user.

[0013] Another aspect of an embodiment of the present disclosure provides a training device for a product recommendation model, comprising: a second indicator module, configured to obtain at least one product indicator for each of n first training users, wherein the n first training users include training users who purchased products within a predetermined time period, and the at least one product indicator is obtained based on information about product purchases within the predetermined time period; a second score module, configured to obtain an evaluation score for each of the first training users based on the at least one product indicator of each first training user; a second classification module, configured to classify m third training users based on the first user features of k second training users to obtain s fourth training users, wherein the k second training users include first training users among the n first training users whose evaluation scores are greater than or equal to a first preset value, the s fourth training users are of the same category as the k second training users, and the m third training users include training users who did not purchase products within the predetermined time period, and n, k, s, and m are respectively integers greater than or equal to 1, with n being greater than or equal to k and m being greater than or equal to s; and a second input module, configured to input the second user features of each of the k second training users and the s fourth training users into the product recommendation model for training.

[0014] Another aspect of an embodiment of the present disclosure provides an electronic device, comprising: one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors execute the method described above.

[0015] Another aspect of the embodiments of the present disclosure further provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, causes the processor to execute the method described above.

[0016] Another aspect of the embodiments of the present disclosure further provides a computer program product, including a computer program, which implements the above method when executed by a processor.

[0017] One or more of the above embodiments have the following beneficial effects: based on at least one product indicator of each first user, an evaluation score of each first user is obtained, and first, the N first users are screened to obtain K second users whose evaluation scores are greater than or equal to a first preset value, and then the M third users are further screened to obtain S fourth users of the same category as the K second users, and finally the second user features of each of the K second users and the S fourth users are input into a pre-trained product recommendation model to obtain a recommendation result. Thus, the two dimensions of users who have purchased the product and users who have not purchased the product are considered, and high-value users among users who have purchased the product are used as a basis to obtain high-value users among users who have not purchased the product. The pre-trained product recommendation model can be used to recommend products to high-value user groups, thereby reducing invalid operations, improving product recommendation efficiency, avoiding model training or prediction based on the entire user base to a certain extent, and reducing the marketing costs generated by recommendations for non-value users in related technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above contents and other objects, features and advantages of the present disclosure will become more apparent through the following description of the embodiments of the present disclosure with reference to the accompanying drawings, in which:

[0019] Figure 1 A diagram schematically illustrates an application scenario of a product recommendation method or a product recommendation model training method according to an embodiment of the present disclosure;

[0020] Figure 2 The following schematically shows a flow chart of a product recommendation method according to an embodiment of the present disclosure;

[0021] Figure 3 A flowchart of aggregating high-intent customer groups for financial management according to an embodiment of the present disclosure is schematically shown;

[0022] Figure 4 Schematically shows a flow chart for obtaining evaluation scores according to an embodiment of the present disclosure;

[0023] Figure 5 Schematically shows a flow chart of feature input according to an embodiment of the present disclosure;

[0024] Figure 6 A flowchart schematically illustrates a method for training a product recommendation model according to an embodiment of the present disclosure;

[0025] Figure 7 Schematically shows a flow chart of feature screening according to an embodiment of the present disclosure;

[0026] Figure 8 A flowchart schematically illustrates a method for training a product recommendation model according to another embodiment of the present disclosure;

[0027] Figure 9 The flowchart of the training and prediction of the product recommendation model according to the embodiment of the present disclosure is schematically shown;

[0028] Figure 10 Schematically shows a structural block diagram of a product recommendation device according to an embodiment of the present disclosure;

[0029] Figure 11 The following schematically shows a structural block diagram of a training device for a product recommendation model according to an embodiment of the present disclosure;

[0030] Figure 12 A block diagram of an electronic device suitable for implementing a product recommendation method or a product recommendation model training method according to an embodiment of the present disclosure is schematically shown. DETAILED DESCRIPTION

[0031] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the detailed description below, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0032] In the technical solutions disclosed herein, user authorization or consent is obtained before obtaining or collecting user personal information. The collection, storage, use, processing, transmission, provision, disclosure, and application of user personal information comply with relevant laws and regulations, employ necessary confidentiality measures, and do not violate public order and good morals.

[0033] As deep learning technology develops, factors such as personnel, hardware, and existing technologies within various industries (such as manufacturing, finance, consumer goods, and the internet) hinder its application. For example, in the banking industry, some companies are at a semi-AI stage, lagging behind in technology. Furthermore, many banks lack sufficient computing power to utilize deep learning. Therefore, when applying AI-based recommendations, it's crucial to minimize computational and memory consumption, taking into account the current characteristics of the banking industry.

[0034] In an embodiment of the present disclosure, an evaluation score of each first user is obtained based on at least one product indicator of each first user, and first, the N first users are screened to obtain K second users whose evaluation scores are greater than or equal to a first preset value, and then the M third users are further screened to obtain S fourth users of the same category as the K second users. Finally, the second user features of each of the K second users and the S fourth users are input into a pre-trained product recommendation model to obtain a recommendation result. Thus, the two dimensions of users who have purchased the product and users who have not purchased the product are considered, and high-value users among users who have purchased the product are used as a basis to obtain high-value users among users who have not purchased the product. The pre-trained product recommendation model can be used to recommend products to high-value user groups, reducing invalid operations, improving product recommendation efficiency, and to a certain extent avoiding the method of model training or prediction based on the entire user base. It also reduces the marketing costs generated by recommendations for non-value users in related technologies.

[0035] Figure 1 The following schematically illustrates an application scenario diagram of a product recommendation method or a product recommendation model training method according to an embodiment of the present disclosure.

[0036] like Figure 1 As shown, the application scenario 100 according to this embodiment may include terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 is used as a medium for providing a communication link between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or optical fiber cables.

[0037] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Various communication client applications can be installed on terminal devices 101, 102, and 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social platform software, etc. (only as examples).

[0038] The terminal devices 101 , 102 , and 103 may be various electronic devices having a display screen and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers.

[0039] The server 105 may be a server that provides various services, such as a background management server (for example only) that supports websites browsed by users using the terminal devices 101, 102, and 103. The background management server may analyze and process received data such as user requests, and feed back processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal device.

[0040] In some embodiments, R&D personnel can use terminal devices 101, 102, and 103 to interact with server 105 to perform product recommendation model training and prediction, as well as data collection, processing, and obtaining recommendation results. In other embodiments, users can use terminal devices 101, 102, and 103 to interact with server 105, and server 105 can collect user information or push products to users.

[0041] It should be noted that the product recommendation method or product recommendation model training method provided in the embodiment of the present disclosure can generally be executed by the server 105. Accordingly, the product recommendation device or product recommendation model training device provided in the embodiment of the present disclosure can generally be set in the server 105. The product recommendation method or product recommendation model training method provided in the embodiment of the present disclosure can also be executed by a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105. Accordingly, the product recommendation device or product recommendation model training device provided in the embodiment of the present disclosure can also be set in a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105.

[0042] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is merely illustrative. Any number of terminal devices, networks and servers may be provided as required.

[0043] The following will be based on Figure 1 The scene described by Figures 2 to 9 The product recommendation method and the training method of the product recommendation model of the embodiment of the present disclosure are described in detail.

[0044] Figure 2 The flowchart of the product recommendation method according to an embodiment of the present disclosure is schematically shown.

[0045] like Figure 2 As shown, the product recommendation method of this embodiment includes operations S210 to S240.

[0046] In operation S210, at least one product indicator of each of N first users is obtained, wherein the N first users include users who purchase products within a predetermined time period, and the at least one product indicator is obtained based on information about the products purchased within the predetermined time period.

[0047] According to an embodiment of the present disclosure, the at least one product indicator includes at least one of a consumption time indicator, a consumption frequency indicator, and a consumption amount indicator. The consumption time indicator includes the time of the most recent purchase of the product. The consumption frequency indicator includes the number of times the product was consumed in a specific sub-period within a predetermined time period. The consumption amount indicator includes the amount of money spent on the product in the specific sub-period.

[0048] Specifically, taking the predetermined time period of one year as an example, in order to avoid the phenomenon that some users make single, small-amount purchases or shift their consumption preferences, resulting in long-term non-purchases, data within one year can be obtained. On this basis, the time of the most recent consumption, the frequency of consumption in the most recent period, and the amount of consumption in the most recent period can be obtained. For example, a year can be divided into four sub-time periods according to quarters. For each first user, the number of consumption times and the amount of purchased products in the most recent quarter can be determined first. If there is no consumption record in the most recent quarter, statistics are collected quarter by quarter in chronological order from the most recent to the most recent. Therefore, the specific sub-time period of each first user may be different. When scoring subsequently, different scores can be given based on the distance between the specific sub-time period and the statistical time. For example, the closer the time period, the higher the score (just an example).

[0049] In operation S220 , an evaluation score of each first user is obtained based on at least one product indicator of each first user.

[0050] For example, the evaluation score can be used to screen out high-value users, where high value refers to the intention to purchase the product. Recommending products to such users has a higher purchase success rate.

[0051] In operation S230, M third users are classified according to the first user characteristics of the K second users to obtain S fourth users, wherein the K second users include first users among the N first users whose evaluation scores are greater than or equal to a first preset value, the S fourth users are of the same category as the K second users, and the M third users include users who have not purchased the product within a predetermined time period. N, K, S, and M are respectively integers greater than or equal to 1, N is greater than or equal to K, and M is greater than or equal to S.

[0052] For example, the first preset value can be set based on actual circumstances, such as taking the user evaluation score ranked in the N*20% order as the first preset value. The first user features may include features derived from basic user information, such as age, gender, star rating, assets, and consumer preferences. The classification process can be implemented using deep learning to build a classification model. Classification can also be performed by calculating the Euclidean distance between the first user features of each third user and each second user. If the Euclidean distance between a third user and any second user meets the requirements, they are considered similar.

[0053] The classifying the M third users according to the first user features of the K second users includes: inputting the first user feature of each of the K second users and the M third users into a support vector machine model to obtain a classification result output by the support vector machine model.

[0054] Support Vector Machine (SVM) is a generalized linear classifier that performs binary classification on data using supervised learning. Its decision boundary is the maximum-margin hyperplane (Maximum-Margin Hyperplane) solved for the learning sample. For example, K second users are taken as positive samples, and the first user features of the second and third users (such as Q features) collected are input. In other words, each second and third user is described as a sample in the Q-dimensional space, and the optimal hyperplane of Q-1 dimension is found in the Q-dimensional space to maximize the spacing, and finally classification is achieved through the hyperplane. The user among the M third users who is classified into the same category as the positive sample is regarded as the fourth user. The objective function of the fitting process is as shown in Formula 1).

[0055]

[0056] in, is the distance from the sample point to the hyperplane. argmax{} is the function that maximizes the independent variable. x' is the given input data (such as the first user feature of each user), and the parameters ω and b are the normal vector and intercept of the hyperplane, respectively. label is the class label. ||ω|| is the L2 norm of the parameter ω.

[0057] In operation S240 , the second user feature of each of the K second users and the S fourth users is input into a pre-trained product recommendation model to obtain a recommended product for each user.

[0058] Exemplarily, the second user characteristic can be obtained based on at least one of the user's basic information, consumer behavior information, and product purchase information. Information on product purchases within a predetermined time period may include consumer behavior information and product purchase information. Consumer behavior information includes user transaction records obtained based on consumer behavior information, including consumer behavior characteristics such as the transaction date, the number of transactions within the predetermined time period, and the transaction amount. Product characteristics, such as the characteristics of the products actually purchased by the user and the quantity of products, can be obtained based on product purchase information. Product characteristics can be obtained based on product attributes. For example, for a fund within the financial product category, attributes may include stock type, bond type, fund manager, average rate of return, and fund investment targets.

[0059] In some embodiments, the product recommendation model can be a recommendation model in the prior art. In other embodiments of the present disclosure, a deep factorization machine model (hereinafter referred to as the DeepFM model) can also be constructed based on the DeepFM algorithm. The model is constructed based on a factorization machine (FM) and a deep neural network (NN). The DeepFM algorithm combines FM and DNN in a parallel structure. The FM and DNN sides share a feature embedding layer, and the model is optimized through joint training.

[0060] For example, the recommended products may be one or more. For example, the financial products recommended to each user can be derived based on the user's second user characteristics. For the first user, financial products can be recommended based on the characteristics of the financial products they have purchased. For the first or second user, similar users can be found based on characteristics related to basic information and consumption behavior, and popular financial products can be recommended. It is also possible to combine purchased product characteristics, basic information characteristics, and consumption behavior characteristics, and during the training process, the product recommendation model learns these combined characteristics to recommend products in a targeted manner. In some embodiments, existing financial products can be recommended, and specific financial products can be generated based on the user's second user characteristics for recommendation.

[0061] According to an embodiment of the present disclosure, N first users are first screened to obtain K second users whose evaluation scores are greater than or equal to a first preset value. Then, M third users are further screened to obtain S fourth users of the same category as the K second users. Finally, the second user features of each of the K second users and the S fourth users are input into a pre-trained product recommendation model to obtain a recommendation result. Thus, the two dimensions of users who have purchased the product and users who have not purchased the product are considered, and high-value users among users who have purchased the product are used as a basis to obtain high-value users among users who have not purchased the product. The pre-trained product recommendation model can be used to recommend products to high-value user groups, reducing invalid calculations, improving product recommendation efficiency, avoiding to a certain extent the method of model training or prediction based on the entire user base, and reducing the marketing costs generated by recommendations for non-value users in related technologies.

[0062] According to the above Figure 1 One or more corresponding embodiments, the following takes the example of a bank recommending financial products, combined with Figure 3 Introduce the process of aggregating high-intent customers for financial management.

[0063] Figure 3 A flowchart for aggregating high-intent customer groups for financial management according to an embodiment of the present disclosure is schematically shown.

[0064] like Figure 3 As shown, the aggregation of high-intent financial management customers in this embodiment includes operations S310 to S370.

[0065] In operation S310, structured data is obtained. Structured data refers to the data already held by the bank. This data has a fixed format and corresponding dictionary meaning, and is often recorded in the databases of different bank business systems. It mainly includes basic user information, product information (such as purchase information or product listing information), and user transaction information (such as consumption behavior information).

[0066] In operation S320, information about users who purchased financial products is obtained. For example, users who purchased financial products in the past year are screened based on structured data and processed into various customer-based features, such as basic information features, consumption behavior features, and product purchase features.

[0067] In operation S330, the user value of the purchased financial product is analyzed. The user value is evaluated based on three indicators: the time of the last consumption, the consumption frequency in the recent period, and the consumption amount in the recent period. The process of obtaining the evaluation score in operation S220 can be referred to.

[0068] In operation S340, high-value users among users who have purchased financial products are screened. For example, a high-value user rating threshold is determined by expert rules, and based on the rating results, users with calculated scores greater than the rating threshold are screened as high-value samples.

[0069] In operation S350, information of users who have not purchased financial products is obtained. For example, users who have not purchased financial products in the past year are screened based on structured data and processed into first user features, which mainly include basic information features such as user age, gender, star rating, assets, and consumption preferences.

[0070] In operation S360, the value analysis of the user who did not purchase the financial management is performed.

[0071] In operation S370, high-value users among users who have not purchased financial management are screened. The classification process is similar to the classification process in operation S230 and will not be described in detail here.

[0072] In operation S380, a customer group with high financial management intention is summarized. For example, based on the samples analyzed in operation S340 and operation S360, a customer group with high financial management intention is summarized.

[0073] Figure 4 The flowchart of obtaining evaluation scores according to an embodiment of the present disclosure is schematically shown.

[0074] like Figure 4 As shown, obtaining the evaluation score of each first user according to at least one product indicator of each first user in operation S220 includes operations S410 to S420.

[0075] In operation S410 , a category to which each first user belongs in each product indicator is obtained, wherein each product indicator is divided into at least one category.

[0076] For example, in the process of assessing and classifying, the three indicators of the time of the most recent consumption, the frequency of consumption in the most recent period, and the amount of consumption in the most recent period are considered to have obvious differences in magnitude. According to business experts and data analysis, each indicator can be equally grouped according to the corresponding dimension (such as the time dimension, the frequency dimension, or the amount dimension) and then standardized into 5 categories. The user group is divided into 5*5*5 categories of users (this is just an example, the number of categories is not limited).

[0077] In operation S420 , an evaluation score of each first user is obtained according to the score and weight corresponding to the category of each product indicator.

[0078] For example, during the calculation process, the importance of each indicator can be determined through business experts and automated data analysis, and different scores can be assigned to each category of each indicator. The weight corresponding to each category can be a uniform weight assigned based on the importance of the indicator, or each category can be assigned a separate weight.

[0079] In some embodiments, the user's index is measured by weighted summation of individual indexes, and the evaluation score is calculated according to formula 2).

[0080]

[0081] Among them, w is the weight of each product indicator, P is the category, and the subscript i represents the i-th product indicator.

[0082] According to an embodiment of the present disclosure, the magnitude differences between various product indicators and the differences between various users under each product indicator dimension may be considered to reasonably determine the evaluation score of each first user.

[0083] Figure 5 The flowchart of feature input according to an embodiment of the present disclosure is schematically shown.

[0084] like Figure 5 As shown, inputting the second user feature of each of the K second users and the S fourth users into the pre-trained product recommendation model in operation S240 includes operations S510 to S520.

[0085] In operation S510, feature screening is performed on the second user feature of each user according to the importance analysis result, wherein the importance analysis result includes the importance of at least one sub-feature in the second user feature, and the importance analysis result is obtained by the importance analysis pre-performed by the decision tree model.

[0086] For example, feature filtering can include removing at least one of basic information features, consumer behavior features, and product purchase features. Alternatively, it can include removing at least one of multiple sub-features of basic information features, consumer behavior features, and product purchase features. Sub-features include user age, gender, star rating, assets, consumer preferences, transaction history, number of transactions within a predetermined time period, transaction amount, product features, and product quantity.

[0087] In some embodiments, the importance analysis can be performed using a decision tree model before training to achieve the purpose of feature screening. The purpose of feature screening in operation S510 can be consistent with the features of the input model during the training process.

[0088] In other embodiments, operation S510 may be performed each time a product recommendation is made to perform feature screening, so as to select appropriate features based on the customer group for each product recommendation. In these embodiments, multiple feature screenings (e.g., training multiple sub-feature combinations) or other adaptive operations may be performed during training to maintain the effectiveness of the model.

[0089] In operation S520 , the filtered second user feature of each user is input into a pre-trained product recommendation model.

[0090] According to the embodiments of the present disclosure, feature screening during product recommendation can not only consider the impact of a single sub-feature on financial purchases, but also automatically implement high-order feature combinations, consider the impact of feature combinations on financial product purchases, reduce invalid calculations, and improve recommendation efficiency.

[0091] The following combination Figures 6 to 8 , introduces the training process of the product recommendation model.

[0092] Figure 6 The flowchart of the training method of the product recommendation model according to the embodiment of the present disclosure is schematically shown.

[0093] like Figure 6 As shown, the product recommendation model training method of this embodiment includes operations S610 to S640.

[0094] In operation S610, at least one product indicator of each of n first training users is obtained, wherein the n first training users include training users who purchase products within a predetermined time period, and the at least one product indicator is obtained based on information of product purchases within the predetermined time period.

[0095] For example, the training user refers to a user used to collect user features to obtain training samples. The product indicators, predetermined time periods, and purchased product information obtained from the training user can refer to operation S210.

[0096] In operation S620 , an evaluation score of each first training user is obtained based on at least one product indicator of each first training user.

[0097] For example, the process of obtaining the evaluation score may refer to the method described in operations S410 to S420.

[0098] In operation S630, m third training users are classified according to the first user characteristics of k second training users to obtain s fourth training users, wherein the k second training users include first training users among the n first training users whose evaluation scores are greater than or equal to a first preset value, the s fourth training users are of the same category as the k second training users, and the m third training users include training users who have not purchased products within a predetermined time period. n, k, s and m are respectively integers greater than or equal to 1, n is greater than or equal to k, and m is greater than or equal to S.

[0099] Exemplarily, the first user feature of each of the k second training users and the m third training users is input into the support vector machine model to obtain a classification result output by the support vector machine model.

[0100] In operation S640 , the second user feature of each of the k second training users and the s fourth training users is input into a product recommendation model for training.

[0101] For example, each training user corresponds to a training sample {X, Y}, where X is a vector representing the second user's characteristics and Y is the product label (or labels) that each training user has purchase intentions for. During training, the product recommendation model outputs (or more) predicted recommended products for each training user. A loss function is used to calculate the degree of difference between the product label and the predicted recommended product. This is used to update the parameters of the product recommendation model to achieve convergence of the loss function.

[0102] According to the embodiments of the present disclosure, the two dimensions of training users who have purchased products and training users who have not purchased products are taken into consideration, and high-value users among training users who have purchased products are used as a basis to obtain high-value users among training users who have not purchased products. This can reduce invalid calculations in the training process, improve training efficiency, and overcome the problem of insufficient computing power.

[0103] Figure 7 The flowchart of feature screening according to an embodiment of the present disclosure is schematically shown.

[0104] like Figure 7 As shown, feature screening may be performed before operation S640. The feature screening in this embodiment includes operations S710 to S730.

[0105] In operation S710 , a user feature set is obtained according to the second user feature of each training user.

[0106] Exemplarily, the user feature set is formed into a data set including the second user features of all training users through operations such as missing value processing, dimension elimination, and outlier processing.

[0107] In operation S720 , the user feature set is input into a decision tree model for importance analysis, wherein the importance analysis includes obtaining the importance of each sub-feature in the second user feature and / or the importance of a combination of at least two sub-features.

[0108] In some embodiments, sub-features may include user age, gender, star rating, assets, consumption preferences, number of transactions since today, number of transactions within a predetermined time period, transaction amount, product features, and product quantity.

[0109] The decision tree model is a tree-based model. For example, age is used as a splitting feature, with different age groups as the root node of a tree. Multiple subtrees are generated, and leaf nodes in each subtree are selected from other subfeatures. For example, the Gini coefficient is used to determine the splitting feature (a subfeature category, such as age) and the split point (the specific content within a subfeature category, such as middle age). The optimal feature and split point are then determined as leaf nodes.

[0110] For example, the Gini coefficient can be used to represent importance. To study the impact of each feature's importance on financial product purchases, considering a combination of features, a gradient boosting decision tree model can be used to determine the impact of each sub-feature on financial product purchases. First, each feature is traversed, and then all possible split points for each feature are traversed to find the most significant and optimal split point.

[0111] In operation S730 , the user feature set is screened based on the importance analysis result to obtain a new user feature set, wherein the screening includes removing at least one sub-feature with low importance in the user feature set.

[0112] According to the embodiments of the present disclosure, in addition to studying and exploring the impact of individual sub-features on product purchases during training, high-level feature combinations are automatically implemented to explore the impact of sub-feature combinations on product purchases, simplifying the current feature engineering work that requires a large amount of resource investment and freeing up a large amount of modeling resources. By filtering features based on a tree structure and considering the impact of features and feature combinations on the model through the tree structure, this effectively avoids the dimensionality curse caused by invalid features and network jitter caused by overfitting of the model in the later stages. It can also solve the problem of insufficient computing power in the existing banking industry.

[0113] According to an embodiment of the present disclosure, in operation S640, inputting the second user features of each of the k second training users and the s fourth training users into the product recommendation model for training includes: inputting a new set of user features into the product recommendation model for training, wherein the product recommendation model includes a deep factor decomposition machine model, and the deep factor decomposition machine model is constructed based on a factor decomposition machine and a deep neural network.

[0114] DeepFM adds DNN to handle the problem of multiple feature intersections in addition to the original FM's problem of handling two feature intersections. Through joint training, the efficiency and performance are greatly improved. Figure 8 Further introduction to the process of training the Deep Factorization Machine (DeepFM) model.

[0115] Figure 8 The flowchart of the method for training a product recommendation model according to another embodiment of the present disclosure is schematically shown.

[0116] like Figure 8 As shown, the product recommendation model training method of this embodiment includes operations S810 to S860.

[0117] In operation S810 , a training set is input, and a new user feature set and a product label corresponding to each training user are input into the model as the training set.

[0118] In operation S820, feature encoding is performed. Categorical feature encoding, also known as one-hot encoding, encodes categorical data that cannot be directly input into the model. The encoder binarizes the categories and expands the values ​​of discrete features into Euclidean space. Each value of the discrete features corresponds to each point in Euclidean space. The feature length is expanded and divided into multiple fields, ultimately concatenating the one-hot encoded categorical features with the numerical features.

[0119] In operation S830, the embedding layer is implemented. This involves embedding the one-hot processed features into vectors. One-hot processed data is often very high-dimensional, with only one index being 1 and the rest being 0s, making it very sparse. The embedding operation allows the input features to be calculated in a lower-dimensional space. The key idea is that each dimension of the one-hot feature and each numerical feature is mapped to a k-dimensional embedding latent vector. Each dense vector is then concatenated horizontally (see Equation 3).

[0120]

[0121] Among them, X' j It represents the jth feature field of the input and has been onehot encoded. The parameters of the embedding layer are reduced by the field operation. Finally, the original numerical features are spliced ​​together and then used as the input of Deep and FM.

[0122] In operation S840, the FM layer is implemented. The key point is to introduce a second-order combination term x on the basis of the original single feature to study the cross-feature situation. i x j , refer to formula 4).

[0123]

[0124] Among them, n represents the number of sample features, x i is the value of the i-th sub-feature, w0, w i , w ij are the parameters of the model.

[0125] x in y(x)i and x j When both are non-zero, the combined features are meaningful, and each parameter w ij Training requires a lot of x i and x j However, in the actual banking financial scenario, due to the large number of customers and product users and few correlations, the data is very sparse. When the data scale is larger, the scoring matrix will generally be sparse, and the model effect will be worse. Therefore, it is easy to have insufficient training samples and cause w ij How does the inaccurate parameters affect the model performance? Therefore, the FM layer introduces the concept of latent vector, namely v i Represents the i-th dimension feature, with <v i , v j > replace w ij , the two-dimensional cross information between features is calculated by replacing the combined features with vector dot products, solving the problems of data sparsity and computational complexity.

[0126] Related art financial product recommendation methods based on collaborative filtering cannot effectively recommend new products or new users without behavioral data, resulting in a cold start problem. The disclosed embodiments overcome the limitation of product and user independence by calculating the dot product of product and user vectors, further breaking the barrier of sparse product-user association matrices and resolving the cold start problem (see Equation 5).

[0127]

[0128] In operation S850, a hidden layer is implemented. All dense vectors implemented by the embedding operation are horizontally concatenated, and then outputted through multiple layers of linear mapping and nonlinear transformation.

[0129] In operation S850, the output is merged. The output of the output layer is the accumulation of the FM layer result and the Hidden layer result, realizing the interactive fusion of low-order and high-order features. Then, it is transformed nonlinearly by the Sigmoid function to obtain the predicted probability output (see Equation 6).

[0130] γ=sigmoid(y FM +y DNN ) Equation 6)

[0131] Among them, γ is the predicted probability output (such as used to represent whether there is an intention to buy a product and the probability value), y FM is the result of FM Layer, y DNN This is the result of Hidden Layer.

[0132] According to the embodiments of the present disclosure, taking the banking industry as an example, the DeepFM model-based recommendation application process can combine the characteristics of the banking industry to reduce computing and memory consumption, and solve the problems of sparse scoring features, cold start, and insufficient computing power in the existing intelligent recommendation scenarios of financial products in the banking industry.

[0133] Figure 9 The flowchart of training and prediction of a product recommendation model according to an embodiment of the present disclosure is schematically shown.

[0134] like Figure 9 As shown, the training and prediction of the product recommendation model in this embodiment includes operations S910 to S950.

[0135] In operation S910, information is acquired, such as various types of customers, products, and transactions. For example, all users of a banking system are acquired and processed into various characteristics based on the customer dimension (basic user information characteristics, user behavior data, user purchase information characteristics, etc.).

[0136] In operation S920, the sample is distilled. Referring to operations S210 to S230, S310 to S370, and S410 to S420, the sample distillation process is described, which will not be elaborated here.

[0137] In operation S930, feature screening: The feature set is screened and sorted to select highly available features. DeepFM models are generally considered robust when dealing with high-dimensional features, enabling implicit feature selection. However, when parameters are too large, current banking computing power is insufficient. Furthermore, insufficient data volume can lead to inadequate learning, inability to achieve optimal feature selection, and subsequent network jitter. Therefore, a gradient boosting decision tree is used before training to consider the impact of each model feature and its combination on the results, thus performing feature screening.

[0138] In operation S940, the model is trained. The processed training samples and features are fed into the DeepFM model for training. In addition to the common one-hot encoding layer, an embedding layer is introduced to convert high-dimensional, sparse data into low-dimensional, dense features. The embedding results are shared between the FM layer and the hidden layer to achieve high- and low-order feature processing. The logistic loss is ultimately used based on the estimated performance of the FM and hidden layer transformations. The training network is continuously optimized using the F1 evaluation metric.

[0139] In operation S950, the model is applied: the predicted list is applied to customer marketing, and customer reach is achieved through strategic approaches. The user interest characteristics captured through training in operation S940 are used to obtain the optimal model. In a real-world environment, the user's purchase probability for financial products is predicted. Based on marketing rules, the top N users are further ranked and a strategy is developed to reach the customer through mobile banking or relationship managers.

[0140] According to the embodiments of the present disclosure, various types of customer, product, and transaction information are obtained through structured data. A feature set is formed through operations such as sample distillation and dimensionality reduction, reducing computing power consumption. The feature set is input into the model for one-hot and embedding operations to make the features low-dimensional and dense, solving the problem of data sparsity. High-order and low-order feature processing is achieved by sharing the embedding results through the FM layer and hidden layer. Finally, real-world data is input into the model to obtain a set of recommended users and their corresponding recommended products, forming a corresponding marketing strategy to reach customers through various channels.

[0141] Based on the above-mentioned product recommendation method and product recommendation model training method, the present disclosure also provides a product recommendation device and a product recommendation model training device, which will be described in detail below.

[0142] Figure 10 The following schematically shows a structural block diagram of a product recommendation device according to an embodiment of the present disclosure.

[0143] like Figure 10 As shown, the product recommendation device 1000 of this embodiment includes a first indicator module 1010 , a first score module 1020 , a first classification module 1030 and a first input module 1040 .

[0144] The first indicator module 1010 can perform operation S210 to obtain at least one product indicator for each of N first users, wherein the N first users include users who purchase products within a predetermined time period, and the at least one product indicator is obtained based on information about the products purchased within the predetermined time period.

[0145] The first score module 1020 may perform operation S220 for obtaining an evaluation score of each first user according to at least one product indicator of each first user.

[0146] According to an embodiment of the present disclosure, the first score module 1020 may perform operations S410 to S420, which will not be described in detail herein.

[0147] The first classification module 1030 can perform operation S230 to classify M third users according to the first user characteristics of K second users to obtain S fourth users, wherein the K second users include first users among the N first users whose evaluation scores are greater than or equal to a first preset value, the S fourth users are of the same category as the K second users, and the M third users include users who have not purchased the product within a predetermined time period, and N, K, S and M are respectively integers greater than or equal to 1, N is greater than or equal to K, and M is greater than or equal to S.

[0148] The first input module 1040 may perform operation S240 for inputting the second user feature of each of the K second users and the S fourth users into a pre-trained product recommendation model to obtain a recommended product for each user.

[0149] According to an embodiment of the present disclosure, the first input module 1040 may perform operations S510 to S520, which will not be described in detail herein.

[0150] Figure 11 The structural block diagram of the training device for the product recommendation model according to an embodiment of the present disclosure is schematically shown.

[0151] like Figure 11 As shown, the product recommendation model training device 1100 of this embodiment includes a second indicator module 1110 , a second score module 1120 , a second classification module 1130 and a second input module 1140 .

[0152] The second indicator module 1110 can perform operation S610 to obtain at least one product indicator for each of n first training users, wherein the n first training users include training users who purchase products within a predetermined time period, and the at least one product indicator is obtained based on information on purchase of products within the predetermined time period.

[0153] The second score module 1120 may perform operation S620 to obtain an evaluation score of each first training user according to at least one product indicator of each first training user.

[0154] The second classification module 1130 can perform operation S630 to classify m third training users according to the first user characteristics of k second training users to obtain s fourth training users, wherein the k second training users include first training users among the n first training users whose evaluation scores are greater than or equal to a first preset value, the s fourth training users are of the same category as the k second training users, and the m third training users include training users who have not purchased products within a predetermined time period, and n, k, s and m are respectively integers greater than or equal to 1, n is greater than or equal to k, and m is greater than or equal to s.

[0155] The second input module 1140 may perform operation S640 for inputting the second user feature of each of the k second training users and the s fourth training users into the product recommendation model for training.

[0156] According to an embodiment of the present disclosure, the training device 1100 may further include a feature screening module, which may perform operations S710 to S730, which will not be described in detail here.

[0157] It should be noted that the implementation methods, technical problems solved, functions implemented, and technical effects achieved of each module / unit / sub-unit in the device part embodiment are the same or similar to the implementation methods, technical problems solved, functions implemented, and technical effects achieved of each corresponding step in the method part embodiment, and will not be repeated here.

[0158] According to an embodiment of the present disclosure, any multiple modules in the product recommendation device 1000 or the product recommendation model training device 1100 can be combined into a single module, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functionality of one or more of these modules can be combined with at least part of the functionality of other modules and implemented in a single module.

[0159] According to an embodiment of the present disclosure, at least one module in the product recommendation device 1000 or the product recommendation model training device 1100 can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or implemented in any one of the three implementation methods of software, hardware and firmware or in an appropriate combination of any of them. Alternatively, at least one module in the product recommendation device 1000 or the product recommendation model training device 1100 can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0160] Figure 12 A block diagram of an electronic device suitable for implementing a product recommendation method or a product recommendation model training method according to an embodiment of the present disclosure is schematically shown.

[0161] like Figure 12As shown, the electronic device 1200 according to an embodiment of the present disclosure includes a processor 1201, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1202 or a program loaded from a storage part 1208 into a random access memory (RAM) 1203. The processor 1201 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or a related chipset and / or a dedicated microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 1201 may also include an onboard memory for caching purposes. The processor 1201 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present disclosure.

[0162] Various programs and data required for the operation of the electronic device 1200 are stored in the RAM 1203. The processor 1201, ROM 1202, and RAM 1203 are connected to each other via a bus 1204. The processor 1201 executes the various operations of the method flow according to the embodiment of the present disclosure by executing the programs in the ROM 1202 and / or RAM 1203. It should be noted that the programs may also be stored in one or more memories other than the ROM 1202 and RAM 1203. The processor 1201 may also execute the various operations of the method flow according to the embodiment of the present disclosure by executing the programs stored in one or more memories.

[0163] According to an embodiment of the present disclosure, the electronic device 1200 may further include an input / output (I / O) interface 1205, which is also connected to the bus 1204. The electronic device 1200 may further include one or more of the following components connected to the I / O interface 1205: an input portion 1206 including a keyboard, a mouse, etc.; an output portion 1207 including a cathode ray tube (CRT), a liquid crystal display (LCD), a speaker, etc.; a storage portion 1208 including a hard disk, etc.; and a communication portion 1209 including a network interface card such as a LAN card or a modem. The communication portion 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to the I / O interface 1205 as needed. A removable medium 1211, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed in the drive 1210 as needed, so that a computer program read therefrom can be installed into the storage portion 1208 as needed.

[0164] The present disclosure also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments. Alternatively, the computer-readable storage medium may exist independently, without being incorporated into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the methods according to the embodiments of the present disclosure.

[0165] According to an embodiment of the present disclosure, a computer-readable storage medium may be a non-volatile computer-readable storage medium, for example, it may include but is not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, apparatus, or device. For example, according to an embodiment of the present disclosure, a computer-readable storage medium may include the ROM 1202 and / or RAM 1203 described above and / or one or more memories other than ROM 1202 and RAM 1203.

[0166] The embodiments of the present disclosure also include a computer program product, which includes a computer program containing program code for executing the method shown in the flowchart. When the computer program product is run in a computer system, the program code is used to enable the computer system to implement the method provided by the embodiments of the present disclosure.

[0167] The computer program executes the above functions defined in the system / device of the embodiment of the present disclosure when the processor 1201 executes the computer program. According to the embodiment of the present disclosure, the system, device, module, unit, etc. described above can be implemented by a computer program module.

[0168] In one embodiment, the computer program may be stored on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may be transmitted and distributed in the form of a signal on a network medium, downloaded and installed via the communication portion 1209, and / or installed from the removable medium 1211. The program code contained in the computer program may be transmitted using any appropriate network medium, including but not limited to wireless, wired, or any suitable combination thereof.

[0169] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 1209 and / or installed from the removable medium 1211. When the computer program is executed by the processor 1201, the above-described functions defined in the system of the embodiment of the present disclosure are performed. According to the embodiment of the present disclosure, the systems, devices, means, modules, units, etc. described above can be implemented by computer program modules.

[0170] According to an embodiment of the present disclosure, the program code for executing the computer program provided by the embodiment of the present disclosure can be written in any combination of one or more programming languages. Specifically, these computer programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, python, "C" or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (for example, using an Internet service provider to connect via the Internet).

[0171] The embodiments of the present disclosure are described above. However, these embodiments are for illustrative purposes only and are not intended to limit the scope of the present disclosure. Although each embodiment has been described separately above, this does not mean that the measures in each embodiment cannot be used in combination to advantage. The scope of the present disclosure is defined by the appended claims and their equivalents. Without departing from the scope of the present disclosure, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present disclosure.

Claims

1. A product recommendation method, comprising: Obtaining at least one product indicator for each of N first users, wherein the N first users include users who purchased a product within a predetermined time period, and the at least one product indicator is obtained based on information about product purchases within the predetermined time period; the at least one product indicator includes at least one of a consumption time indicator, a consumption frequency indicator, and a consumption amount indicator; Obtaining an evaluation score of each first user based on at least one product indicator of each first user, including: obtaining a category to which each first user belongs in each product indicator, wherein each product indicator is divided into at least one category; and obtaining an evaluation score of each first user based on a score and a weight corresponding to the category of each product indicator; Classifying M third users according to the first user features of the K second users to obtain S fourth users, wherein the K second users include first users among the N first users whose evaluation scores are greater than or equal to a first preset value, the S fourth users are of the same category as the K second users, and the M third users include users who have not purchased the product within the predetermined time period, N, K, S, and M are respectively integers greater than or equal to 1, N is greater than or equal to K, and M is greater than or equal to S; The second user feature of each of the K second users and the S fourth users is input into a pre-trained product recommendation model to obtain a recommended product for each user.

2. The method according to claim 1, wherein The consumption time indicator includes the time of the most recent purchase of the product; The consumption frequency index includes the number of times consumption is made in a specific sub-time period within the predetermined time period; The consumption amount indicator includes the amount of money spent on products purchased during the specific sub-time period.

3. The method according to claim 1, wherein Inputting the second user feature of each of the K second users and the S fourth users into the pre-trained product recommendation model includes: Performing feature screening on the second user feature of each user according to an importance analysis result, wherein the importance analysis result includes the importance of at least one sub-feature of the second user feature, and the importance analysis result is obtained by performing importance analysis in advance on a decision tree model; The filtered second user feature of each user is input into a pre-trained product recommendation model.

4. The method according to claim 1, wherein The classifying the M third users according to the first user features of the K second users includes: The first user feature of each of the K second users and the M third users is input into a support vector machine model to obtain a classification result output by the support vector machine model.

5. A method for training a product recommendation model, comprising: Obtaining at least one product indicator for each of n first training users, wherein the n first training users include training users who purchased a product within a predetermined time period, and the at least one product indicator is obtained based on information about product purchases within the predetermined time period; the at least one product indicator includes at least one of a consumption time indicator, a consumption frequency indicator, and a consumption amount indicator; Obtaining an evaluation score for each first training user based on at least one product indicator of each first training user, including: obtaining a category to which each first training user belongs in each product indicator, wherein each product indicator is divided into at least one category; and obtaining an evaluation score for each first training user based on a score and a weight corresponding to the category of each product indicator; Classifying m third training users according to the first user features of the k second training users to obtain s fourth training users, wherein the k second training users include first training users among the n first training users whose evaluation scores are greater than or equal to a first preset value, the s fourth training users are of the same category as the k second training users, and the m third training users include training users who have not purchased the product within the predetermined time period, and n, k, s, and m are respectively integers greater than or equal to 1, n is greater than or equal to k, and m is greater than or equal to s; The second user feature of each of the k second training users and the s fourth training users is input into the product recommendation model for training.

6. The method according to claim 5, wherein: Before inputting the second user feature of each of the k second training users and the s fourth training users into the product recommendation model for training, the method further includes: Obtaining a user feature set according to the second user feature of each training user; Inputting the user feature set into a decision tree model for importance analysis, wherein the importance analysis includes obtaining the importance of each sub-feature in the second user feature and / or the importance of at least two sub-features combined; The user feature set is screened based on the importance analysis result to obtain a new user feature set, wherein the screening includes removing at least one sub-feature with low importance in the user feature set.

7. The method according to claim 6, wherein: Inputting the second user feature of each of the k second training users and the s fourth training users into the product recommendation model for training includes: The new user feature set is input into a product recommendation model for training, wherein the product recommendation model includes a deep factorization machine model, and the deep factorization machine model is constructed based on a factorization machine and a deep neural network.

8. A product recommendation device, comprising: A first indicator module is configured to obtain at least one product indicator for each of N first users, wherein the N first users include users who purchased a product within a predetermined time period, and the at least one product indicator is obtained based on information about product purchases within the predetermined time period; the at least one product indicator includes at least one of a consumption time indicator, a consumption frequency indicator, and a consumption amount indicator; A first score module is configured to obtain an evaluation score of each first user based on at least one product indicator of each first user, comprising: obtaining a category to which each first user belongs in each product indicator, wherein each product indicator is divided into at least one category; and obtaining an evaluation score of each first user based on a score and a weight corresponding to the category of each product indicator; a first classification module, configured to classify M third users based on first user features of K second users to obtain S fourth users, wherein the K second users include first users among the N first users whose evaluation scores are greater than or equal to a first preset value, the S fourth users are of the same category as the K second users, and the M third users include users who have not purchased the product within the predetermined time period, wherein N, K, S, and M are respectively integers greater than or equal to 1, N is greater than or equal to K, and M is greater than or equal to S; The first input module is used to input the second user feature of each of the K second users and the S fourth users into a pre-trained product recommendation model to obtain a recommended product for each user.

9. A training device for a product recommendation model, comprising: a second indicator module, configured to obtain at least one product indicator for each of n first training users, wherein the n first training users include training users who purchased products within a predetermined time period, and the at least one product indicator is obtained based on information about product purchases within the predetermined time period; the at least one product indicator includes at least one of a consumption time indicator, a consumption frequency indicator, and a consumption amount indicator; A second scoring module is configured to obtain an evaluation score of each first training user based on at least one product indicator of each first training user, comprising: obtaining a category to which each first training user belongs in each product indicator, wherein each product indicator is divided into at least one category; and obtaining an evaluation score of each first training user based on a score and a weight corresponding to the category of each product indicator; a second classification module, configured to classify m third training users according to the first user features of the k second training users to obtain s fourth training users, wherein the k second training users include first training users among the n first training users whose evaluation scores are greater than or equal to a first preset value, the s fourth training users are of the same category as the k second training users, and the m third training users include training users who have not purchased the product within the predetermined time period, and n, k, s, and m are respectively integers greater than or equal to 1, n is greater than or equal to k, and m is greater than or equal to s; The second input module is used to input the second user feature of each of the k second training users and the s fourth training users into the product recommendation model for training.

10. An electronic device comprising: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors are enabled to execute the method according to any one of claims 1 to 7.

11. A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, causes the processor to perform the method according to any one of claims 1 to 7.

12. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images