Commodity information recommendation method and device, equipment, medium and program product

Abstract

The invention discloses a commodity information recommendation method and device, equipment, a medium and a program product, and belongs to the field of data processing. The method comprises the following steps: training a recommendation model according to application tendency information of sample users for commodities to obtain a preference commodity information set of the sample users predicted by the recommendation model; based on the application tendency information, the preference commodity information set and a large language model, performing iterative training on the recommendation model until performance parameters of the recommendation model meet preset conditions, and determining the recommendation model as a target recommendation model, the large language model is used for extracting sample information used for carrying out iterative training on the recommendation model from the preference commodity information set; and utilizing the target recommendation model to obtain recommended commodity information according to the application tendency information of the tested user for the commodity, and pushing the recommended commodity information to the tested user. According to the embodiment of the invention, the accuracy of recommending the commodity information to the user by the recommendation model can be improved.

Description

Technical Field

[0001] The present application belongs to the field of data processing, and in particular, relates to a method, device, equipment, medium and program product for recommending commodity information. Background Art

[0002] With the continuous development of electronic information technology, users' demand for information acquisition is also increasing. In order to meet users' needs, information service providers can analyze users' behaviors and preferences, recommend personalized information to users, and improve user experience.

[0003] In the scenario where users apply for products, most users will only apply for a few types of products, and a large number of types of products will only be applied for by a few users. This power-law distribution between users and products makes the collaborative signals between users and products obtained by the recommendation model based on user behavior and preferences very sparse, making it difficult to identify similar users who apply for similar products, making it difficult for the recommendation model to accurately recommend products to users. Summary of the invention

[0004] The embodiments of the present application provide a method, apparatus, device, medium and program product for recommending product information, which can improve the accuracy of a recommendation model in recommending product information to users.

[0005] In a first aspect, an embodiment of the present application provides a method for recommending product information, including: training a recommendation model based on sample users' application tendency information for products to obtain a set of preferred product information of sample users predicted by the recommendation model; iteratively training the recommendation model based on the application tendency information, the set of preferred product information and a large language model until the performance parameters of the recommendation model meet preset conditions, determining the recommendation model as a target recommendation model, and using the large language model to extract sample information for iterative training of the recommendation model from the set of preferred product information; and using the target recommendation model to obtain recommended product information based on the application tendency information of the tested user for products and push it to the tested user.

[0006] In a second aspect, an embodiment of the present application provides a device for recommending product information, including: a prediction result acquisition module, used to train a recommendation model according to the sample user's application tendency information for products, and obtain a set of preferred product information of the sample user predicted by the recommendation model; an enhanced training module, used to iteratively train the recommendation model based on the application tendency information, the preferred product information set and the large language model, until the performance parameters of the recommendation model meet the preset conditions, and the recommendation model is determined as the target recommendation model, and the large language model is used to extract sample information for iterative training of the recommendation model from the preferred product information set; an information push module, used to use the target recommendation model to obtain recommended product information based on the application tendency information of the tested user for products and push it to the tested user.

[0007] In a third aspect, an embodiment of the present application provides an electronic device, comprising: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the method for recommending product information of the first aspect is implemented.

[0008] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon, and when the computer program instructions are executed by a processor, the method for recommending product information of the first aspect is implemented.

[0009] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which implements the method for recommending product information in the first aspect when the computer program is executed by a processor.

[0010] The embodiments of the present application provide a method, apparatus, device, medium and program product for recommending product information. A recommendation model is trained based on application tendency information that can reflect the sample user's application tendency for products, and a set of sample user's preferred product information predicted by the recommendation model is obtained. Based on the application tendency information, the inductive reasoning ability of the large language model can be used to further extract information from the set of preferred product information as sample information to train the recommendation model, until the performance parameters of the recommendation model meet the preset conditions, and the recommendation model is put into use as a target recommendation model to recommend product information to the tested user. In the iterative process of training the recommendation model using the output of the recommendation model as the input of the large language model and the output of the large language model as sample information, the collaborative signal is continuously enhanced by using the output of the large language model as sample information to improve the problem of lack of collaborative signals, thereby enhancing the recommendation model and improving the accuracy of the recommendation model in recommending product information to users. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solution of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0012] Figure 1 A flowchart of a method for recommending product information provided in an embodiment of the present application;

[0013] Figure 2 A flowchart of a method for recommending product information provided by another embodiment of the present application;

[0014] Figure 3 A schematic diagram of an example of a logical framework of a large language model enhanced recommendation algorithm based on a course learning paradigm provided in an embodiment of the present application;

[0015] Figure 4 A schematic diagram of an example of embedded representation of user-product interaction information provided in an embodiment of the present application;

[0016] Figure 5 A schematic diagram of the structure of a device for recommending product information provided in an embodiment of the present application;

[0017] Figure 6 A schematic diagram of the structure of an electronic device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0018] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are intended only to explain the present application, rather than to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present application by showing examples of the present application. It should be noted that the acquisition, storage, use, and processing of information and data in the embodiments of the present application are authorized by users or relevant agencies, in compliance with the relevant provisions of national laws and regulations.

[0019] With the continuous development of electronic information technology, users' demand for information acquisition is also increasing. In order to meet users' needs, information service providers can recommend personalized information to users by analyzing users' behaviors and preferences to improve user experience. In the scenario where users apply for products, most users will only apply for a few types of products, and a large number of types of products will only be applied for by a few users. This power-law distribution between users and products makes the collaborative signals between users and products obtained by the recommendation model based on user behavior and preferences very sparse, while users who have not applied for products or products that have not been applied for do not have collaborative signals. Collaborative signals can reflect the similarity of users' behaviors in applying for products. In the absence of collaborative signals, it is difficult to determine similar users who apply for similar products, making it difficult for the recommendation model to accurately recommend products to users.

[0020] The present application provides a method, device, equipment, medium and program product for recommending product information, which uses a large model based on the curriculum learning paradigm to enhance the recommendation model. In the model training process of the recommendation model, the powerful inductive reasoning ability of the large language model can be used to extract product information that meets the user's preferences from the product information output by the recommendation model based on the interaction information between the user and the product or the user portrait information, etc., and use the original sample and the pseudo sample together to train the recommendation model, and repeat the process until the recommendation model meets the requirements and is put into use. Pseudo samples can enhance the collaborative signal and enhance the recommendation model, thereby improving the accuracy of the recommendation model in recommending product information to users.

[0021] The following describes the recommendation method, device, equipment, medium and program product for the product information provided by this application.

[0022] A first aspect of the present application provides a method for recommending product information, which can be applied to a device, equipment, etc. for recommending product information. That is, the method for recommending product information is executed by a device, equipment, etc. for recommending product information. Figure 1 A flowchart of a method for recommending product information provided in an embodiment of the present application, such as Figure 1 As shown, the product information recommendation method may include steps S101 to S103.

[0023] In step S101, the recommendation model is trained according to the sample user's application tendency information for commodities, and a set of the sample user's preferred commodity information predicted by the recommendation model is obtained.

[0024] The sample users include users corresponding to the samples used in the model training process for the recommendation model. In the embodiment of the present application, the goods may include physical goods, services, businesses, etc. For example, the goods may include common food, daily necessities, etc., and the goods may also include credit cards, membership cards, etc. The application for goods can be understood as purchase, handling, etc. For example, for physical goods, the application for goods is to purchase goods; for services, businesses, etc., the application for goods is to handle services and handle businesses. The sample user's application tendency information for goods is used to characterize the user's application tendency for goods, such as the user tends to apply for a certain or certain goods, and the user tends not to apply for a certain or certain goods. In some examples, the application tendency information may include but is not limited to user product interaction information and user portrait information. The user product interaction information characterizes the user's historical application for goods, that is, the user product interaction information can characterize the goods that have application interactions with the user, such as goods purchased by the user, services and businesses handled by the user, etc. User portrait information is label information abstracted from information such as the user's social attributes, living habits, and consumption behaviors. The user's preferences can be determined to a certain extent based on the user portrait information.

[0025] The recommendation model is used to output product information recommended to the user. By using the application tendency information to train the recommendation model, the recommendation model can learn the user's preference for products, so that the recommendation model can predict the product information of the products that the user tends to apply for. The recommendation model can output the sample user's preferred product information set based on the sample user's application tendency information. The preferred product information set includes the product information of the user's preferred products predicted by the recommendation model. The number of product information in the preferred product information set can be one or more than two, which is not limited here.

[0026] For example, if the original sample used to train the recommendation model, that is, the application tendency information, is represented as in, is the original sample, is a positive sample, is a negative sample. A negative sample can be a sample constructed using negative sampling. The set of preferred product information output by the recommendation model based on the application tendency information can be expressed as in, is a collection of preferred product information. The product information of the product that the recommendation model believes the sample user prefers most. The product information of the product that the recommendation model considers the sample user to be the second most preferred. is the product information of the product that the recommendation model believes the sample user prefers, where C is a positive integer. In other words, the set of preferred product information It includes the product information of the first C products arranged from high to low according to the preference of the sample users.

[0027] In some examples, when the sample user has a history of applying for goods, that is, the sample user has applied for goods before, and has user-goods interaction information, the application tendency information includes the user-goods interaction information, and the recommendation model may include a model for making recommendations based on the similarity of user behaviors. For example, the recommendation model may include but is not limited to a collaborative filtering (CF) model, which predicts content that the user may be interested in by analyzing user behaviors and preferences. The collaborative filtering model may specifically adopt but is not limited to the SimpleX model. The historically applied goods are goods that the user has applied for.

[0028] In other examples, when the sample user has no history of applying for products, that is, the user has never applied for products, and there is no user product interaction information, the application tendency information includes user portrait information, and the recommendation model may include a model for recommending based on the probability prediction of product applications. For example, the recommendation model may include but is not limited to a click-through rate prediction (CTR) model, which predicts the probability of a certain content being selected by analyzing user behavior, historical interest data and other related features. The click-through rate prediction model may adopt but is not limited to the EulerNet model.

[0029] In step S102, based on the application tendency information, the preferred product information set and the large language model, the recommendation model is iteratively trained until the performance parameters of the recommendation model meet the preset conditions, and the recommendation model is determined as the target recommendation model.

[0030] The large language model is used to extract sample information from the preferred product information for iterative training of the recommendation model. The large language model can use the application tendency information as a basis, and use its own summary and reasoning ability to further extract product information from the preferred product information set to construct a pseudo sample for training the recommendation model. The steps of retraining the recommendation model with pseudo samples, the recommendation model outputting the preferred product information, and extracting pseudo samples from the preferred product information with the large language model can be performed cyclically until the performance parameters of the recommendation model meet the preset conditions. Retraining the recommendation model with pseudo samples will adjust the model parameters of the recommendation model accordingly to update the recommendation model. The preset conditions include the conditions that the recommendation model meets the usage standards. In some examples, the preset conditions may include that the difference between the performance parameters of the recommendation model after this training and the performance parameters of the recommendation model after the previous training is less than a preset threshold, that is, the training of the recommendation model until the performance parameters of the recommendation model after training no longer continue to significantly improve. When the performance parameters of the recommendation model meet the preset conditions, the current recommendation model is determined as the target recommendation model. The target recommendation model is a model that can be put into use.

[0031] In step S103, the target recommendation model is used to obtain recommended product information based on the application tendency information of the tested user for the product and push it to the tested user.

[0032] Input information can be obtained based on the application tendency information of the tested user for the product, and the input information is input into the target recommendation model. The result information output by the target recommendation model is the recommended product information. The recommended product information includes information about the products recommended by the target recommendation model for the tested user. The obtained recommended product information can be pushed to the tested user. The method of pushing the recommended product information to the tested user is not limited here. For example, the recommended product information can be pushed to the tested user through short messages, applications, etc.

[0033] In some examples, the target recommendation model may include a model for making recommendations based on the similarity of user behaviors and a model for making recommendations based on the probability prediction of product applications. If the user being tested has a history of applying for products, that is, the user being tested is an old user who has applied for products, then the application tendency information of the user being tested for products includes user-product interaction information, and the target recommendation model selected is a model for making recommendations based on the similarity of user behaviors. If the user being tested does not have a history of applying for products, that is, the user being tested is a new user who has not applied for products, then the application tendency information of the user being tested for products includes user portrait information, and the target recommendation model selected is a model for making recommendations based on the probability prediction of product applications.

[0034] In some embodiments, the sample user's application tendency information for commodities can be divided into two parts, one part as a training sample for model training, and the other part as a verification sample for model verification. When the performance parameters of the recommendation model meet the preset conditions, the verification sample can be used to verify the recommendation model to evaluate the training results of the model training. If the training results meet the preset requirements, the recommendation model is determined as the target recommendation model and put into use.

[0035] For ease of understanding, here we take the example of a user applying for a product as an example of a user applying for a credit card. If the sample user is a user who has opened a credit card, the application tendency information of the sample user may include the credit card benefit description information of the credit card that the sample user has opened. The recommendation model is trained based on the credit card benefit description information of the credit card that the sample user has opened, and the credit card benefit description information of the credit card that the sample user prefers is obtained, which is predicted by the recommendation model based on the credit card benefit description information of the credit card that the sample user has opened. Based on the credit card benefit description information of the credit card that the sample user has opened, the recommendation model is trained again using the large language model to extract information from the credit card benefit description information of the credit card that the sample user prefers predicted by the recommendation model as sample information until the performance parameters of the recommendation model meet the standards, and the recommendation model is put into use as the target recommendation model. The target recommendation model can push credit card recommendation information to the tested user based on the application tendency information of the tested user who has opened a credit card, so as to recommend a credit card to the tested user. Similarly, if the sample user is a user who has not opened a credit card, the application tendency information of the sample user may include the user portrait information of the sample user. The model training process of the recommendation model is similar to the model training process of the recommendation model related to the sample user being a user who has opened a credit card, and will not be repeated here. The trained target recommendation model can push credit card recommendation information to the tested user who has not opened a credit card, so as to recommend a credit card to the tested user.

[0036] The product information recommendation method in the embodiment of the present application can be used to recommend product information in a given user group. Product Collection And user-product interaction information ε + In the case of implicit feedback, implicit feedback is used to recommend product information and find out products that users may be interested in. test Product information. represents the Cartesian product, Represents the possible interaction relationship between users and products, minus the user-product interaction information ε + The content other than the above is the content that the embodiments of the present application hope to discover.

[0037] In an embodiment of the present application, a recommendation model is trained according to application tendency information that can reflect the application tendency of sample users for commodities, and a set of preferred commodity information of sample users predicted by the recommendation model is obtained. Based on the application tendency information, the inductive reasoning ability of the large language model can be used to further extract information from the set of preferred commodity information as sample information to train the recommendation model, until the performance parameters of the recommendation model meet the preset conditions, and the recommendation model is put into use as the target recommendation model to recommend commodity information to the tested user. In the iterative process of training the recommendation model with the output of the recommendation model as the input of the large language model and the output of the large language model as sample information, the synergy signal is continuously enhanced by using the output of the large language model as sample information to improve the problem of lack of synergy signal, enhance the recommendation model, and improve the accuracy of the recommendation model in recommending commodity information to users. Moreover, by gradually using the output of the large language model as sample information to iteratively train the recommendation model until the performance of the recommendation model is no longer significantly improved, the noise introduced into the model training can be reduced while enhancing the synergy signal, so that the enhancement of the synergy signal and the reduction of the introduction of noise can be balanced, maximizing the performance gain of the model training of the recommendation model.

[0038] In order to verify the effectiveness of the method for recommending product information in the embodiment of the present application, three different methods of credit card information recommendation were carried out using credit card application as an example, and the effects of the three different methods were compared. Method 1: The method for recommending product information in the embodiment of the present application is used to push the credit card information of 10 credit cards output by the target recommendation model to the user under test. Method 2: Professionals obtain a recommendation strategy based on experience summary, select the 10 card issuing institutions with the highest credit card application volume, and push the credit card information of the credit cards promoted by these 10 card issuing institutions to the user under test according to the card approval success rate of each card issuing institution. Method 3: The credit card information of 10 credit cards is randomly selected from all credit cards and pushed to the user under test. After pushing the credit card information to the user under test in accordance with the above three methods, the statistical data is shown in Table 1 below:

[0039] Table 1

[0040]

[0041]

[0042] Among them, old users are users who have applied for products in history, that is, users who have applied for credit cards, and new users are users who have not applied for products in history, that is, users who have not applied for credit cards. The exposure volume indicates the number of credit card information pushed to users. The number of applications indicates the number of users who apply for credit cards. The number of users who apply for recommended cards is the number of users who apply for credit cards corresponding to the pushed credit card information. In method 1, the ratio of the number of users who apply for recommended cards corresponding to old users to the number of applications is 46.51%, and the ratio of the number of users who apply for recommended cards corresponding to new users to the number of applications is 56.53%. In method 2, the ratio of the number of users who apply for recommended cards corresponding to old users to the number of applications is 37.57%, and the ratio of the number of users who apply for recommended cards corresponding to new users to the number of applications is 43.20%. It can be obtained from the above that the accuracy of recommending credit card information to users using method 1 of the product information recommendation method in this application is higher than that of method 2. In the scenario of old users, the recommendation accuracy of method 1 is 23.8% higher than that of method 2; in the scenario of new users, the recommendation accuracy of method 1 is 30.86% higher than that of method 2. It can be seen that the product information recommendation method provided in this application can improve the recommendation accuracy of the recommendation model.

[0043] In some embodiments, the process of extracting sample information from the preferred product information set using a large language model and training the recommendation model using the extracted sample information can be performed cyclically to gradually improve the performance of the recommendation model. Figure 2 This is a flowchart of a method for recommending product information provided by another embodiment of the present application. Figure 2 and Figure 1 The difference is that Figure 1 Step S102 in the above example can be specifically broken down into the following steps: Figure 2 Steps S1021 to S1023 in .

[0044] In step S1021, based on the application tendency information and the preferred product information set, the preferred product information and the rejected product information of the sample user are extracted from the preferred product information set using a large language model.

[0045] The application tendency information can be used as the basis for reasoning with the large language model, and the set of sample user's preferred product information predicted by the recommendation model can be used as a candidate set. The sample user's preferred product information and excluded product information can be extracted from the candidate set through the large language model. The sample user's preferred product information can represent the products that the large language model infers that the sample user is interested in in the set of preferred product information. In some examples, the sample user's preferred product information can represent one or more products that the large language model infers that the sample user is most interested in in the set of preferred product information. The sample user's excluded product information can represent the products that the large language model infers that the sample user is not interested in in the set of preferred product information. In some examples, the sample user's preferred product information can represent one or more products that the large language model infers that the sample user is least interested in in the set of preferred product information.

[0046] In some examples, the application tendency information of the sample user includes positive samples and negative samples, the positive samples represent the goods or types of goods that the sample user tends to apply for, and the negative samples represent the goods or types of goods that the sample user does not tend to apply for. In the process of the large language model using the application tendency information as a basis to infer the preferred product information and the excluded product information, the application tendency information used as a basis is the positive sample, that is, based on the positive sample and the preferred product information set, the preferred product information and the excluded product information are extracted from the preferred product information set using the large language model.

[0047] In some examples, the prompt word of the large language model can be set to enable the large language model to input the preferred product information and excluded product information that meet the requirements. Specifically, a first prompt word can be constructed based on the application tendency information and the preferred product information set, and the first prompt word includes extraction requirement information, application tendency information, and a preferred product information set; the first prompt word is input into the large language model to obtain the preferred product information and excluded product information output by the large language model. The extraction requirement information in the first prompt word is used to give the requirements for reasoning the large language model, the application tendency information in the first prompt word can be used as the basis for reasoning the large language model, and the preferred product information set in the first prompt word can be used as a candidate set for the large language model to extract the preferred product information and excluded product information.

[0048] For example, if the product is a credit card and the user has a history of applying for the product, the first prompt word input to the large language model and the output of the large language model can be as follows:

[0049]

[0050]

[0051] Among them, "Please select one credit card that the user is most likely to like and one credit card that the user is least likely to like from the candidate set based on the user's historical credit card application records. When making a decision, you should analyze and reason based on the credit card benefits information I provide. When outputting the results, only output the serial numbers in [] in the candidate set, and do not provide any other information such as analysis and explanation. The serial numbers of the most favorite card and the least favorite card are separated by a space" is the extraction requirement information in the first prompt word. The content of the historical application record is the application tendency information in the first prompt word, specifically the user product interaction information. The content of the candidate set is the set of preferred product information in the first prompt word. The large language model performs reasoning based on the first prompt word, and the output preferred product information is "495", indicating that the credit card indicated by

[495] in the candidate set is the credit card that the user is most interested in after reasoning by the large language model; the large language model performs reasoning based on the first prompt word, and the output excluded product information is "430", indicating that the credit card indicated by

[430] in the candidate set is the credit card that the user is least interested in after reasoning by the large language model. After obtaining the user preference information and user rejection information, you can input the prompt word "Why give this result" to get the analysis result or reply given by the large language model.

[0052] For another example, if the product is a credit card and the user has no history of applying for the product, the first prompt word input to the large language model and the output of the large language model may be as follows:

[0053]

[0054]

[0055] Among them, "Please select one credit card that the user is most likely to like and one credit card that the user is least likely to like from the candidate set based on the user portrait. When making a decision, you should analyze and reason based on the user portrait information I provide. When outputting the results, only output the serial number in [] in the candidate set, and do not provide any other information such as analysis and explanation. The serial numbers of the most favorite card and the least favorite card are separated by a space" is the extraction requirement information in the first prompt word. The content of the user portrait is the application tendency information in the first prompt word, specifically the user portrait information. The content of the candidate set is the set of preferred product information in the first prompt word. The large language model performs reasoning based on the first prompt word, and the output preferred product information is "2049", indicating that the credit card indicated by

[2049] in the candidate set is the credit card that the user is most interested in after reasoning by the large language model; the large language model performs reasoning based on the first prompt word, and the output excluded product information is "1897", indicating that the credit card indicated by

[1897] in the candidate set is the credit card that the user is least interested in after reasoning by the large language model. After obtaining the user preference information and user rejection information, you can input the prompt word "Why give this result" to get the analysis result or reply given by the large language model.

[0056] In step S1022, the recommendation model is trained using the preferred product information, the excluded product information and the application tendency information to obtain performance parameters of the trained recommendation model.

[0057] The application tendency information can be regarded as the original sample, the preferred product information and the excluded product information output by the large language model are the newly added samples, the preferred product information can be used as the newly added positive sample, and the excluded product information can be used as the newly added negative sample. The preferred product information, the excluded product information and the application tendency information are used as samples to train the recommendation model, adjust the model parameters of the recommendation model, and obtain the performance parameters of the trained recommendation model to determine the performance of the recommendation model.

[0058] In step S1023, it is determined whether the performance parameter meets the preset condition. If the performance parameter meets the preset condition, step S1024 is executed; if the performance parameter does not meet the preset condition, the process returns to step S2021.

[0059] In step S1024, the recommended model is determined as the target recommended model.

[0060] If the performance parameters meet the preset conditions, it means that the recommendation model has met the use requirements and can be put into use. If the performance parameters do not meet the preset conditions, it means that the recommendation model has not met the use requirements and needs to be further improved. Execute step S1021 again. The set of preferred product information in step S1021 is the set of preferred product information output by the trained recommendation model based on the preferred product information, excluded product information and application tendency information. Repeat the steps of extracting the preferred product information and excluded product information by the large language model, training the recommendation model, and obtaining the set of preferred product information output by the trained recommendation model until the performance parameters of the recommendation model meet the preset conditions and the recommendation model is determined as the target recommendation model. That is, if the performance parameters do not meet the preset conditions, the recommendation model can be trained again using the application tendency information, the preferred product information and the excluded product information extracted by the large language model from the set of preferred product information output by the trained recommendation model, until the performance parameters of the recommendation model meet the preset conditions and the recommendation model is determined as the target recommendation model.

[0061] Figure 3 A schematic diagram of an example of a logical framework of a large language model enhanced recommendation algorithm based on a course learning paradigm provided in an embodiment of the present application, such as Figure 3 As shown, the original user-product interaction information 21 can be used to train the recommendation model 22, and the recommendation model 22 predicts the preferred product information set 23, which can be regarded as the predicted user-product interaction information. The original user-product interaction information 21 and the preferred product information set 23 are input into the large language model 24, and the large language model 24 extracts the preferred product information and the excluded product information 25 from the preferred product information set 23. The preferred product information and the excluded product information 25 can be regarded as the user-product interaction information selected by the large language model 24, and the selected user-product interaction information can be regarded as the sample amplified by the large language model 24. The preferred product information and the excluded product information 25 can be used as samples to retrain the recommendation model.

[0062] Through the powerful indication reserve and understanding and analysis capabilities of the large language model, the original application tendency information is amplified. The large language model is oriented to natural language. The large language model can be placed in the back end, and the candidate set input into the large language model is reduced through the recommendation model, reducing the calculation amount of the large language model. The large language model is used to generate pseudo-labels, that is, pseudo-labels, to enhance the recommendation algorithm of the recommendation model. The pseudo-labels here are the preferred product information and excluded product information output as samples by the large language model in the above embodiment. The course learning method adopted in the embodiment of the present application is a machine learning training strategy, which trains the model by imitating the easy-to-difficult and step-by-step approach in human education. The set of preferred product information predicted by the recommendation model is the information of the products that the recommendation model believes the user is interested in. The excluded product information extracted from the set of preferred product information using the large model is a difficult sample, and the difficult sample is a sample that is difficult to distinguish by the recommendation model. The recommendation model can be first trained based on data that is easy to distinguish, that is, application tendency information, and gradually transferred to training the recommendation model using difficult data, that is, difficult samples, thereby improving the recommendation model's ability to distinguish products that the user is interested in.

[0063] In some embodiments, the application tendency information includes user product interaction information, and the user has a history of applying for products. In practice, the description information of historically applied products has problems such as lengthy narration, messy format, and low information density. If the description information of historically applied products is directly used to train the recommendation model, it is difficult to obtain ideal results. The refining and summarizing capabilities of the large language model can be used to refine the description information of historically applied products. For example, if the product is a credit card, the credit card's rights and interests information may contain a large amount of detailed information such as numbers and time, but for credit card applications, users are mainly concerned about the category to which the rights and interests belong. For example, the credit card's rights and interests information "Free video application A, video application B, video application C membership for 1 year (choose 1 from 3)" can be summarized as "video website membership". Specifically, a second prompt word can be generated, and the second prompt word includes the refinement requirement information and the description information of the historically applied products; the second prompt word is input into the large language model, and the large language model extracts the user product interaction information from the description information.

[0064] The second prompt word can guide the large language model to generate a high-quality summary based on the description information of the historical application product. For example, if the product is a credit card and the description information of the historical application product is the credit card's benefits information, the second prompt word can be as follows:

[0065]

[0066]

[0067] Among them, the rights and interests details are the description information of the historically applied products. The content in the reply is the user-product interaction information extracted by the large language model, and the user-product interaction information is the information obtained after being refined by the large language model.

[0068] The application tendency information may be text information. In order to facilitate model training, the text information may be converted into a vector. Specifically, the application tendency information may be converted into a first feature vector using a sentence vector extraction model; the first feature vector is subjected to dimensionality reduction processing to obtain a second feature vector; and the recommendation model is trained using the second feature vector. The application tendency information here may include user-product interaction information. The sentence vector extraction model may be used to convert text information into a vector, and the sentence vector extraction model may include but is not limited to the Sentence Transformer model. The first feature vector is a feature vector obtained by the sentence vector extraction model for the application tendency information. However, the first feature vector is a high-dimensional vector with dense representation, so dimensionality reduction processing is required. The dimensionality reduction processing can be achieved by a dimensionality reduction algorithm, and the dimensionality reduction algorithm may include but is not limited to the UMAP (full name Uniform Manifold Approximation and Projection) algorithm. The second feature vector is a feature vector obtained after dimensionality reduction processing, which is a low-dimensional embedding vector. The second feature vector can be used as a feature vector of the application tendency information.

[0069] Figure 4 A schematic diagram of an example of an embedded representation of user-product interaction information provided in an embodiment of the present application, such as Figure 4 As shown, the original credit card benefit description text 31 is input into the large language model 32, and the large language model 32 can output the refined user-product interaction information, i.e., the refined credit card benefit description text 33. The refined credit card benefit description text is processed through the sentence vector extraction model 34 to obtain the first feature vector. The first feature vector is processed through dimensionality reduction 35 to obtain the second feature vector, i.e., the embedded representation 36 of the credit card benefit description text.

[0070] A second aspect of the present application provides a device for recommending product information. Figure 5 A schematic diagram of the structure of a device for recommending product information provided by an embodiment of the present application, such as Figure 5 As shown, the product information recommendation device 400 may include a prediction result acquisition module 401 , an enhanced training module 402 and an information push module 403 .

[0071] The prediction result acquisition module 401 may be used to train the recommendation model according to the sample user's application tendency information for commodities, and obtain a set of the sample user's preferred commodity information predicted by the recommendation model.

[0072] The enhanced training module 402 can be used to iteratively train the recommendation model based on the application tendency information, the preferred product information set and the large language model until the performance parameters of the recommendation model meet the preset conditions and the recommendation model is determined as the target recommendation model. The large language model is used to extract sample information for iterative training of the recommendation model from the preferred product information set.

[0073] The information push module 403 may be used to obtain recommended product information and push it to the user under test based on the user under test's application tendency information for the product using the target recommendation model.

[0074] In some embodiments, the enhanced training module 402 can be specifically used to: extract the sample user's preferred product information and excluded product information from the preferred product information set using a large language model based on the application tendency information and the preferred product information set; train the recommendation model using the preferred product information, excluded product information and application tendency information to obtain the performance parameters of the trained recommendation model; if the performance parameters meet the preset conditions, determine the recommendation model as the target recommendation model; if the performance parameters do not meet the preset conditions, train the recommendation model again using the application tendency information, the preferred product information and excluded product information extracted from the preferred product information set output by the trained recommendation model using the large language model, until the performance parameters of the recommendation model meet the preset conditions, and determine the recommendation model as the target recommendation model.

[0075] In some examples, the application tendency information includes positive samples and negative samples. The enhanced training module 402 can be specifically used to extract preferred product information and excluded product information from the preferred product information set using a large language model according to the positive samples and the preferred product information set.

[0076] In some examples, the enhanced training module 402 can be specifically used to: construct a first prompt word based on the application tendency information and the preferred product information set, the first prompt word including extraction requirement information, application tendency information and the preferred product information set; input the first prompt word into the large language model to obtain the preferred product information and excluded product information output by the large language model.

[0077] In some embodiments, when the sample user has a history of applying for products, the application tendency information includes user product interaction information, the user product interaction information represents the user's historical application for products, and the recommendation model includes a model for making recommendations based on similarity of user behaviors.

[0078] In the case where the sample user has no history of applying for products, the application tendency information includes user portrait information, and the recommendation model includes a model for making recommendations based on probability predictions of product applications.

[0079] In some embodiments, the application tendency information includes user-product interaction information. The product information recommendation device 400 may also include a refinement module. The refinement module may be used to: generate a second prompt word, the second prompt word includes refinement requirement information and description information of historical application products; input the second prompt word into the large language model, and obtain the user-product interaction information extracted by the large language model from the description information.

[0080] In some embodiments, the prediction result acquisition module 401 can be specifically used to: use a sentence vector extraction model to convert application tendency information into a first feature vector; perform dimensionality reduction processing on the first feature vector to obtain a second feature vector; and use the second feature vector to train a recommendation model.

[0081] It should be noted that the product information recommendation device 400 is a device corresponding to the product information recommendation method described above. All implementation methods in the above method embodiments are applicable to the embodiments of the device and can achieve the same technical effects, which will not be described in detail here.

[0082] The third aspect of the present application also provides an electronic device. Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application is shown in FIG. Figure 6 As shown, the electronic device 500 includes a memory 501 , a processor 502 , and a computer program stored in the memory 501 and executable on the processor 502 .

[0083] In some examples, the processor 502 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0084] The memory 501 may include a read-only memory (ROM), a random access memory (RAM), a disk storage medium device, an optical storage medium device, a flash memory device, an electrical, optical or other physical / tangible memory storage device. Therefore, generally, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., a memory device) encoded with software including computer executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method for recommending product information in the embodiment of the present application.

[0085] The processor 502 runs a computer program corresponding to the executable program code by reading the executable program code stored in the memory 501 , so as to implement the method for recommending product information in the above embodiment.

[0086] In some examples, the electronic device 500 may further include a communication interface 503 and a bus 504. Figure 6 As shown, the memory 501, the processor 502, and the communication interface 503 are connected via a bus 504 and communicate with each other.

[0087] The communication interface 503 is mainly used to realize the communication between the modules, devices, units and / or equipment in the embodiment of the present application. The communication interface 503 can also be used to access input devices and / or output devices.

[0088] The bus 504 includes hardware, software, or both, coupling the components of the electronic device 500 to each other. By way of example and not limitation, the bus 504 may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a Memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-E) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local Bus (VLB) bus, or other suitable buses or a combination of two or more of these. Where appropriate, the bus 504 may include one or more buses. Although embodiments of the present application describe and illustrate a particular bus, the present application contemplates any suitable bus or interconnect.

[0089] The fourth aspect of the present application provides a computer-readable storage medium, on which a computer program instruction is stored, and when the computer program instruction is executed by a processor, the method for recommending product information in the above-mentioned embodiment can be implemented, and the same technical effect can be achieved. To avoid repetition, it is not repeated here. Among them, the above-mentioned computer-readable storage medium may include a non-transitory computer-readable storage medium, such as a read-only memory (Read-Only Memory, referred to as ROM), a random access memory (Random Access Memory, referred to as RAM), a disk or an optical disk, etc., which is not limited here.

[0090] The fifth aspect of the present application provides a computer program product, which includes a computer program. When the computer program is executed by a processor, it implements the method for recommending product information in the above-mentioned embodiment and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0091] It should be clear that each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. For device embodiments, equipment embodiments, computer-readable storage medium embodiments, and computer program product embodiments, the relevant parts can refer to the description part of the method embodiment. The present application is not limited to the specific steps and structures described above and shown in the figures. Those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present application. In addition, for the sake of brevity, a detailed description of known method technologies is omitted here.

[0092] The above reference is according to the method of the embodiment of the present application, the flow chart of the device (system) and the computer program product and / or the block diagram described various aspects of the present application.It should be understood that each square box in the flow chart and / or the block diagram and the combination of each square box in the flow chart and / or the block diagram can be realized by computer program instructions.These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device to produce a machine so that these instructions executed by the processor of the computer or other programmable data processing device enable the realization of the function / action specified in one or more square boxes of the flow chart and / or the block diagram.Such a processor can be but is not limited to a general-purpose processor, a special-purpose processor, a special application processor or a field programmable logic circuit.It can also be understood that each square box in the block diagram and / or the flow chart and the combination of the square boxes in the block diagram and / or the flow chart can also be realized by the dedicated hardware that performs the specified function or action, or can be realized by the combination of dedicated hardware and computer instructions.

[0093] Those skilled in the art should understand that the above embodiments are exemplary rather than restrictive. Different technical features appearing in different embodiments can be combined to achieve beneficial effects. Based on the study of the drawings, the specification and the claims, those skilled in the art should be able to understand and implement other variations of the disclosed embodiments. In the claims, the term "comprising" does not exclude other devices or steps; the quantifier "one" does not exclude multiple; the terms "first" and "second" are used to indicate names rather than to indicate any specific order. Any figure marks in the claims should not be understood as limiting the scope of protection. The functions of multiple parts appearing in the claims can be implemented by a single hardware or software module. The appearance of certain technical features in different dependent claims does not mean that these technical features cannot be combined to achieve beneficial effects.

Claims

1. A method for recommending product information, characterized in that: include: The recommendation model is trained according to the sample user's application tendency information for commodities, and a set of the sample user's preferred commodity information predicted by the recommendation model is obtained; Based on the application tendency information, the preferred product information set and the large language model, iteratively train the recommendation model until the performance parameter of the recommendation model meets the preset conditions, and determine the recommendation model as the target recommendation model, and the large language model is used to extract sample information for iteratively training the recommendation model from the preferred product information set; The target recommendation model is used to obtain recommended product information based on the application tendency information of the tested user for the product and push it to the tested user.

2. The method according to claim 1, characterized in that The iterative training of the recommendation model based on the application tendency information, the preferred product information set and the large language model until the performance parameter of the recommendation model meets the preset conditions, and determining the recommendation model as the target recommendation model includes: According to the application tendency information and the preferred product information set, using the large language model to extract the preferred product information and the excluded product information of the sample user from the preferred product information set; Using the preferred product information, the excluded product information and the application tendency information to train the recommendation model, and obtain performance parameters of the trained recommendation model; If the performance parameter meets the preset conditions, the recommendation model is determined as the target recommendation model; If the performance parameters do not meet the preset conditions, the recommendation model is trained again using the application tendency information, the preferred product information and the excluded product information extracted by the large language model from the preferred product information set output by the trained recommendation model, until the performance parameters of the recommendation model meet the preset conditions, and the recommendation model is determined as the target recommendation model.

3. The method according to claim 2, characterized in that The application tendency information includes positive samples and negative samples; The step of extracting the sample user's preferred product information and the excluded product information from the preferred product information set using the large language model according to the application tendency information and the preferred product information set includes: According to the positive samples and the preferred product information set, the preferred product information and the excluded product information are extracted from the preferred product information set using the large language model.

4. The method according to claim 2, characterized in that: According to the application tendency information and the preferred product information set, the preferred product information and the excluded product information of the sample user are extracted from the preferred product information set using the large language model, including: constructing a first prompt word according to the application tendency information and the preferred product information set, wherein the first prompt word includes extraction requirement information, the application tendency information and the preferred product information set; The first prompt word is input into the large language model to obtain the preferred product information and the excluded product information output by the large language model.

5. The method according to claim 1, characterized in that In the case where the sample user has a history of applying for commodities, the application tendency information includes user commodity interaction information, the user commodity interaction information represents the user's historical application for commodities, and the recommendation model includes a model for making recommendations based on the similarity of user behaviors; In the case that the sample user has no history of applying for products, the application tendency information includes user portrait information, and the recommendation model includes a model for making recommendations based on probability prediction of product applications.

6. The method according to claim 5, characterized in that The application tendency information includes the user-product interaction information; The method further comprises: generating a second prompt word, wherein the second prompt word includes the refinement requirement information and the description information of the historical application product; The second prompt word is input into the large language model, so that the large language model extracts the user-product interaction information from the description information.

7. The method according to claim 1, characterized in that The training of the recommendation model according to the sample user's application tendency information for the product includes: Using a sentence vector extraction model to convert the application tendency information into a first feature vector; Performing dimensionality reduction processing on the first eigenvector to obtain a second eigenvector; The recommendation model is trained using the second feature vector.

8. A device for recommending product information, characterized in that: include: A prediction result acquisition module is used to train the recommendation model according to the sample user's application tendency information for commodities, and obtain the set of preferred commodity information of the sample user predicted by the recommendation model; An enhanced training module, used to iteratively train the recommendation model based on the application tendency information, the preferred product information set and the large language model until the performance parameters of the recommendation model meet the preset conditions, and determine the recommendation model as the target recommendation model, and the large language model is used to extract sample information for iteratively training the recommendation model from the preferred product information set; The information push module is used to use the target recommendation model to obtain recommended product information based on the application tendency information of the tested user for the product and push it to the tested user.

9. An electronic device, characterized in that: include: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, the method for recommending product information according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by a processor, the method for recommending product information according to any one of claims 1 to 7 is implemented.

11. A computer program product, characterized in that It comprises a computer program, which, when executed by a processor, implements the method for recommending product information as described in any one of claims 1 to 7.

Images