Product recommendation method and device based on intelligent agent and medium

Abstract

The invention discloses a product recommendation method and device based on an intelligent agent and a medium, and belongs to the technical field of artificial intelligence, and the method comprises the steps: obtaining a user portrait of a target user in response to a product recommendation request for the target user; if the target user is a new user, the user portrait is an initial portrait constructed by performing question and answer interaction with the target user by using an interaction agent in the large model agent; based on an analysis agent in the large model agent, determining a matching degree between the user portrait vector of the target user and each product vector in a product vector library of the to-be-recommended product set, and generating a product recommendation result for the target user according to the determined matching degree; under the condition of insufficient user data, the analysis of the recommendation capability of the agent is formed by simulating feedback of different users on recommendation results by using a large model and training based on simulated data. The product recommendation technology is optimized, and the performance of the recommendation system and the user experience are improved.

Description

Technical Field

[0001] The present application belongs to the field of artificial intelligence technology, and in particular, relates to a product recommendation method, device and medium based on an intelligent agent. Background Art

[0002] Product recommendation is a core technology in the field of artificial intelligence. It uses comprehensive data analysis to intelligently predict and recommend content or items that users may be interested in. This field brings together cutting-edge research results from multiple disciplines such as machine learning, data mining, and social network analysis. It not only greatly improves the efficiency of personalized information delivery and optimizes user experience, but also creates significant economic benefits for content providers and e-commerce platforms.

[0003] The development of product recommendation technology has gone from simple content-based recommendation and collaborative filtering algorithms to complex systems that integrate advanced technologies such as deep learning, graph neural networks, and reinforcement learning. In particular, the application of deep learning technology enables recommendation systems to automatically learn and extract features from a large amount of unstructured data, greatly improving the accuracy and intelligence of recommendations. The introduction of deep learning technology, especially the application of large models, has brought new breakthroughs to recommendation systems.

[0004] However, there are still many problems with product recommendation technology in the known technologies, mainly including the cold start problem of the recommendation system when data is scarce in the initial stage, the difficulty of the recommendation system to efficiently process large-scale data sets, the lack of personalized recommendations and persuasiveness, and poor adaptability to dynamic changes in user behavior. How to solve at least some of these technical problems in order to optimize the recommendation technology has become a research hotspot in this field. Summary of the invention

[0005] In view of this, the present application provides a product recommendation method, device and medium based on an intelligent agent to solve at least some technical problems of product recommendation technology in the known technology, optimize the recommendation technology, and improve the performance of the recommendation system and user experience.

[0006] The specific technical solutions are as follows:

[0007] An agent-based product recommendation method, comprising:

[0008] In response to obtaining a product recommendation request for a target user, a user portrait of the target user is obtained; if the target user is a new user, the user portrait of the target user is an initial portrait constructed by using a large model based on a question-and-answer interaction with the target user by using an interactive agent in a pre-set large model agent;

[0009] Based on the analysis agent in the large model agent, determining the matching degree between the user portrait vector of the target user and each product vector in the product vector library of the product set to be recommended; the user portrait vector is a vector corresponding to the user portrait of the target user;

[0010] Based on the analysis agent, generating a product recommendation result for the target user according to the matching degree between the user portrait vector and each product vector;

[0011] In the case of insufficient user data, the recommendation capability of the analysis agent is formed by using a large model to simulate the feedback of different users on the recommendation results, and training the recommendation capability based on the simulated data.

[0012] Optionally, if the target user is not a new user, the user portrait of the target user is an updated portrait obtained by updating the initial portrait based on the information data generated by the target user for the product, or is a portrait of a user of the category to which the target user belongs; the process of constructing the initial portrait of the target user includes:

[0013] Collecting basic information and preference information of the target user through question-answer interaction between the interactive agent and the target user;

[0014] Inputting the basic information and preference information of the target user and the guidance information for instructing the large model to construct a user portrait into the large model, and obtaining an initial portrait of the target user output by the large model;

[0015] The updating of the initial portrait based on the information data generated by the target user for the product includes:

[0016] The initial portrait is updated based on the target user's historical interaction data and / or behavior pattern with the product.

[0017] Optionally, the process of constructing the product vector library includes:

[0018] Determine each key feature of the product to be recommended according to the product information of the product to be recommended in the product set to be recommended; the key feature of the product to be recommended is a feature selected from the product information of the product to be recommended, or a feature obtained by processing the selected feature;

[0019] Performing vectorization processing on each key feature of the product to be recommended, respectively, to obtain feature vectors corresponding to each key feature;

[0020] Merge the feature vectors corresponding to the key features into one vector to obtain a product vector corresponding to the product to be recommended;

[0021] The product vector library includes product vectors corresponding to each product to be recommended in the product set to be recommended; each product vector is mapped to the corresponding product to be recommended through a corresponding index in a preset index structure.

[0022] Optionally, the user portrait vector of the target user is a vector corresponding to the initial portrait of the target user, or a comprehensive vector constructed by integrating the historical interaction data and / or behavior pattern of the target user on the product with the initial portrait, or a vector corresponding to the portrait of the user of the category to which the target user belongs;

[0023] The determining of the matching degree between the user portrait vector of the target user and each product vector in the product vector library of the product set to be recommended includes:

[0024] quantifying the degree of match between the user portrait vector of the target user and each product vector in the product vector library using a predefined matching function, and obtaining the degree of match between the user portrait vector of the target user and each product vector in the product vector library;

[0025] Among them, the matching function includes: the matching degree between the user portrait vector and the product vector in the product vector library, and the functional relationship between the multidimensional vector representing user characteristics contained in the user portrait vector and the multidimensional vector representing product characteristics contained in the product vector.

[0026] Optionally, generating a product recommendation result for the target user according to the matching degree between the user portrait vector and each product vector includes:

[0027] Determine, from the set of products to be recommended, at least one product whose matching degree between a corresponding product vector and a user portrait vector of the target user satisfies a matching degree condition;

[0028] Generate a product recommendation result including the product information of the at least one product; or, select at least one target product from the at least one product that has a correlation with the target user's historical interaction data on the product, and generate a product recommendation result including the product information of the at least one target product.

[0029] Optionally, the agent-based product recommendation method further includes:

[0030] Based on the user portrait of the target user, the matching between the user portrait vector and the product vectors, and the semantic understanding of at least part of the historical interaction data of the target user on the product, recommendation reason explanation information for the product recommendation result is generated.

[0031] Optionally, the agent-based product recommendation method further includes:

[0032] Obtaining feedback information from the target user on the product recommendation result;

[0033] The model parameters of the recommendation model are adjusted according to the feedback information; the recommendation model is a model constructed according to the recommendation process of the analysis agent and used to implement the recommendation function of the analysis agent.

[0034] Optionally, the recommendation model is constructed by determining model parameters based on a predefined optimization problem;

[0035] The goal of the optimization problem includes: determining a set of model parameters for the recommendation model while satisfying computing resource constraints so as to minimize the weighted sum of the prediction error and model complexity of the recommendation model.

[0036] A product recommendation device based on an intelligent agent, comprising:

[0037] An acquisition module, configured to acquire a user portrait of the target user in response to obtaining a product recommendation request for the target user; if the target user is a new user, the user portrait of the target user is an initial portrait constructed by using the big model based on the question-and-answer interaction by using an interactive agent in a pre-set big model agent to conduct a question-and-answer interaction with the target user;

[0038] A determination module, configured to determine, based on the analysis agent in the large model agent, a degree of match between a user portrait vector of the target user and each product vector in a product vector library of a set of products to be recommended; the user portrait vector is a vector corresponding to the user portrait of the target user;

[0039] A generating module, configured to generate a product recommendation result for the target user based on the analysis agent and according to the matching degree between the user portrait vector and each product vector;

[0040] In the case of insufficient user data, the recommendation capability of the analysis agent is formed by using a large model to simulate the feedback of different users on the recommendation results, and training the recommendation capability based on the simulated data.

[0041] A computer-readable medium having a computer program stored thereon, which, when executed by a processor, can be used to implement the agent-based product recommendation method as described in any one of the above.

[0042] According to the above scheme, the agent-based product recommendation method, device and medium provided by the present application pre-set a large model agent including an interactive agent and an analysis agent for the product recommendation scenario. When a new user joins, the interactive agent in the large model agent is used to interact with the new user through question-and-answer interaction to support the large model to generate a preliminary user portrait for the new user, that is, the initial portrait, to help the system make more accurate recommendations even without a large amount of personal data. At the same time, in the early stage when user data is insufficient, the large model is used to simulate the feedback of different users on the recommendation results, and these simulated data are used to train the recommendation ability of the analysis agent, so that the analysis agent can better understand the preferences of different types of users before actually accumulating enough user data. Through the above scheme, the present application can effectively solve the cold start problem of the recommendation system in the early stage when new user data is scarce, thereby optimizing the product recommendation technology and improving the performance and user experience of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0044] Figure 1 It is a flow chart of the agent-based product recommendation method provided by this application;

[0045] Figure 2 This is a large model agent framework diagram of the recommendation system provided by this application;

[0046] Figure 3 It is a schematic diagram of the construction process of the product vector library provided by this application;

[0047] Figure 4 is an exemplary implementation flow chart of building a product vector library provided by this application;

[0048] Figure 5 is another flow chart of the agent-based product recommendation method provided in this application;

[0049] Figure 6 It is another flow chart of the agent-based product recommendation method provided by the present application;

[0050] Figure 7 This is a structural diagram of the composition of the intelligent agent-based product recommendation device provided in this application. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0052] The problems of product recommendation technology in the known technologies mainly include the cold start problem of the recommendation system in the case of lack of data in the initial stage, the difficulty of the recommendation system to efficiently process large-scale data sets, the lack of personalized recommendations and persuasiveness, and the poor adaptability to dynamic changes in user behavior. The applicant further analyzes each problem in detail as follows:

[0053] 1. Cold start problem

[0054] The cold start problem refers to how to provide effective recommendations for new users or new items in the absence of sufficient historical data. For new users, the recommendation system needs to quickly mine the user's potential interests and preferences from limited information. Similarly, for newly added items (such as new items, new content), due to the lack of user interaction data, the system cannot evaluate its popularity or match with existing users. The difficulty of this problem lies in how to quickly establish an effective recommendation model for new users or new items in the absence of historical data.

[0055] 2. Handling Large-Scale Datasets

[0056] With the popularity of the Internet and mobile devices, the amount of data that recommendation systems need to process has exploded. As the number of users and projects increases, the dimensionality and complexity of the interactive data also increase. This not only requires the recommendation system to have efficient data processing capabilities, but also to be able to extract valuable information from massive data, which poses a huge challenge to the storage, processing and computing capabilities of the recommendation system. How to efficiently process and analyze large-scale data sets while ensuring the quality of recommendations is a technical problem that must be faced in the design of recommendation systems.

[0057] 3. Lack of personalized recommendations and persuasiveness

[0058] The ultimate goal of a recommendation system is to provide personalized recommendations to improve user satisfaction and engagement. Personalized recommendations must not only be accurate, but also be able to explain the reasons for the recommendations to users in order to enhance their trust in the system. However, if the recommendation system only provides a list of recommendations without explaining why these items are recommended, users may be confused or distrustful of the recommendations. Unconvincing recommendations not only reduce user satisfaction, but may also affect the overall effectiveness of the recommendation system. Therefore, how to improve the transparency and credibility of recommendation results is a key issue in improving the performance of recommendation systems.

[0059] 4. Capturing User Dynamic Preferences

[0060] User preferences are not fixed, but often change over time, with changes in life stages, social environment, and other factors. This dynamic preference requires the recommendation system to be able to capture and adapt to these changes in real time or near real time. Traditional recommendation systems are often trained based on static data sets, which makes it difficult to cope with the rapid changes in user preferences, resulting in outdated or irrelevant recommendation results. Therefore, how to design a recommendation algorithm that can respond to changes in user preferences in a timely manner is a problem that needs to be analyzed and solved in depth.

[0061] In summary, these four problems respectively involve the lack of data in the initial stage of recommendation technology, the processing capacity of large-scale data, the personalization and persuasiveness of recommendation results, and the adaptability to dynamic changes in user behavior / preferences. Solving these problems is crucial to improving the performance and user experience of the recommendation system. Based on this, the embodiments of the present application provide a product recommendation method, device and medium based on an intelligent agent to solve at least some of the technical problems in the above series of problems.

[0062] See also Figure 1 As shown in the method flow chart, the agent-based product recommendation method provided in the embodiment of the present application may include at least the following steps 101 to 103, which are described in detail below.

[0063] Step 101, in response to obtaining a product recommendation request for a target user, obtaining a user portrait of the target user; if the target user is a new user, the user portrait of the target user is an initial portrait constructed by using an interactive agent in a pre-set big model agent to conduct a question-and-answer interaction with the target user and using the big model based on the question-and-answer interaction.

[0064] The product recommendation in the embodiments of the present application may include, but is not limited to, recommending content (such as movies, music), goods / articles and / or services to users, which may depend on the actual application scenario.

[0065] When building a recommendation system to provide product recommendations to users, refer to Figure 2The big model agent framework diagram of the recommendation system in the embodiment of the present application establishes a big model agent framework for the recommendation system. The core of the framework is composed of two key components: an interactive agent and an analytical agent. The two agents are based on a common basic big model, but each has different roles and functions. Among them, the role and function of the interactive agent mainly revolves around question-and-answer interaction with the user, and enables the big model to build a user portrait for the user based on the question-and-answer interaction results. The role and function of the analytical agent are mainly based on recommending products to users based on user portraits and product information, and combined with the semantic understanding of the big model to generate recommendation reason explanation information for the recommendation results. The two agents can specifically maintain different prompt word templates for the common basic big model respectively, and the prompt word templates maintained by the two agents are respectively used to specify the roles required for the big model and guide the big model to call the role functions required by each. In this way, the performance of the recommendation system can be optimized while ensuring the effective use of computing resources.

[0066] Based on the big model agent framework, when a new user joins, the interactive agent can conduct a few simple question-and-answer interactions with the user, allowing the big model to generate a preliminary user profile based on the user's answers. The questions involved in the question-and-answer interaction can be designed around the user's preferred content types, daily activities, etc. Based on this information, combined with the pre-trained knowledge base of the big model, it can predict the areas or topics that the user may be interested in, helping the system to make more accurate recommendations to new users without a large amount of personal data. More specifically, the interactive agent can construct the input data for the big model based on the user's answer information, using the pre-stored prompt word template used to guide the big model to generate the user profile, and input it into the big model to obtain the user profile data output by the big model.

[0067] The above process can be dynamic. During implementation, multiple rounds of questions and answers based on a quick question-and-answer method can be set up under the interactive agent. As the user interacts with the interactive agent, new user data is continuously collected and the user portrait is updated in real time to ensure the timeliness and accuracy of the initial recommendation to new users.

[0068] Subsequently, the interactive agent can also obtain information data generated by users regarding products to update the user's initial profile, so as to continuously adjust and improve the user profile and support more accurate product recommendations for users.

[0069] The information data generated by users for products can be further subdivided into, but not limited to, historical interaction data and / or behavior patterns between users and products. The historical interaction data may include, but is not limited to, products booked by users, products unbooked, products of no interest, products browsed / clicked / purchased, etc.; behavior patterns generally refer to the behavioral sequence of users' interactions with products. These behaviors may be explicit, such as ratings and comments, or implicit, such as browsing, clicking, purchasing, etc.

[0070] After obtaining the above information data generated by users for the product, you can optionally first perform preprocessing such as deleting duplicate data, filling missing data, and correcting erroneous data. On this basis, extract useful features from the preprocessed data (useful features can be determined manually based on business understanding and experience, or by machines based on set judgment logic) for use in establishing / adjusting user portraits.

[0071] In addition, optionally, users with similar characteristics can be grouped into the same category, and each category can be regarded as having a user portrait. In this implementation, each user portrait is used to describe the typical behavior and attributes of a class of users, and correspondingly reflects the behavior patterns and preferences of a class of users. That is to say, in implementation, for non-new users (i.e., users who have collected enough data), one class of users can correspond to one portrait, or one user can correspond to one portrait. For new users, an interactive agent conducts question-and-answer interactions with new users to generate an initial portrait for the new user. In addition, users' interests and behavior patterns usually change over time, so user portraits can be updated regularly to reflect these changes. In actual applications, new user portraits can be generated by re-executing the user portrait generation process, or user portraits can be updated in real time by designing a dynamic adjustment mechanism for user portraits (such as initial portraits), and there is no restriction on this.

[0072] Based on this, in this step 101, when responding to obtaining a product recommendation request for a target user and thereby obtaining a user portrait of the target user, if the target user is a new user, the user portrait obtained is an initial portrait constructed by using the big model based on the question-and-answer interaction by using the interactive agent in the big model agent to conduct question-and-answer interaction with the target user; if the target user is not a new user, the user portrait of the target user may be an updated portrait obtained by updating the initial portrait of the target user based on the information data generated by the target user for the product, or may be a portrait of a user of the category to which the target user belongs.

[0073] In the case where the target user is a new user, when constructing an initial profile for the target user, the basic information and preference information of the target user can be first collected through question-and-answer interaction between the interactive agent and the target user; then, the basic information and preference information of the target user and the guidance information for instructing the large model to construct the user profile are input into the large model to obtain the initial profile of the target user output by the large model. Subsequently, the interactive agent can also obtain information data generated by the target user for the product, such as historical interaction data and / or behavior patterns of the product, and dynamically update the initial profile of the target user based on this information data.

[0074] During implementation, the analysis agent may specifically obtain a user profile of a target user in response to obtaining a product recommendation request for the target user.

[0075] Step 102: Based on the analysis agent in the large model agent, determine the degree of match between the user portrait vector of the target user and each product vector in the product vector library of the product set to be recommended; the user portrait vector is a vector corresponding to the user portrait of the target user.

[0076] The product vector library includes a product vector corresponding to each product to be recommended in the product set to be recommended. The product vector corresponding to the product to be recommended is specifically a vector obtained by vectorizing the product information of the product to be recommended.

[0077] The product information may include but is not limited to product description, specifications, evaluation and other information.

[0078] After obtaining the user portrait of the target user, the analysis agent can further convert the user portrait of the target user into a vector form to obtain the user portrait vector of the target user, and determine the matching degree between the user portrait vector of the target user and each product vector in the product vector library of the product set to be recommended. For example, the matching degree between the user portrait vector and the product vector can be evaluated by calculating the similarity between the user portrait vector and the product vector, etc. Subsequently, the product recommendation results for the target user can be further generated based on the matching degree between the user portrait vector and each product vector.

[0079] It is easy to understand that the user portrait vector of the target user is a vector corresponding to the initial portrait of the target user, or a comprehensive vector constructed by integrating the target user's historical interaction data and / or behavior pattern with the target user's initial portrait, or a vector corresponding to the portrait of the user category to which the target user belongs, depending on the actual application situation.

[0080] In the embodiment of the present application, the user portrait vector (U) is defined as a multidimensional vector of user characteristics: (U(u1,u2,...un)), where each dimension (ui) can represent different attributes of the user, such as industry categories, key categories, purchase history, browsing behavior, etc. The product vector (P) is defined as a multidimensional vector of product characteristics: (P(p1,p2,...pn)), where each dimension (pj) represents different characteristics of the product, such as price, category, user rating, etc. 1≤i, j≤n, and i, j, n are integers.

[0081] Optionally, you can first use a data normalization method, such as the Z-score normalization method, to normalize the user portrait vector and the product vector to ensure that the user portrait vector and the product vector are compared on the same scale to avoid certain features having a disproportionate impact on the results due to differences in dimensions or orders of magnitude.

[0082] Then, the analysis agent calculates the matching degree between the user portrait vector (U) and each product vector (P) in the product vector library of the product set to be recommended, for example, by calculating the cosine similarity between the two to measure the matching degree between the two.

[0083] Exemplarily, this embodiment defines a matching function M(U, P) to quantify the matching degree between the user portrait vector and the product vector, and obtains the matching degree between the user portrait vector and each product vector in the product vector library.

[0084] Among them, the matching function includes: the matching degree between the user portrait vector and the product vector in the product vector library, and the functional relationship between the multidimensional vector representing user characteristics contained in the user portrait vector and the multidimensional vector representing product characteristics contained in the product vector.

[0085] The matching function can be expressed as, but not limited to:

[0086]

[0087] in, represents the dot product of vectors, Represents the norm of a vector.

[0088] Optionally, based on the consideration of feature weights, a feature weight vector ( W = (w_1, w_2, ..., w_m) ) can be further introduced into the matching function, where ( w_j ) represents the importance of each feature in the product vector, 1≤j≤m, and j and m are integers. These weights can be predicted or adjusted by machine learning models to capture the user's implicit preferences. That is, when modeling user preferences, machine learning models (such as logistic regression, SVM or neural network) can be used to predict or adjust the weights in W to better reflect the user's implicit preferences.

[0089] After integrating the weight vector into the matching function, the following matching function can be further obtained:

[0090]

[0091] If there are multiple matching scores between the same pair of user portrait vector and product vector (for example, matching scores are calculated based on different categories or dimensions of users / products, or based on different calculation methods), the multiple matching scores are merged into a final matching score. The specific fusion method can be, but is not limited to, the weighted average method: , where M k is the kth matching score, α k is the corresponding fusion weight. This method allows to consider multiple factors while adjusting the relative importance of each factor according to the actual situation.

[0092] Step 103: Based on the analysis agent, generate a product recommendation result for the target user according to the matching degree between the user portrait vector and the product vectors.

[0093] In the case of insufficient user data, the recommendation capability of the analysis agent is formed by using a large model to simulate the feedback of different users on the recommendation results, and training the recommendation capability based on the simulated data.

[0094] That is to say, in the early stage when user data is insufficient, the interaction agent's base model (i.e., the common model of the interaction agent and the analysis agent) can be used to simulate different users' feedback on recommended content. Use these simulated data to train the recommendation capabilities / recommendation algorithms in the analysis agent, so that the analysis agent can better understand different types of user preferences before actually accumulating enough user data.

[0095] Specifically, we can set different personalities for the big model, let it answer interactive questions designed around "different users' feedback on recommended content", and collect the feedback generated by the big model based on different personalities, so as to simulate the feedback of different users on recommended content through the simulated behavior of the big model. On this basis, these simulated data are used to train the recommendation ability / recommendation algorithm of the analysis agent.

[0096] On the basis of step 102, in this step 103, the analysis agent may further generate product recommendation results for the target user according to the matching degree between the user portrait vector of the target user and each product vector.

[0097] Among them, the analysis agent can first determine at least one product from the set of products to be recommended, whose matching degree between the corresponding product vector and the user portrait vector of the target user meets the matching condition; on this basis, generate a product recommendation result containing the product information of the at least one product.

[0098] The matching condition may be, but is not limited to, set to: the matching degree between the corresponding product vector and the user portrait vector of the target user is the top_k (k ≥ 1, and k is an integer) of the matching degrees between each product vector and the user portrait vector. For example, assuming that the matching condition is set to: the matching degree between the corresponding product vector and the user portrait vector of the target user is the top_1 (i.e., the highest matching degree) of the matching degrees between each product vector and the user portrait vector, then the product information of the product with the highest matching degree may be specifically included in the generated product recommendation result. That is, the product information of the product with the highest matching degree is recommended to the target user.

[0099] However, it is not limited to this. In other implementations, at least one target product that has an association relationship with the historical interaction data of the target user on the product can also be selected from at least one product that meets the matching condition based on the historical interaction data between the target user and the product, and a product recommendation result containing the product information of the at least one target product is generated. In this implementation, the target product recommended to the target user not only meets the matching condition (such as the matching degree between the corresponding product vector and the user portrait vector of the target user is the top_k of the matching degrees between each product vector and the user portrait vector), but also has an association relationship with the historical interaction data of the target user on the product, for example, the recommended target product is the style, type, etc. preferred in the historical interaction record of the target user.

[0100] According to the above scheme, the agent-based product recommendation method provided by the present application pre-sets a large model agent including an interactive agent and an analysis agent for the product recommendation scenario. When a new user joins, the interactive agent in the large model agent is used to interact with the new user through question-and-answer interaction to support the large model to generate a preliminary user portrait for the new user, that is, the initial portrait, to help the system make more accurate recommendations even without a large amount of personal data. At the same time, in the early stage when user data is insufficient, a large model is used to simulate the feedback of different users on the recommendation results, and these simulated data are used to train the recommendation ability of the analysis agent, so that the analysis agent can better understand the preferences of different types of users before actually accumulating enough user data. Through the above scheme, the present application can effectively solve the cold start problem of the recommendation system in the case of a lack of new user data in the early stage, thereby optimizing the product recommendation technology and improving the performance and user experience of the recommendation system.

[0101] For the set of products to be recommended, the embodiment of the present application pre-constructs the product vector of each product in the set of products to be recommended to form a product vector library of products, so as to support the recommendation system to efficiently retrieve and process large-scale data sets.

[0102] In an alternative embodiment, see Figure 3 , the process of constructing the product vector library of the product set to be recommended can be implemented as follows:

[0103] Step 301: determine each key feature of the product to be recommended based on the product information of the product to be recommended in the product set to be recommended; the key feature of the product to be recommended is a feature selected from the product information of the product to be recommended, or a feature obtained by processing the selected feature.

[0104] Specifically, the product information of each product to be recommended in the recommended product set can be collected first, such as product description, specifications, evaluation and other information. Different products have different information. For example, the information of telecommunications government and enterprise products usually includes product name, product code, product application scenario, product function, product advantage, success case, etc. This information may come from different sources, such as databases, files, networks, etc.

[0105] Before determining the key features of each product to be recommended in the product set to be recommended based on the collected product information, optionally, the collected original product information may be first cleaned and preprocessed. In particular, irrelevant information such as advertisements, comments, etc. may be removed through data cleaning. After that, the corresponding preprocessing is continued, and the preprocessing performed may include but is not limited to any one or more of the following:

[0106] 1) Handle missing values, such as filling missing values ​​with mean, median, mode or specific value;

[0107] 2) Handling outliers, such as using the quartile rule to detect and handle outliers;

[0108] 3) Handle duplicate values, such as deleting duplicate records;

[0109] 4) Normalize the data and convert it to a uniform numerical range, such as the [0,1] interval or the standard normal distribution. Categorical variables are also converted to numerical data for subsequent processing;

[0110] 5) Data merging: merging different data tables into one data table based on common keys.

[0111] See also Figure 4 FIG. 1 is an exemplary implementation flowchart of constructing a product vector library. After data cleaning and preprocessing of the collected original product information, the key features of the products to be recommended can be further determined based on feature engineering according to the product information after data cleaning and preprocessing.

[0112] Optionally, important features may be selected from the preprocessed product information data by statistical analysis, correlation analysis and / or principal component analysis. The selected features should be able to effectively express the characteristics of the product and have a significant impact on the prediction accuracy (recommendation accuracy) of the recommendation system. Specifically, the important features may be some important parameter data selected from the product information parameters based on correlation analysis and other methods.

[0113] After selecting important features, new features can be created by combining, converting, etc. the selected features. For example, the embodiment of the present application multiplies or divides two or more related features to create a new feature, or a continuous feature is discretized to create a new feature. If the number of features is too large, feature dimensionality reduction can also be achieved through principal component analysis and other methods. On this basis, the important features selected from the product information of the product to be recommended, and / or the features obtained by processing the selected features, can be used as the key features of the product to be recommended.

[0114] In addition, the quality of key features can be evaluated by cross-validation, AUC (Area Under the Curve), F1 score and other indicators. If the quality of key features does not meet expectations, you can return to the feature selection step to re-select important features and generate subsequent new features to ensure the quality of key features.

[0115] Step 302: vectorize each key feature of the product to be recommended to obtain a feature vector corresponding to each key feature.

[0116] After obtaining the key features of the products to be recommended, these features are converted into vectors through vectorization processing.

[0117] In implementation, optionally, a word embedding model can be trained, and the word embedding model can be used to map the words / phrases in the key features of the recommended products from the vocabulary to the vector space to obtain the feature vectors of the key features of the recommended products. This mapping can be learned through a neural network model. Assuming that w represents the words or phrases in the key features, V represents the vocabulary, and θ represents the model parameters, the word embedding can be expressed as a function: ,Right now , where e is an n-dimensional real vector and R represents the real space. Based on the word embedding model Word2Vec, it is learned by optimizing the following formula:

[0118]

[0119] Among them, D is the training corpus, c is the context word, and w is the word or phrase in the key feature. Calculated by the softmax function:

[0120] .

[0121] After completing the training of the word embedding model, use the trained word embedding model to convert each key feature of the product to be recommended into a vector. This vector is the vector representation of the key feature, which can retain the semantic information of the feature and reduce the dimension of the data.

[0122] Specifically, suppose there is a sample x of products to be recommended, whose key features are (f1, f2, f3, ..., fn), then the feature vectorization can be defined as a function , that is, F(x) = (f1(x), f2(x), f3(x), ..., fn(x)). In text processing, a feature function tf-idf (w, d) is defined for each word w, where d is a document. The calculation formula of this function is: tf-idf(w, d) = tf(w, d)×idf(w), where:

[0123] tf(w, d) = (number of times #w appears in d) / (all words in #d);

[0124] idf(w) = log(#document set) / (#total number of documents containing w).

[0125] Step 303: merge the feature vectors corresponding to the key features into one vector to obtain a product vector corresponding to the product to be recommended.

[0126] For each product to be recommended, after obtaining the vector representation of its key features based on feature engineering, refer to Figure 4 , further executing the process of generating product vectors, wherein, for each product to be recommended, the feature vectors of its various key features can be merged into one vector through, but not limited to, averaging, weighted averaging, maximum pooling and other methods, and the merged vector is the vector representation of the product to be recommended, that is, the product vector of the product to be recommended.

[0127] Each product to be recommended corresponds to a corresponding product vector, which captures the key elements of product information and allows products to be compared mathematically.

[0128] See also Figure 4 After generating the product vector, the storage problem of the product vector is further faced. The generated product vector can be stored in a database to form a product vector library of the product set to be recommended, so as to facilitate subsequent retrieval and use. The formed product vector library includes product vectors corresponding to each product to be recommended in the product set to be recommended, and each product vector is mapped to the corresponding product to be recommended through the corresponding index in the preset index structure.

[0129] In implementation, a database suitable for storing and retrieving a large number of vectors can be selected, and an efficient index structure can be established based on, but not limited to, an inverted index to map product vector values ​​to their corresponding products, thereby maximizing retrieval efficiency. Specifically, each product vector can be inserted into the database, and some optimizations can be performed during the insertion process, such as batch insertion, to increase the insertion speed and reduce database pressure. Once the product vector is stored in the database and the index is built, subsequent queries and retrieval can be supported. At the same time, optionally, frequently accessed data is retained in memory to implement an intelligent caching mechanism, reduce access pressure on the storage system, and improve retrieval efficiency.

[0130] After completing the storage of product vectors to form a product vector library, optionally, refer to Figure 4 , and the product vector library can also be optimized and updated.

[0131] The update process of the product vector library can be based on the continuous addition of new products and / or changes in existing product information to update the product vectors in the database. This update process can be real-time, that is, whenever a new product is added or the existing product information changes, the corresponding product vector is immediately updated synchronously (adding a new product vector or modifying / deleting an existing product vector); or it can be updated periodically, such as updating the product vector library once a day or a week. In either case, it is sufficient to ensure that the product vectors in the database can always correctly reflect the actual product information.

[0132] For the optimization of the product vector library, optionally, the pre-set performance indicators can be monitored according to the usage of the product vector in actual applications, such as query frequency, query response time, etc. According to the monitoring results, the generation method of the product vector and the storage and retrieval strategy are optimized. The optimization process can specifically, but is not limited to, improve the accuracy of the generated product vector by improving the algorithm of vector generation, and / or improve the storage and retrieval speed by adjusting the structure or parameters of the product vector library.

[0133] Optionally, after each optimization is completed, the optimization effect can be evaluated to verify the optimization effect, and if the optimization effect does not meet expectations, the optimization can be iterated or the optimization strategy can be adaptively adjusted.

[0134] The evaluation can be completed by comparing the performance indicators before and after the optimization. The performance indicators may include but are not limited to the response time and hit accuracy of the product vector library.

[0135] Through the above implementation, this embodiment realizes the construction of a product vector library as a long-term memory module of a large model agent, which can not only store and retrieve a large amount of data and support efficient processing of large-scale data sets, but also provide richer and deeper insights for the recommendation system, thereby achieving more accurate and personalized recommendation services. Therefore, based on this embodiment, large-scale data sets can be effectively processed, while improving the accuracy and efficiency of the recommendation system.

[0136] In addition, the embodiment of the present application proposes and implements vectorized matching technology to measure the matching between users and products. By converting user portraits and product information into high-dimensional vectors, and using cosine similarity and other methods to analyze the matching between user portrait vectors and product vectors, it is not only possible to accurately identify the subtle differences between user preferences and product characteristics, but also to significantly improve the matching accuracy and processing speed. This vectorization method shows its powerful performance and efficiency when processing large-scale data.

[0137] In an alternative embodiment, see Figure 5 As shown in the method flow chart, the agent-based product recommendation method provided by the present application may further include the following processing after step 103:

[0138] Step 104: Generate recommendation reason explanation information for the product recommendation result based on the user profile of the target user, the matching between the user profile vector and the product vectors, and the semantic understanding of at least part of the target user's historical interaction data on the product.

[0139] In addition to selecting suitable products for personalized recommendations to target users based on the matching degree between the user portrait vector and the product vector through the analysis agent, in this embodiment, the analysis agent also selects suitable products based on the matching degree between the user image, the user portrait vector and the product vector and / or the historical interaction data of the target user (such as historical interaction records of the products browsed / clicked / purchased, products unsubscribed by the target user, and corresponding behavior patterns, etc.), and combines the big model to perform semantic analysis and understanding of at least part of these data, so as to explain the reasons for the personalized recommendation results of the products and generate recommendation reason explanation information for the product recommendation results.

[0140] Optionally, during implementation, product recommendation results and corresponding explanation information of recommendation reasons may be passed to an interactive agent, which will display diversified product recommendation results and recommendation reasons to target users. The interactive agent may be specifically responsible for displaying the product recommendation results in a user-friendly manner and providing detailed recommendation reasons to increase the transparency and credibility of the recommendation.

[0141] Based on the embodiment of the present application, not only can personalized product recommendations be provided to users, but also the explanation of the reasons for the recommendations can be realized by combining the semantic analysis capabilities of the analysis agent with the large model, and complete and diverse reasons for the recommendations can be obtained and presented to the user. This explanatory recommendation method of the embodiment of the present application not only improves the user's trust in the recommendation system, but also enhances the user's understanding and acceptance of the recommended content. By explaining the reasons for the recommendation, the user can better understand the logic behind the recommendation, thereby improving the user experience.

[0142] In an alternative embodiment, see Figure 6 As shown in the method flow chart, the agent-based product recommendation method provided by the present application may further include the following processing after step 103:

[0143] Step 105: Obtain feedback information from the target user on the product recommendation result.

[0144] Step 106: Adjust model parameters of a recommendation model according to the feedback information; the recommendation model is a model constructed according to the recommendation process of the analysis agent and used to implement a recommendation function for the analysis agent.

[0145] In this embodiment, a recommendation model for implementing the recommendation function of the analysis agent is constructed in advance according to the recommendation process of the analysis agent.

[0146] The recommendation model is specifically constructed by determining model parameters based on a predefined optimization problem. The goal of the optimization problem includes: determining a set of model parameters for the recommendation model to minimize the weighted sum of the prediction error and model complexity of the recommendation model while satisfying computing resource constraints.

[0147] When designing a recommendation system, we are faced with the dual constraints of model complexity and computing resources. In order to find the best balance between the two, this application defines the above optimization problem, the goal of which is to minimize the prediction error of the model while controlling the complexity of the model while satisfying the computing resource constraints. Model complexity can be measured by indicators such as the number of model parameters and the number of model layers, while computing resources can include but are not limited to computing time and memory usage. By adjusting the model hyperparameters, a trade-off can be made between prediction accuracy and computational efficiency to select the most suitable model parameters θ.

[0148] Specifically, for the above optimization goal, i.e. minimizing the prediction error of the model, taking into account the complexity of the model and the limitations of computing resources, the following optimization problem can be defined:

[0149]

[0150]

[0151] in, is the prediction error of the model, y is the true value, is the predicted value of the model, X is the input data, are model parameters. It is a measure of model complexity, such as the number of model parameters or the depth of the model. is a hyperparameter that is used to balance the trade-off between prediction error and model complexity. are the computational resources required by the model, such as computation time or memory usage. Represents the limit of available computing resources. The goal of this optimization problem is to find a set of model parameters θ that minimizes the weighted sum of the model's prediction error and model complexity while satisfying the computing resource limit. By making an optimal comparison between model complexity and computing resources, the most suitable model parameters are selected to form a prediction model, so as to provide the analysis agent with prediction-based product recommendation capabilities.

[0152] For example, for the telecom government business recommendation scenario, x can include two aspects: one is telecom government business information (which may include product name, product code, product application scenario, product function, product advantage, success case), and the other is enterprise information (which may include enterprise name, industry category, industry subcategory, enterprise location, etc.). y is the product code. It is the appropriate product code predicted based on the input x information.

[0153] After obtaining the product recommendation results and even the reasons for the recommendation based on the designed interactive agent and analysis agent and presenting them to the user, information such as the recommended products and the reasons for the recommendation will be displayed on the user interface, and the user can give feedback on this.

[0154] This embodiment further collects user (such as the target user) feedback information on product recommendation results and reasons for recommendations, including but not limited to collecting user satisfaction with the recommended products and whether the users actually purchased the recommended products, etc., and dynamically adjusts the recommendation strategy based on the designed algorithm, and adjusts the model parameters of the recommendation model according to the user feedback information, so as to update and optimize the model and improve the recommendation quality and user experience.

[0155] That is to say, the overall recommendation architecture proposed in this application is not static, but can be continuously optimized and iterated based on user feedback. The analysis agent can continuously monitor the performance indicators of the recommendation system and make adjustments based on the real-time feedback and long-term behavior patterns of users. In this way, the recommendation system can adapt to the dynamic changes of user preferences and provide more personalized and user-friendly recommendation services.

[0156] In implementation, optionally, the product data can be classified first, with U representing the user and P representing the product, and a preference function f can be defined: U×P→{must choose, will choose, not choose}, where "(U, P)" represents the set of user U and product P, and serves as the input of function f, and the output is one of {must choose, will choose, not choose}, representing the different selection preferences of user U for product P. This function can be trained by a machine learning algorithm.

[0157] Then, the accuracy of recommendation is evaluated. Here, we assume that R represents the recommended product set and A represents the product set actually selected by the user. Then, we collect user feedback and define a feedback function f': U×P→{satisfaction, purchase or not}. Here, "(U, P)" also represents the set of users U and products P, and serves as the input of f'. The output is one of {satisfaction, purchase or not}, which represents the user U's satisfaction with product P and his purchase behavior. Satisfaction can be obtained by directly asking users or monitoring user evaluations. Purchase behavior can be obtained by monitoring users' online shopping records.

[0158] After training an initialized recommendation model (M) according to the recommendation process of the analysis agent, the historical dataset (H) of user satisfaction and purchase behavior is initialized according to user feedback. For each user (u) and recommended product (p), satisfaction and purchase behavior are collected as feedback (f'(u, p)), and the historical dataset (H) is updated in real time. Then, the model parameters of the recommendation model are dynamically adjusted according to the latest user feedback through online stochastic gradient descent using the online learning method. At the same time, the performance of the adjusted model is evaluated to ensure the improvement of recommendation accuracy (RA). After that, based on the updated recommendation model (M'), a recommendation list (R) is generated, and user satisfaction and purchase behavior feedback on the recommendation list are collected. The above update steps are repeated to iteratively update the recommendation model.

[0159] Specifically, the recommendation accuracy (RA) can be obtained by calculating the size of the intersection between the recommendation set R and the product set A actually selected by the user, and dividing it by the size of the recommendation set. Calculation formula:

[0160] RA= |R∩A| / |R|

[0161] Among them, “|R∩A|” represents the number of intersections between recommended products and products actually selected by users, and “|R|” represents the number of recommended products. The ratio between the two reflects the proportion of products provided by the recommendation system that users are actually interested in and choose.

[0162] The optimization of the recommendation model mainly involves adjusting model parameters, changing model structure and updating training data. After receiving a user's feedback on the recommended product (f'(u, p)), the loss function (L) can be calculated and updated online. The error L can be calculated using the result of the feedback function f', and the model parameter θ can be updated in the following way according to the gradient descent rule:

[0163]

[0164] Among them, α represents the learning rate, is the gradient of L with respect to θ_old. Considering the problems of overfitting and underfitting, the regularization term λ is introduced and the parameter update formula is modified as follows:

[0165]

[0166] The optimization of the recommendation system is an iterative process. It can be achieved by repeatedly collecting user feedback and optimizing the model, and using the recommendation accuracy evaluation formula to measure the effect of the model until the model parameters converge or the preset number of iterations is reached. Through this iterative process, the recommendation system can gradually adapt to changes in user behavior and preferences, thereby providing more personalized and accurate recommendation services.

[0167] This embodiment uses online learning and incremental update technology to enable the recommendation model to adjust the recommendation strategy in real time according to user feedback. This mechanism ensures that the recommended content can quickly adapt to changes in user preferences, thereby continuously improving the quality of recommendations. Accordingly, the recommendation system in the embodiment of the present application has the ability to adjust the recommendation strategy in real time according to user feedback. This dynamic adjustment mechanism enables the recommendation system to quickly adapt to changes in user behavior and preferences, and provide users with more relevant and satisfactory recommendation content.

[0168] Corresponding to the above method, the present application embodiment also provides a product recommendation device based on an intelligent agent, see Figure 7 The schematic diagram of the composition structure of the device comprises:

[0169] The acquisition module 701 is used to obtain a user portrait of the target user in response to obtaining a product recommendation request for the target user; if the target user is a new user, the user portrait of the target user is an initial portrait constructed by using the big model based on the question-and-answer interaction by using an interactive agent in a pre-set big model agent to interact with the target user;

[0170] A determination module 702 is used to determine the matching degree between the user portrait vector of the target user and each product vector in the product vector library of the product set to be recommended based on the analysis agent in the large model agent; the user portrait vector is a vector corresponding to the user portrait of the target user;

[0171] A generating module 703, configured to generate a product recommendation result for the target user based on the analysis agent and according to the matching degree between the user portrait vector and each product vector;

[0172] In the case of insufficient user data, the recommendation capability of the analysis agent is formed by using a large model to simulate the feedback of different users on the recommendation results, and training the recommendation capability based on the simulated data.

[0173] In an optional implementation manner, if the target user is not a new user, the user portrait of the target user is an updated portrait obtained by updating the initial portrait based on the information data generated by the target user for the product, or is a portrait of a user of the category to which the target user belongs;

[0174] When constructing the initial profile of the target user, the interaction agent is specifically used to:

[0175] Collecting basic information and preference information of the target user through question-answer interaction between the interactive agent and the target user;

[0176] Inputting the basic information and preference information of the target user and the guidance information for instructing the large model to construct a user portrait into the large model, and obtaining an initial portrait of the target user output by the large model;

[0177] When the interaction agent updates the initial portrait based on the information data generated by the target user for the product, it is specifically used to: update the initial portrait based on the historical interaction data and / or behavior pattern of the target user for the product.

[0178] In an optional implementation manner, the device further includes a construction module for constructing a product vector library. When constructing the product vector library, the construction module is specifically used to:

[0179] Determine each key feature of the product to be recommended according to the product information of the product to be recommended in the product set to be recommended; the key feature of the product to be recommended is a feature selected from the product information of the product to be recommended, or a feature obtained by processing the selected feature;

[0180] Performing vectorization processing on each key feature of the product to be recommended, respectively, to obtain feature vectors corresponding to each key feature;

[0181] Merge the feature vectors corresponding to the key features into one vector to obtain a product vector corresponding to the product to be recommended;

[0182] The product vector library includes product vectors corresponding to each product to be recommended in the product set to be recommended; each product vector is mapped to the corresponding product to be recommended through a corresponding index in a preset index structure.

[0183] In an optional implementation manner, the user portrait vector of the target user is a vector corresponding to the initial portrait of the target user, or a comprehensive vector constructed by integrating the historical interaction data and / or behavior pattern of the target user on the product with the initial portrait, or a vector corresponding to the portrait of the category of users to which the target user belongs;

[0184] The determination module 702 is specifically configured to:

[0185] quantifying the degree of match between the user portrait vector of the target user and each product vector in the product vector library using a predefined matching function, and obtaining the degree of match between the user portrait vector of the target user and each product vector in the product vector library;

[0186] Among them, the matching function includes: the matching degree between the user portrait vector and the product vector in the product vector library, and the functional relationship between the multidimensional vector representing user characteristics contained in the user portrait vector and the multidimensional vector representing product characteristics contained in the product vector.

[0187] In an optional implementation manner, the generating module 703 is specifically configured to:

[0188] Determine, from the set of products to be recommended, at least one product whose matching degree between a corresponding product vector and a user portrait vector of the target user satisfies a matching degree condition;

[0189] Generate a product recommendation result including the product information of the at least one product; or, select at least one target product from the at least one product that has a correlation with the target user's historical interaction data on the product, and generate a product recommendation result including the product information of the at least one target product.

[0190] In an optional implementation manner, the generating module 703 is further configured to:

[0191] Based on the user portrait of the target user, the matching between the user portrait vector and the product vectors, and the semantic understanding of at least part of the historical interaction data of the target user on the product, recommendation reason explanation information for the product recommendation result is generated.

[0192] In an optional implementation, the adjustment module is used to:

[0193] Obtaining feedback information from the target user on the product recommendation result;

[0194] The model parameters of the recommendation model are adjusted according to the feedback information; the recommendation model is a model constructed according to the recommendation process of the analysis agent and used to implement the recommendation function of the analysis agent.

[0195] In an optional embodiment, the recommendation model is constructed by determining model parameters based on a predefined optimization problem;

[0196] The goal of the optimization problem includes: determining a set of model parameters for the recommendation model while satisfying computing resource constraints so as to minimize the weighted sum of the prediction error and model complexity of the recommendation model.

[0197] As for the agent-based product recommendation device disclosed in the embodiment of the present application, since it corresponds to the agent-based product recommendation method disclosed in the above method embodiment, the description is relatively simple. For the relevant similarities, please refer to the description of the above method embodiments, which will not be described in detail here.

[0198] In addition, the present application also provides a computer-readable medium having a computer program stored thereon, wherein the computer program includes a program code for executing the agent-based product recommendation method disclosed in any of the method embodiments above.

[0199] In the context of the present application, a computer readable medium (machine readable medium) may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine readable medium may be a machine readable signal medium or a machine readable storage medium. A machine readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or any suitable combination of the foregoing. A more specific example of a machine readable storage medium may include an electrical connection based on one or more lines, 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), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0200] In summary, the agent-based product recommendation method, device, and medium provided in the embodiments of the present application have at least the following technical advantages over the known technologies:

[0201] 1) Highly interactive dual-agent large model framework: The highly interactive dual-agent large model framework proposed in this application represents a significant advancement in recommendation system technology, especially in the dynamic construction of user portraits and the accuracy of personalized recommendations. Through natural language communication between interactive agents and users, this application can not only capture users' immediate feedback in real time, thereby dynamically adjusting the recommended content, but also further refine and optimize user portraits by learning users' long-term behavior patterns. The dynamic update mechanism of user portraits can reflect users' latest preferences in real time, providing continuously updated data support for the recommendation system. This method greatly improves the flexibility and adaptability of the recommendation system and can more accurately reflect and meet users' personalized needs. In addition, the introduced analysis agent can process and analyze huge data sets and extract deep insights that are crucial to recommendation decisions, which not only improves the efficiency and accuracy of data processing, but also enhances the decision support capabilities of the recommendation system. Through this highly interactive and intelligent dual-agent structure, this application significantly improves the performance of the recommendation system, especially in terms of the accuracy of personalized recommendations and user satisfaction.

[0202] This application uses an interactive agent to interact with users in a two-way manner, dynamically building and updating user portraits. This interactive user portrait construction method is relatively rare in existing recommendation systems. Through real-time interaction and feedback, the recommendation system can better adapt to users' personalized needs and behavioral changes, and provide users with recommended content that is more in line with their current needs.

[0203] 2) Comprehensive product vectorization processing flow and advanced vectorization matching technology: From data cleaning, feature engineering to the generation and storage of product vectors, this application proposes a complete set of product vectorization processing flows. This process not only improves data processing efficiency, but also provides strong support for subsequent vectorization matching and recommendation algorithms. In this way, the recommendation system can quickly and accurately retrieve the products that best match the user profile from the product library, thereby improving the accuracy and efficiency of recommendations. Advanced vectorization matching technology is another highlight of this application. Through efficient algorithms, user portraits and product information are converted into high-dimensional vectors, and user portrait vectors are matched with product vectors using methods such as cosine similarity. It can not only accurately identify subtle differences between user preferences and product features, but also significantly improve processing speed and matching accuracy. This vectorization method has shown its powerful performance and efficiency when processing large-scale data.

[0204] 3) Combination of vectorized recommendation and long-term memory module: This application not only improves the processing speed and efficiency of the recommendation system by building a product vector library as a long-term memory module for the large model proxy, but also enables long-term storage and retrieval of user data and product information. This method of combining vectorized recommendation and long-term memory module provides users with continuous personalized recommendation services while ensuring the timeliness and accuracy of the recommended content.

[0205] 4) Explanatory recommendation algorithm: This application not only provides personalized recommendations, but also explains the reasons for recommendations by combining the semantic analysis capabilities of the analytical agent with the large model. This explanatory recommendation algorithm not only improves users' trust in the recommendation system, but also enhances users' understanding and acceptance of the recommended content. By explaining the reasons for the recommendation, users can better understand the logic behind the recommendation, thereby improving the user experience.

[0206] 5) Dynamically adjusted recommendation strategy, recommendation strategy optimization mechanism based on real-time feedback drive: The recommendation system of this application can capture user feedback in real time and has the ability to adjust the recommendation strategy in real time based on user feedback. This optimization mechanism based on real-time feedback drive is advanced in the field of recommendation systems, enabling the recommendation system to quickly adapt to changes in user behavior and preferences, thereby providing users with more relevant and satisfactory recommendation content. Through online learning and incremental update technology, the recommendation model can adjust the recommendation strategy in real time based on user feedback, thereby achieving continuous optimization of the model. This mechanism ensures that the recommended content can quickly adapt to changes in user preferences, thereby continuously improving the quality of recommendations.

[0207] 6) Comprehensively consider the optimization of computing resources and design a recommendation model based on resource optimization: In terms of computing resource optimization, this application achieves the best balance between minimizing prediction errors and model complexity and computing resources through carefully designed optimization problems. This design takes into account resource constraints in different computing environments, ensuring that the recommendation model not only runs efficiently on large servers, but also runs smoothly on resource-constrained mobile devices. This strategy of comprehensive consideration of computing resources significantly improves the practicality and scalability of the model, allowing the recommendation service to cover a wider range of user groups.

[0208] By considering the limitations of computing resources at the model design stage, this application can provide users with recommendation services that can run efficiently on different devices and platforms.

[0209] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0210] For the convenience of description, the above system or device is described by dividing it into various modules or units according to its functions. Of course, when implementing the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0211] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0212] Finally, it should be noted that, in this article, relational terms such as first, second, third and target are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the statement "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0213] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A product recommendation method based on an agent, characterized in that: include: In response to obtaining a product recommendation request for a target user, obtaining a user profile of the target user; If the target user is a new user, the user profile of the target user is an initial profile constructed by using the big model based on the question-and-answer interaction by using an interactive agent in a pre-set big model agent to conduct question-and-answer interaction with the target user; Determine, based on the analysis agent in the large model agent, the degree of match between the user portrait vector of the target user and each product vector in the product vector library of the product set to be recommended; The user portrait vector is a vector corresponding to the user portrait of the target user; Based on the analysis agent, generating a product recommendation result for the target user according to the matching degree between the user portrait vector and each product vector; In the case of insufficient user data, the recommendation capability of the analysis agent is formed by using a large model to simulate the feedback of different users on the recommendation results, and training the recommendation capability based on the simulated data.

2. The agent-based product recommendation method according to claim 1, characterized in that: If the target user is not a new user, the user portrait of the target user is an updated portrait obtained by updating the initial portrait based on the information data generated by the target user for the product, or is a portrait of a user of the category to which the target user belongs; The process of constructing the initial portrait of the target user includes: Collecting basic information and preference information of the target user through question-answer interaction between the interactive agent and the target user; Inputting the basic information and preference information of the target user and the guidance information for instructing the large model to construct a user portrait into the large model, and obtaining an initial portrait of the target user output by the large model; The updating of the initial portrait based on the information data generated by the target user for the product includes: The initial portrait is updated based on the target user's historical interaction data and / or behavior pattern with the product.

3. The agent-based product recommendation method according to claim 1, characterized in that: The process of constructing the product vector library includes: Determine each key feature of the product to be recommended according to the product information of the product to be recommended in the product set to be recommended; the key feature of the product to be recommended is a feature selected from the product information of the product to be recommended, or a feature obtained by processing the selected feature; Performing vectorization processing on each key feature of the product to be recommended, respectively, to obtain feature vectors corresponding to each key feature; Merge the feature vectors corresponding to the key features into one vector to obtain a product vector corresponding to the product to be recommended; The product vector library includes product vectors corresponding to each product to be recommended in the product set to be recommended; each product vector is mapped to the corresponding product to be recommended through a corresponding index in a preset index structure.

4. The agent-based product recommendation method according to claim 2, characterized in that: The user portrait vector of the target user is a vector corresponding to the initial portrait of the target user, or a comprehensive vector constructed by integrating the historical interaction data and / or behavior pattern of the target user on the product with the initial portrait, or a vector corresponding to the portrait of the category of users to which the target user belongs; The determining of the matching degree between the user portrait vector of the target user and each product vector in the product vector library of the product set to be recommended includes: quantifying the degree of match between the user portrait vector of the target user and each product vector in the product vector library using a predefined matching function, and obtaining the degree of match between the user portrait vector of the target user and each product vector in the product vector library; Among them, the matching function includes: the matching degree between the user portrait vector and the product vector in the product vector library, and the functional relationship between the multidimensional vector representing user characteristics contained in the user portrait vector and the multidimensional vector representing product characteristics contained in the product vector.

5. The agent-based product recommendation method according to claim 1, characterized in that: The generating a product recommendation result for the target user according to the matching degree between the user portrait vector and each product vector includes: Determine, from the set of products to be recommended, at least one product whose matching degree between a corresponding product vector and a user portrait vector of the target user satisfies a matching degree condition; Generate a product recommendation result including the product information of the at least one product; or, select at least one target product from the at least one product that has a correlation with the target user's historical interaction data on the product, and generate a product recommendation result including the product information of the at least one target product.

6. The agent-based product recommendation method according to claim 1, characterized in that: Also includes: Based on the user portrait of the target user, the matching between the user portrait vector and the product vectors, and the semantic understanding of at least part of the historical interaction data of the target user on the product, recommendation reason explanation information for the product recommendation result is generated.

7. The agent-based product recommendation method according to claim 1, characterized in that: Also includes: Obtaining feedback information from the target user on the product recommendation result; The model parameters of the recommendation model are adjusted according to the feedback information; the recommendation model is a model constructed according to the recommendation process of the analysis agent and used to implement the recommendation function of the analysis agent.

8. The agent-based product recommendation method according to claim 7, characterized in that: The recommendation model is constructed by determining model parameters based on a predefined optimization problem; The goal of the optimization problem includes: determining a set of model parameters for the recommendation model while satisfying computing resource constraints so as to minimize the weighted sum of the prediction error and model complexity of the recommendation model.

9. A product recommendation device based on an intelligent agent, characterized in that: include: An acquisition module, configured to acquire a user portrait of a target user in response to obtaining a product recommendation request for the target user; If the target user is a new user, the user profile of the target user is an initial profile constructed by using the big model based on the question-and-answer interaction by using an interactive agent in a pre-set big model agent to conduct question-and-answer interaction with the target user; A determination module, configured to determine, based on the analysis agent in the large model agent, a degree of match between the user portrait vector of the target user and each product vector in the product vector library of the product set to be recommended; The user portrait vector is a vector corresponding to the user portrait of the target user; A generating module, configured to generate a product recommendation result for the target user based on the analysis agent and according to the matching degree between the user portrait vector and each product vector; In the case of insufficient user data, the recommendation capability of the analysis agent is formed by using a large model to simulate the feedback of different users on the recommendation results, and training the recommendation capability based on the simulated data.

10. A computer readable medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it can be used to implement the agent-based product recommendation method as described in any one of claims 1-8.

Images