Method and system for commodity recommendation based on agent and hotspot information cooperation

Abstract

The application discloses a kind of based on agent and hot information cooperation's commodity recommendation method and system, it is related to intelligent recommendation technical field, including: through hot information collection tool, commodity information collection tool and user information collection tool parallel acquisition structured hot spot data, commodity dynamic information and current user data;Based on large language model, multi-source information fusion is generated including target user, hot spot associated direction, commodity constraint condition, user preference matching point and recommendation priority structured recommendation demand;Matching and screening the commodity that meets demand, then according to hot spot associated density, user preference matching degree, commodity performance ratio and inventory adequacy multidimensional priority ranking, generate optimized commodity recommendation list.The application is through the depth cooperation of hot information and user demand, based on large language model carries out semantic understanding and information fusion, improves the accuracy, timeliness and adaptability of commodity recommendation.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology, and more specifically, to a product recommendation method and system based on the collaboration between intelligent agents and trending information. Background Technology

[0002] With the rapid development of e-commerce and Internet technology, product recommendation systems have become the core bridge connecting users and massive amounts of goods. Their performance directly affects the user experience and business benefits of the platform. At present, the mainstream product recommendation technologies are mainly based on collaborative filtering algorithms, content recommendation algorithms and deep learning recommendation models. These methods have achieved personalized recommendations to a certain extent, but in practical applications, there are still the following prominent problems: (1) Insufficient timeliness of recommendations: Traditional recommendation technologies rely on users' historical consumption data and static interest tags, making it difficult to capture social hotspots, industry dynamics and temporary changes in users' interests in real time. (2) Weak cross-scenario information collaboration capability: The behavioral data (such as social interaction, browsing history and consumption evaluation) generated by users on multiple platforms (such as social platforms, e-commerce APPs and local life services) has not been effectively integrated. Existing recommendation systems are usually limited to single-scenario data mining and lack cross-platform data collaborative analysis, resulting in incomplete user profiles and limited recommendation accuracy. (3) Low efficiency of recommendation strategy iteration: Existing systems rely on manual adjustment or offline training, making it difficult to dynamically update recommendation strategies based on real-time user feedback (such as clicks, favorites, purchases and evaluations). When user needs or trending information change, the system cannot adjust the recommendation weights in a timely manner, resulting in delayed recommendation results, which affects user experience and platform revenue.

[0003] For example, when a user searches for "toys," incorporating trending information such as popular movies and cartoon characters into the recommendation process would significantly improve the recommendation effect. However, traditional recommendation algorithms have obvious shortcomings in acquiring and integrating external trending information. Therefore, there is an urgent need for a product recommendation method and corresponding intelligent agent system that can achieve deep collaboration between trending information and user needs, efficient integration of cross-scenario data, and real-time iteration of recommendation strategies to meet the development needs of modern e-commerce platforms. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention proposes a product recommendation method and system based on the collaboration between intelligent agents and trending information. By deeply collaborating trending information with user needs, and performing semantic understanding and information fusion based on a large language model, the accuracy, timeliness, and adaptability of product recommendations are improved.

[0005] The first aspect of this invention provides a product recommendation method based on the collaboration of intelligent agents and hotspot information, comprising the following steps:

[0006] S101: Obtain structured hotspot data, product dynamic information, and current user data in parallel through hotspot information collection tools, product information collection tools, and user information collection tools;

[0007] S102: The acquired structured hotspot data, product dynamic information and user data are fused with multi-source information based on a large language model to generate structured recommendation requirements that include target users, hotspot association directions, product constraints, user preference matching points and recommendation priorities;

[0008] S103: Input the generated structured recommendation requirements into the recommendation module based on the large language model, match and filter products that meet the requirements, and then sort them in multiple dimensions according to the closeness of hot topic association, user preference matching degree, product cost performance and inventory sufficiency to generate an optimized product recommendation list.

[0009] S104: The product recommendation list is visualized to the user through the user interaction module, and the user's real-time feedback on the recommendation results is collected. The real-time feedback is used to optimize and iterate the subsequent recommendation strategy.

[0010] In this solution, step S101 involves acquiring structured hotspot data, product dynamic information, and current user data in parallel using a hotspot information collection tool, a product information collection tool, and a user information collection tool, including:

[0011] By using hot topic information collection tools, we can capture social hot topics and industry dynamics information in real time from external search engines and online platforms, extract themes and analyze sentiment trends to obtain structured hot topic feature words and generate structured hot topic data.

[0012] By using product information collection tools, dynamic product information related to the user's current search and historical behavior is obtained from the product supply platform, and a preliminary recommended product pool is generated based on the semantic similarity calculation of hot feature words and product attributes.

[0013] By using user information collection tools, static basic information, dynamic behavioral data, historical consumption records and preference tags of users are collected from user data sources. The collected raw data is integrated into a complete user profile and then cleaned and de-identified to generate current user data.

[0014] In this solution, step S102 involves fusing the acquired structured hotspot data, product dynamic information, and current user data using a large language model, including:

[0015] The collected structured hotspot data, product dynamic information and current user data are aligned, and the predefined information fusion prompt template in the prompt template management module is called to integrate the aligned multi-source information and input it into the large language model;

[0016] By understanding the semantic relationship between structured hotspot data and current user data through the large language model, and understanding and applying real-time business rules in product dynamic information as hard constraints, the system combines users' long-term historical preferences, short-term behaviors and current real-time intentions to generate structured recommendation requirements that include target user attributes, hotspot association directions, product constraints, user preference matching points and recommendation priorities.

[0017] In this scheme, in step S103, the generated structured recommendation requirements are input into the recommendation module based on a large language model to generate a product recommendation list, including:

[0018] The structured recommendation request is accepted, and the pre-trained large language model is used to parse and decompose the structured recommendation request into a search and filtering instruction, including identifying target users, extracting core product categories, listing hard constraints, clarifying soft preference weights, and understanding ranking strategies.

[0019] Based on the core product category, the retrieval interface of the product information management is called to obtain an initial candidate product set that matches the core product category from the product index library, and the product title, detailed description, attribute tags, and user review summary are extracted from the initial candidate product set;

[0020] The product title, detailed description, attribute tags, user review summary, and all elements of the structured recommendation requirements are submitted to the large language model, which is then used to perform hard constraints, semantic association analysis, and soft preference matching degree evaluation.

[0021] The products selected by the large language model are evaluated from multiple dimensions, including: hot topic relevance, user preference matching, product cost-effectiveness and inventory adequacy. Priority scores are calculated based on the multi-dimensional comprehensive evaluation to generate a comprehensive priority ranking list.

[0022] Based on the comprehensive priority ranking list, a correlation analysis is performed between products. When the correlation score is less than a preset threshold, diversity processing is performed based on a preset dimension to avoid homogeneous product recommendations. In addition, a recommendation reason is generated for each finally selected product, and a final product recommendation list with sorted and optimized results and accompanying recommendation reasons is output.

[0023] In this solution, virtual scenario construction and product narrative matching are introduced into the recommendation module based on a large language model, including:

[0024] The structured recommendation requirements are analyzed using a pre-defined large language model. At least one virtual consumption scenario is constructed based on the discrete elements in the recommendation requirements, which matches the hot topic association direction and user preference matching point. Each candidate product in the initial candidate product set is then substituted into the virtual consumption scenario.

[0025] Based on the virtual consumption scenario, a preset large language model is used to generate a narrative evaluation of the candidate products in a specific virtual scenario according to the product title, detailed description, attribute tags, and user review summary of the candidate products, and output the scenario suitability score according to the narrative evaluation.

[0026] The scenario suitability score is combined with the priority score to generate a comprehensive priority ranking list, and a recommendation reason is generated based on the narrative evaluation.

[0027] In this solution, step S104 displays the product recommendation list and iteratively optimizes the recommendation strategy based on feedback data, including:

[0028] In the user interaction module interface, the most prominent position is selected as the core recommendation area to display a preset number of recommended products. In the core recommendation area, each product card uses a combination of images and text to present the main product image, name, current price, promotional tags, and recommendation reasons generated by a preset large language model.

[0029] A real-time feedback option is set for each product card, and a feedback option is also set in the recommendation reason to directly evaluate the quality of the recommendation reason generated by the preset large language model;

[0030] Obtain explicit feedback based on the user's direct expression and implicit feedback based on the user's behavior trajectory. Generate feedback data from the explicit and implicit feedback, and perform short-term adaptation, mid-term model optimization and long-term strategy evolution of the recommendation module based on the feedback data.

[0031] In this solution, the recommendation module undergoes short-term adaptation, medium-term model optimization, and long-term strategy evolution based on feedback data, including:

[0032] In a single user session, the subsequent recommended content is adjusted based on the user's real-time behavior. By continuously listening to all user interaction events, each event is encapsulated into a structured feedback signal, which is then synchronized to the real-time user profile. Based on the feedback signal, the user's short-term interest vector is updated in real time according to preset rules.

[0033] When a user triggers the next request, the recommendation module based on the big language model will first query the updated real-time user profile before generating a new product recommendation list. It will then use the latest short-term interest vector as a strong constraint to re-filter and sort the product recommendation list and output the adjusted new product recommendation list.

[0034] On a daily basis, the recommendation module of the big language-based model is fine-tuned using batch feedback data. Batch user feedback data from the past 24 hours is extracted from the data lake according to a preset timestamp to generate a training dataset. The training data in the training dataset is cleaned and labeled to construct an instruction dataset for fine-tuning. Each data item includes: {old recommendation requests, products rejected by users} as negative examples, and {new recommendation requests, products accepted by users} as positive examples.

[0035] The LoRA algorithm is used to perform lightweight fine-tuning of the large language model using the instruction dataset, learn the user's preference model and platform business rules, and then replace the original recommendation module after verifying the fine-tuned recommendation module.

[0036] On a monthly basis, obtain a comprehensive evaluation report containing monthly data, including: analysis of the traffic-driving effect of trending events, return on investment of different recommendation strategies, and changes in user lifetime value;

[0037] Based on the comprehensive evaluation report, the priority of hot topic capture and filtering is adjusted in the hot topic information collection tool, the ranking weight is adjusted in the recommendation module based on the big language model, and the interaction design is optimized in the user interaction module to achieve global strategy adjustment.

[0038] The second aspect of the present invention provides a product recommendation system based on the collaboration of intelligent agents and hotspot information. The system includes: an intelligent agent scheduling strategy module, a hotspot information collection tool management module, a product information collection tool management module, a user information collection tool management module, a recommendation demand information fusion and generation module, a recommendation module, a prompt template management module, a product information management module, and a user interaction module.

[0039] The intelligent agent scheduling strategy module is responsible for presetting the startup order, execution priority, and data transmission path of each module according to different recommendation scenarios and business needs; monitoring the load of each module in real time, dynamically allocating computing resources, and establishing a module fault detection mechanism to automatically trigger a backup scheduling scheme when a module is abnormal; recording the scheduling log of the entire system and generating scheduling efficiency reports periodically.

[0040] The hotspot information collection tool management module accesses and manages the hotspot data sources, configuring their respective crawling rules and frequencies. It cleans, deduplicates, and classifies the crawled raw hotspot data, extracts structured hotspot feature words, and synchronizes the processed high-quality hotspot information to other modules in real time.

[0041] The product information collection tool management module connects with various product supply platforms to obtain the static attributes and dynamic information of products, verify and clean the product data, mark abnormal data, set attribute classification tags for products, and establish a product data index library for the recommendation module to quickly retrieve.

[0042] The user information collection tool management module collects user static data, dynamic behavior, historical preferences and feedback data, cleans the user static data, desensitizes sensitive information, and continuously updates user profiles.

[0043] The recommendation demand information fusion generation module uses user ID and timestamp as a basis to align multi-source information in time and space. Based on LLM fusion analysis, it uses a large language model to analyze the correlation between hotspots and users, and combines product constraints to perform logical reasoning to output structured recommendation demands that include target user attributes, hotspot association direction, product constraints, user preference matching points, and recommendation priorities.

[0044] The recommendation module understands the various instructions and constraints in the recommendation requirements, retrieves candidate products from the product library, and uses the semantic understanding capabilities of the large language model to deeply filter products that meet the requirements. It also sorts the products comprehensively according to multiple dimensions, generates recommendation reasons for each recommended product, and obtains the final product recommendation list.

[0045] The prompt template management module designs prompt templates for different tasks of the large language model, and continuously iterates and optimizes the template content based on the output effect of the large language model and user feedback.

[0046] The product information management module establishes and maintains a structured product database, provides product query interfaces for other modules, and performs product data statistical analysis.

[0047] The user interaction module displays the final product recommendation list and the reasons for the recommendation to the user through visualization methods, collects user feedback data, and allows users to actively input natural language requirements for customized recommendations.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] This invention enhances the timeliness and contextual adaptability of recommendations. Through a dedicated hot topic information collection module, it proactively and in real-time captures dynamic changes in the external world, including social hotspots, festivals, and popular trends. This ensures that recommendation results are no longer static inferences based on past user behavior, but resonate with the current social atmosphere and market trends. When a hot event occurs, the system can respond quickly, accurately pushing relevant products or services to potentially interested users, thereby effectively seizing fleeting consumption opportunities and significantly improving the attractiveness and conversion potential of recommended content.

[0050] Secondly, it achieves deep cross-domain information fusion and true personalization. Through intelligent agent scheduling, it collaboratively analyzes users' historical preferences, real-time behavior, and external trending information. Leveraging the powerful semantic understanding and logical reasoning capabilities of large language models, the system can deeply interpret the correlation between trending topics and individual users, and make comprehensive decisions by combining real-time product inventory and promotional information. Breaking away from the limitations of traditional recommendation systems that are single-dimensional and superficial in their insights, it can generate highly contextualized recommendation reasons and product lists that truly match users' current potential needs.

[0051] Furthermore, the system possesses self-evolution and continuous optimization capabilities. Every user response to the recommendation results, whether an explicit click or an implicit browsing depth, is captured in real time and synchronized back to the system. This feedback data drives the system's self-iteration at different time scales: in the short term, it instantly adjusts the recommendation strategy to adapt to changes in the user's interests during the current session; in the medium term, it optimizes the core algorithm and suggestion templates to correct model biases; and in the long term, it guides the evolution of global strategies and interaction design. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0053] Figure 1 A flowchart of a product recommendation method based on the collaboration of intelligent agents and hotspot information is shown;

[0054] Figure 2 A flowchart illustrating the multi-source information fusion and structured recommendation requirement generation based on LLM is shown.

[0055] Figure 3 A flowchart illustrating the detailed execution scheme of the LLM-based recommendation module is shown.

[0056] Figure 4A block diagram of a product recommendation system based on the collaboration of intelligent agents and hotspot information is shown. Detailed Implementation

[0057] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0058] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0059] Figure 1 A flowchart of a product recommendation method based on the collaboration of intelligent agents and hotspot information is shown.

[0060] like Figure 1 As shown, this embodiment provides a product recommendation method based on the collaboration of intelligent agents and hotspot information, including:

[0061] S101: Obtain structured hotspot data, product dynamic information, and current user data in parallel through hotspot information collection tools, product information collection tools, and user information collection tools;

[0062] S102: The acquired structured hotspot data, product dynamic information and user data are fused with multi-source information based on a large language model to generate structured recommendation requirements that include target users, hotspot association directions, product constraints, user preference matching points and recommendation priorities;

[0063] S103: Input the generated structured recommendation requirements into the recommendation module based on the large language model, match and filter products that meet the requirements, and then sort them in multiple dimensions according to the closeness of hot topic association, user preference matching degree, product cost performance and inventory sufficiency to generate an optimized product recommendation list.

[0064] S104: The product recommendation list is visualized to the user through the user interaction module, and the user's real-time feedback on the recommendation results is collected. The real-time feedback is used to optimize and iterate the subsequent recommendation strategy.

[0065] It should be noted that in step S101, the hot topic information collection tool captures and processes dynamic hot topic information from the Internet in real time. This tool accesses various types of real-time data sources, including but not limited to trending topics from general search engines, social media platforms, news headlines from news websites, and dynamic information from specific industry vertical websites. Independent crawling parameters are configured for each data source, including crawling frequency, keyword filtering rules, and data format specifications. Social hot topics and industry dynamics are crawled in real time from external search engines and network platforms. The connection status, crawling success rate, and whether anti-crawler mechanisms are triggered are monitored in real time to ensure the continuity and stability of data acquisition. The crawled raw data is preprocessed, including deduplication, noise reduction, and classification. Structured feature words representing the hot topic are extracted. For example, from "This year's Dragon Boat Festival rice dumpling sales skyrocketed," "Dragon Boat Festival," "rice dumpling," and "sales" are extracted, along with positive, negative, and neutral sentiment and popularity values. The processed high-quality, structured hot topic data is synchronized in real time, awaiting subsequent collaborative analysis with product and user information.

[0066] By integrating with APIs provided by various product supply platforms through a product information collection tool, and adapting to different data specifications, the tool retrieves comprehensive product information, including static attributes (product ID, name, category, brand, specifications, material, and image / text description) and dynamic attributes (real-time inventory, current price, information on all ongoing promotions, user ratings, and latest reviews). It obtains dynamic product information related to the user's current search and historical behavior from the product supply platform and generates an initial recommended product pool based on the semantic similarity calculation between trending keywords and product attributes. The product information collection tool does not apply the same update frequency to all products but implements a differentiated strategy based on product characteristics. For example, for fresh produce and digital products with frequent inventory and price fluctuations, a high-frequency (e.g., every minute) update is set; for products with relatively stable information, such as large appliances and furniture, a low-frequency (e.g., daily) update is set, thus optimizing system resource consumption while ensuring real-time information. Strict verification rules are applied to the acquired product data to check its compliance, accuracy, and completeness. When abnormal or incomplete data is detected, a retry mechanism is automatically triggered to re-acquire the data. If multiple attempts fail, the data is marked as untrusted, and the administrator is notified. The verified product data will be categorized and organized according to its attributes (category, brand, price range, applicable scenarios, etc.). At the same time, tags will be set for the products (such as "Dragon Boat Festival Limited Edition", "Low Sugar", "Summer Must-Have", "High Cost Performance"). The tags will be automatically generated by the system based on the product information. All product data will be stored in a structured product database and an efficient index will be created to facilitate subsequent millisecond-level retrieval.

[0067] By using user information collection tools, static basic information (age, gender, geographical location), dynamic behavioral data (real-time conversation behavior, short-term behavior), historical consumption records (long-term order history, browsing history, spending power), and preference tags (preferences explicitly expressed through likes, favorites, and high ratings, and interest tendencies inferred by analyzing long-term behavioral data) of users are collected from user data sources. The collected raw data is integrated into a complete user profile and then cleaned and de-identified to generate current user data, making it possible to generate highly real-time personalized recommendations.

[0068] It should be noted that, as Figure 2 As shown, in step S102, the acquired structured hotspot data, product dynamic information, and current user data are fused using a Large Language Model (LLM) to integrate multi-source information. The structured hotspot data comes from the hotspot information collection tool management module: for example, the hotspot theme is "Dragon Boat Festival," keywords include "zongzi," "gift-giving," "dragon boat," and "health," with a positive sentiment, and includes a popularity value and timestamp; the product dynamic information comes from the product information collection tool management module: for example, real-time inventory status of zongzi products, current promotional activities, price fluctuations, and user review summaries; the current user data comes from the user information collection tool management module: for example, user ID, geographic location, historical behavior, real-time conversation, and explicit preference tags (preference for low-sugar foods, trusted brands). Based on the user's unique identifier and a unified timestamp, the collected structured hotspot data, product dynamic information, and current user data are aligned. A predefined information fusion-specific prompt template from the prompt template management module is then called to integrate the aligned multi-source information and input it into the Large Language Model.

[0069] The large language model understands the semantic relationships between structured trending data and current user data. For example, it might infer that the "Dragon Boat Festival" trending topic is relevant to all users, but the sub-topic under the "low sugar" trending topic is highly relevant to a user with a history of "low sugar" preferences. Conversely, for a user without a fitness habit, the "sports equipment" trending topic might have a weaker relevance. The large language model understands and applies real-time business rules from product dynamic information as hard constraints. For example, it logically excludes products that are out of stock, not participating in promotions, or whose delivery is not supported in the user's region, prioritizing products with sufficient stock, strong promotions, and delivery availability to ensure the feasibility of recommendations. By combining long-term user preferences (e.g., consistently purchasing brand A), short-term behavior (searching for Dragon Boat Festival gifts), and current real-time intent (browsing gifts), it might conclude that the user's Dragon Boat Festival shopping is likely a continuation of their trust in brand A and its low-sugar attributes, while also indicating a desire to give gifts. After completing the above analysis, the large language model does not output a simple list of products, but generates a highly structured, machine-readable description of recommendation requirements. This means generating a structured recommendation requirement that includes target user attributes, hot topic associations, product constraints, user preference matching points, and recommendation priorities.

[0070] It should be noted that, as Figure 3 As shown, in step S103, the generated structured recommendation requirements are input into the recommendation module based on a large language model to generate a product recommendation list. The structured recommendation request is accepted, for example: "Recommend [Zongzi] products for the [Dragon Boat Festival gift] scenario to target user [ID:123, Beijing]. The products must be [in stock, support delivery, and participate in promotions]. Priority should be given to matching the user's [low sugar] preference and the historical preference of [Brand A]. When ranking, [hot topic relevance] and [promotional strength] should be the primary considerations." A pre-trained large language model is used to parse and decompose the structured recommendation request, converting it into retrieval and filtering instructions. This includes identifying the target user, extracting core product categories, listing hard constraints, clarifying soft preference weights, and understanding the ranking strategy. Identifying the target user confirms the service recipient and their geographical location, among other key attributes. Extracting the core product category clarifies that the core recommended product category is Zongzi. A list of hard constraints is generated to satisfy the filter list, including inventory status > 0, delivery area including Beijing, and promotion status being true. Clarifying soft preference weights identifies low sugar and Brand A as priority matching bonuses. Understanding the ranking strategy determines that the results should be primarily ranked according to "relevance to the Dragon Boat Festival hot topic" and "promotional strength."

[0071] Based on the core product category, the retrieval interface of the product information management is invoked to obtain an initial candidate product set that matches the core product category from the product index. The product title, detailed description, attribute tags, and user review summaries from the initial candidate product set are extracted. All elements of the product title, detailed description, attribute tags, user review summaries, and structured recommendation requirements are submitted to the large language model. The large language model performs hard constraint analysis, semantic association analysis, and soft preference matching evaluation, logically excluding all products that do not meet the conditions of "in stock," "supports delivery," or "participates in promotions." This deeply understands the semantic association between product content and recommendation requirements. For example, even if a product's title does not contain the words "Dragon Boat Festival," but the details page describes "Dragon Boat Festival gift box packaging" or user reviews mention "bought as a gift for the Dragon Boat Festival," the LLM can still identify its high relevance to the "Dragon Boat Festival gift-giving" scenario. The matching degree between each candidate product and user soft preferences is evaluated. For example, regarding the preference for "low sugar," the large language model will comprehensively judge whether the product is labeled "low sugar," whether the product name or description emphasizes "sugar-free" or "healthy formula," and whether user reviews mention positive feedback such as "not too sweet." Similarly, it will identify products from "Brand A" and its sub-brands.

[0072] The products selected by the large language model are comprehensively evaluated from multiple dimensions. These dimensions include: hot topic relevance (products with the highest semantic relevance to hot keywords like "Dragon Boat Festival" and "gift-giving" are prioritized), user preference matching (products that perfectly match "low sugar" and are from brand A receive the highest weight; those matching "low sugar" but not from brand A receive the next highest weight), product cost-effectiveness (evaluated based on promotional efforts (discount range), user reviews, recent sales data, etc.), and inventory adequacy (products with sufficient inventory are prioritized to avoid poor user experience due to stockouts). Priority scores are calculated based on these multi-dimensional evaluations to generate a comprehensive priority ranking list. Correlation analysis is performed between products based on this comprehensive priority ranking list. When the correlation score is less than a preset threshold, diversity processing is applied based on preset dimensions, appropriately introducing diversity in categories, brands, and prices to avoid homogeneous product recommendations. Furthermore, recommendation reasons are generated for each finally selected product, and a final product recommendation list, sorted and optimized with accompanying reasons, is output.

[0073] It should be noted that the introduction of virtual scenario construction and product narrative matching into the recommendation module based on a large language model breaks through the traditional attribute-based matching mode. By dynamically constructing a virtual consumption or usage scenario for users and evaluating the suitability and attractiveness of products within this scenario, more imaginative, experiential, and persuasive recommendation results are generated. The pre-set large language model is used to analyze the structured recommendation requirements. Based on the discrete elements in the recommendation requirements, at least one virtual consumption scenario is constructed that matches the trending topics and user preferences. For example, for the requirement of "user [ID:123, Beijing] giving low-sugar rice dumplings as gifts for Dragon Boat Festival," the pre-set large language model constructs the following scenario:

[0074] Scenario A (Business Gift Giving): "A professional living in Beijing wants to choose a tasteful and respectable gift for a business partner who values ​​health during the Dragon Boat Festival. The gift should reflect care and respect, while avoiding the health risks associated with high sugar content."

[0075] Scenario B (Family Reunion): "A Beijing family plans to hold a family gathering for the Dragon Boat Festival. There is an elder in the family who loves traditional festivals but needs to control their blood sugar. They need to choose a healthy food that retains the traditional flavor of the festival (zongzi) and is safe for the elder to eat as the centerpiece of the banquet."

[0076] Each candidate product in the initial candidate product set is substituted into the virtual consumption scenario. Based on the virtual consumption scenario, a preset large language model is used to generate a narrative evaluation of the candidate product in a specific virtual scenario, according to the product title, detailed description, attribute tags, and user review summary. For example, for a "premium black truffle low-sugar rice dumpling gift box," the large language model's evaluation in scenario A might be: "The product packaging is luxurious and elegant, the black truffle ingredient highlights its prestige, and the low-sugar formula precisely matches the recipient's pursuit of health. As a business gift, it can not only demonstrate the giver's sincerity and taste but also express concern for the health of partners, which is very suitable for the scenario." The evaluation of the same product in scenario B might be: "Although the product is high-end, the black truffle flavor may be too unique and not suitable for the general taste of family gatherings. Moreover, its expensive price is more suitable for gift scenarios than for family use, and its fit with this warm family scenario is average." By combining the narrative evaluations of each product in different scenarios, a scenario suitability score is output based on the narrative evaluation. The scenario suitability score is combined with the priority score to generate a comprehensive priority ranking list, and recommendation reasons are generated based on the narrative evaluation.

[0077] It should be noted that in step S104, the product recommendation list is displayed, and the recommendation strategy is iteratively optimized through feedback data. A core recommendation area, located in the most prominent position within the user interaction module interface, displays a preset number of recommended products. Each product card in this core area uses a combination of images and text, presenting the product's main image, name, current price, promotional tags, and a recommendation reason generated by a preset large language model. Below the core recommendation area, more alternative products are displayed in a grid or flow layout. This area reflects the diversity of recommendations, including products from different brands, price points, or with different features, providing users with space for comparison and exploration. Based on the recommendation logic, products are grouped for display. For example, they are divided into tag groups such as "Selected Gifts for Dragon Boat Festival," "Healthy and Low-Sugar Choices," and "Brand A You Trust," helping users understand the logic behind the recommendations. Real-time feedback options are set for each product card, such as "Like / Thumbs Up," "Not Interested," "Add to Favorites," "Add to Cart," and "Buy Now." The system includes feedback options such as "reason is accurate" and "reason is inconsistent" in the recommendation reasons to directly evaluate the quality of the recommendation reasons generated by the preset large language model. The interface retains an active demand input channel, allowing users to actively input new demands in natural language descriptions, which directly incorporates users into the starting point of the recommendation loop, greatly improving the system's flexibility and user control.

[0078] Explicit feedback is obtained from direct user input, including actions such as clicking "Like," "Not Interested," and "Reason Does Not Match." Implicit feedback is also obtained from user behavior patterns, including the duration of time spent on a recommended product, whether the user clicks to view product details, whether the product is added to the cart and then removed, and whether a purchase is ultimately made. This explicit and implicit feedback is used to generate feedback data, which is then used for short-term adaptation, mid-term model optimization, and long-term strategy evolution of the recommendation module.

[0079] It should be noted that the recommendation module undergoes short-term adaptation, mid-term model optimization, and long-term strategy evolution based on feedback data. Within a single user session, subsequent recommendations are adjusted based on the user's real-time behavior. Throughout a user session, the system can instantly adjust subsequent recommendations based on real-time user feedback. For example, if a user clicks "not interested" on several dessert recommendations, the system will immediately reduce the number of desserts in subsequent recommendations and increase the number of other product categories.

[0080] By continuously monitoring all user interaction events, including clicks (likes, disinterest), browsing duration, adding to cart, placing orders, and entering new queries, each event is encapsulated as a structured feedback signal. This feedback signal is synchronized to the real-time user profile, and the user's short-term interest vector is updated in real-time according to preset rules based on the feedback signal. For example, clicking "like" significantly increases the weight of the product's category, brand, and attributes; clicking "disinterest" immediately reduces the weight of related features and filters out similar products in the current session; prolonged browsing moderately increases the weight of related interest points; entering a new query uses the query term as a strong interest signal for the current session, resetting or dominating the short-term interest direction. When the user triggers the next request, the recommendation module based on the large language model queries the updated real-time user profile before generating a new product recommendation list. Using the latest short-term interest vector as a strong constraint, it re-filters and sorts the product recommendation list, outputting an adjusted new product recommendation list, completing one real-time adaptive loop.

[0081] On a daily basis, batch feedback data is used to fine-tune the recommendation module based on the big language model, correcting model biases and improving overall recommendation accuracy. For example, if it is found that the big language model frequently and incorrectly associates a certain type of product with a certain trending topic (such as users generally expressing no interest in "recommended electronic products for the Dragon Boat Festival"), the backend can use this feedback data to fine-tune the LLM and correct its comprehension biases.

[0082] Based on a preset timestamp, batch user feedback data from the past 24 hours is extracted from the data lake to generate a training dataset. The training data in the training dataset is cleaned and labeled to construct an instruction dataset for fine-tuning. Each data entry includes: {old recommendation requests, products rejected by the user} as negative examples, and {new recommendation requests, products accepted by the user} as positive examples. The instruction dataset is used to perform lightweight fine-tuning of the large language model based on the LoRA algorithm, learning the user's preference model and platform business rules. After the fine-tuned recommendation module is verified, it replaces the original recommendation module.

[0083] On a monthly basis, conduct macro-level strategic reviews and adjustments to optimize the entire recommendation system's operational strategies, algorithm weights, and interaction design. If the conversion rate from certain trending topics remains consistently low, adjust the filtering rules of the trending topic information collection tool or the association weight between trending topics and products. If the "user reviews" dimension has a greater impact on purchase decisions than expected, increase its weight in the ranking strategy of the large language model. By analyzing the heatmap and flow of user interactions with the interface, a high click-through rate was found in the "Recommendation Reasons" display area, which should be strengthened in the next UI redesign.

[0084] Obtain a comprehensive evaluation report covering monthly data, including: analysis of the traffic-driving effect of trending events, return on investment (ROI) of different recommendation strategies, and changes in user lifetime value; discover deeper trends and insights through data mining. For example: Insight 1: User dwell time and average order value brought by health and diet-related trending topics are consistently higher than those brought by entertainment-related trending topics. Insight 2: Products with displayed recommendation reasons have significantly higher trust levels and conversion rates than products that are not displayed.

[0085] Based on the comprehensive evaluation report, the priority of hot topic capture and filtering in the hot topic information collection tool was adjusted, increasing the collection weight of long-term value hot topics such as "health," "wellness," and "fitness," while reducing resource investment in low-value hot topics. In the recommendation module based on a large language model, the ranking weight was adjusted, globally increasing the weight of "user rating score" and "inventory turnover rate," while decreasing the weight of "promotional intensity," optimizing the product ecosystem with the platform's long-term interests at its core. In the user interaction module, the interaction design was optimized, implementing a global strategy adjustment, strengthening the visual display of "recommendation reasons," and designing more feedback entry points to further collect data and optimize the quality of reason generation.

[0086] Figure 4 The diagram illustrates the architecture of a product recommendation system based on the collaboration between intelligent agents and hotspot information.

[0087] The second embodiment of the present invention provides a product recommendation system based on the collaboration of intelligent agents and hotspot information. The system includes: an intelligent agent scheduling strategy module, a hotspot information collection tool management module, a product information collection tool management module, a user information collection tool management module, a recommendation demand information fusion and generation module, a recommendation module, a prompt template management module, a product information management module, and a user interaction module.

[0088] The intelligent agent scheduling strategy module is responsible for presetting the startup order, execution priority, and data transmission path of each module according to different recommendation scenarios and business needs; monitoring the load of each module in real time, dynamically allocating computing resources, and establishing a module fault detection mechanism to automatically trigger a backup scheduling scheme when a module is abnormal, ensuring that the recommendation service is not interrupted; recording the scheduling log of the entire system and generating scheduling efficiency reports periodically.

[0089] The hotspot information collection tool management module accesses and manages the hotspot data sources, configuring their respective crawling rules and frequencies. It cleans, deduplicates, and classifies the crawled raw hotspot data, extracts structured hotspot feature words, and synchronizes the processed high-quality hotspot information to other modules in real time.

[0090] The product information collection tool management module connects with various product supply platforms to obtain the static attributes and dynamic information of products, verify and clean the product data, mark abnormal data such as price errors and inaccurate inventory, set attribute classification tags for products such as Dragon Boat Festival limited edition and high cost performance, and establish a product data index library for the recommendation module to quickly retrieve.

[0091] The user information collection tool management module collects user static data, dynamic behavior, historical preferences and feedback data, cleans the user static data, removes invalid records, desensitizes sensitive information to ensure privacy and security, and continuously updates user profiles to ensure that the latest user behavior can be reflected in the profile immediately, providing support for real-time recommendations.

[0092] The recommendation demand information fusion generation module uses user ID and timestamp as a basis to align multi-source information in time and space to form a consistent context; the LLM-based fusion analysis uses a large language model to analyze the correlation between hotspots and users, and combines product constraints to perform logical reasoning, outputting structured recommendation demands that include target user attributes, hotspot association directions, product constraints, user preference matching points, and recommendation priorities.

[0093] The recommendation module understands the various instructions and constraints in the recommendation requirements, retrieves candidate products from the product library, and uses the semantic understanding capabilities of the large language model to deeply filter products that meet the requirements. It also sorts the products comprehensively according to multiple dimensions, generates recommendation reasons for each recommended product, and obtains the final product recommendation list.

[0094] The prompt template management module designs prompt templates for different tasks of the large language model, and continuously iterates and optimizes the template content based on the output effect of the large language model and user feedback to improve the accuracy and stability of the model performance; it manages the version history of the templates, controls modification permissions, and ensures system stability; it automatically matches and pushes the corresponding prompt template to the LLM according to the current task of the system to ensure that the LLM can process tasks according to a unified standard.

[0095] The product information management module establishes and maintains a structured product database, regularly cleans up invalid data, provides product query interfaces for other modules, supports multi-condition, fuzzy, and tag-based searches, and performs statistical analysis of product data to generate product popularity lists, slow-moving product lists, etc., providing data insights for operation and recommendation strategies.

[0096] The user interaction module uses a visual approach to present the final product recommendation list and reasons for recommendation to users in a clear and organized manner with both text and images. It provides convenient access for users to provide real-time feedback, records implicit user behavior throughout the process, collects user feedback data, and allows users to actively input natural language requests for customized recommendations. In addition, based on user habits, such as the preference for a concise display on mobile devices and detailed information on PC users, the layout of the interactive interface is adjusted; font size and operation buttons are optimized for specific user groups; the loading speed and smoothness of the interactive interface are monitored in real time, and issues such as lag and crashes are promptly fixed to improve the user experience.

[0097] A third embodiment of the present invention provides a computer-readable storage medium, which includes a product recommendation method program based on the collaboration of intelligent agents and hotspot information. When the product recommendation method program based on the collaboration of intelligent agents and hotspot information is executed by a processor, it implements the steps of the product recommendation method based on the collaboration of intelligent agents and hotspot information.

[0098] In the several embodiments provided in this application, it should be understood that the disclosed methods and systems can be implemented in other ways. The system embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, and can be electrical, mechanical, or other forms. Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0099] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0100] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A product recommendation method based on the collaboration of intelligent agents and hotspot information, characterized in that, Includes the following steps: S101: Obtain structured hotspot data, product dynamic information, and current user data in parallel through hotspot information collection tools, product information collection tools, and user information collection tools; S102: The acquired structured hotspot data, product dynamic information and user data are fused with multi-source information based on a large language model to generate structured recommendation requirements that include target users, hotspot association directions, product constraints, user preference matching points and recommendation priorities; S103: Input the generated structured recommendation requirements into the recommendation module based on the large language model, match and filter products that meet the requirements, and then sort them in multiple dimensions according to the closeness of hot topic association, user preference matching degree, product cost performance and inventory sufficiency to generate an optimized product recommendation list. S104: The product recommendation list is visualized to the user through the user interaction module, and the user's real-time feedback behavior on the recommendation results is collected. The real-time feedback behavior is used to optimize and iterate the subsequent recommendation strategy. In step S101, structured hotspot data, product dynamic information, and current user data are acquired in parallel using hotspot information collection tools, product information collection tools, and user information collection tools, including: By using hot topic information collection tools, we can capture social hot topics and industry dynamics information in real time from external search engines and online platforms, extract themes and analyze sentiment trends to obtain structured hot topic feature words and generate structured hot topic data. By using product information collection tools, dynamic product information related to the user's current search and historical behavior is obtained from the product supply platform, and a preliminary recommended product pool is generated based on the semantic similarity calculation of hot feature words and product attributes. By using user information collection tools, static basic information, dynamic behavioral data, historical consumption records and preference tags of users are collected from user data sources. The collected raw data is integrated into a complete user profile and then cleaned and de-identified to generate current user data.

2. The product recommendation method based on the collaboration of intelligent agents and hotspot information according to claim 1, characterized in that, In step S102, the acquired structured hotspot data, product dynamic information, and current user data are fused using a large language model to perform multi-source information fusion, including: The collected structured hotspot data, product dynamic information and current user data are aligned, and the predefined information fusion prompt template in the prompt template management module is called to integrate the aligned multi-source information and input it into the large language model; By understanding the semantic relationship between structured hotspot data and current user data through the large language model, and understanding and applying real-time business rules in product dynamic information as hard constraints, the system combines users' long-term historical preferences, short-term behaviors and current real-time intentions to generate structured recommendation requirements that include target user attributes, hotspot association directions, product constraints, user preference matching points and recommendation priorities.

3. The product recommendation method based on the collaboration of intelligent agents and hotspot information according to claim 1, characterized in that, In step S103, the generated structured recommendation requirements are input into the recommendation module based on a large language model to generate a product recommendation list, including: The structured recommendation request is accepted, and the pre-trained large language model is used to parse and decompose the structured recommendation request into a search and filtering instruction, including identifying target users, extracting core product categories, listing hard constraints, clarifying soft preference weights, and understanding ranking strategies. Based on the core product category, the retrieval interface of the product information management is called to obtain an initial candidate product set that matches the core product category from the product index library, and the product title, detailed description, attribute tags, and user review summary are extracted from the initial candidate product set; The product title, detailed description, attribute tags, user review summary, and all elements of the structured recommendation requirements are submitted to the large language model, which is then used to perform hard constraints, semantic association analysis, and soft preference matching degree evaluation. The products selected by the large language model are evaluated from multiple dimensions, including: hot topic relevance, user preference matching, product cost-effectiveness and inventory adequacy. Priority scores are calculated based on the multi-dimensional comprehensive evaluation to generate a comprehensive priority ranking list. Based on the comprehensive priority ranking list, a correlation analysis is performed between products. When the correlation score is less than a preset threshold, diversity processing is performed based on a preset dimension to avoid homogeneous product recommendations. In addition, a recommendation reason is generated for each finally selected product, and a final product recommendation list with sorted and optimized results and recommendation reasons is output.

4. The product recommendation method based on the collaboration of intelligent agents and hotspot information according to claim 3, characterized in that, The recommendation module based on a large language model introduces virtual scenario construction and product narrative matching, including: The structured recommendation requirements are analyzed using a pre-defined large language model. At least one virtual consumption scenario is constructed based on the discrete elements in the recommendation requirements, which matches the hot topic association direction and user preference matching point. Each candidate product in the initial candidate product set is then substituted into the virtual consumption scenario. Based on the virtual consumption scenario, a preset large language model is used to generate a narrative evaluation of the candidate products in a specific virtual scenario according to the product title, detailed description, attribute tags, and user review summary of the candidate products, and output the scenario suitability score according to the narrative evaluation. The scenario suitability score is combined with the priority score to generate a comprehensive priority ranking list, and a recommendation reason is generated based on the narrative evaluation.

5. The product recommendation method based on the collaboration of intelligent agents and hotspot information according to claim 1, characterized in that, In step S104, the recommended product list is displayed, and the recommendation strategy is iteratively optimized based on feedback data, including: In the user interaction module interface, the most prominent position is selected as the core recommendation area to display a preset number of recommended products. In the core recommendation area, each product card uses a combination of images and text to present the main product image, name, current price, promotional tags, and recommendation reasons generated by a preset large language model. A real-time feedback option is set for each product card, and a feedback option is also set in the recommendation reason to directly evaluate the quality of the recommendation reason generated by the preset large language model; Obtain explicit feedback based on the user's direct expression and implicit feedback based on the user's behavior trajectory. Generate feedback data from the explicit and implicit feedback, and perform short-term adaptation, mid-term model optimization and long-term strategy evolution of the recommendation module based on the feedback data.

6. The product recommendation method based on the collaboration of intelligent agents and hotspot information according to claim 5, characterized in that, Based on feedback data, the recommendation module undergoes short-term adaptation, medium-term model optimization, and long-term strategy evolution, including: In a single user session, the subsequent recommended content is adjusted based on the user's real-time behavior. By continuously listening to all user interaction events, each event is encapsulated into a structured feedback signal, which is then synchronized to the real-time user profile. Based on the feedback signal, the user's short-term interest vector is updated in real time according to preset rules. When a user triggers the next request, the recommendation module based on the large language model will first query the updated real-time user profile before generating a new product recommendation list. It will then use the latest short-term interest vector as a strong constraint to re-filter and sort the product recommendation list and output the adjusted new product recommendation list. On a daily basis, the recommendation module based on the large language model is fine-tuned using batch feedback data. Batch user feedback data from the past 24 hours is extracted from the data lake according to a preset timestamp to generate a training dataset. The training data in the training dataset is cleaned and labeled to construct an instruction dataset for fine-tuning. Each data item includes: {old recommendation requests, products rejected by users} as negative examples, and {new recommendation requests, products accepted by users} as positive examples. The LoRA algorithm is used to perform lightweight fine-tuning of the large language model using the instruction dataset, learn the user's preference model and platform business rules, and then replace the original recommendation module after verifying the fine-tuned recommendation module. On a monthly basis, obtain a comprehensive evaluation report containing monthly data, including: analysis of the traffic-driving effect of trending events, return on investment of different recommendation strategies, and changes in user lifetime value; Based on the comprehensive evaluation report, the priority of hot topic capture and filtering is adjusted in the hot topic information collection tool, the ranking weight is adjusted in the recommendation module based on the large language model, and the interaction design is optimized in the user interaction module to achieve global strategy adjustment.

7. A product recommendation system based on the collaboration of intelligent agents and hotspot information, characterized in that, To implement the product recommendation method based on the collaboration of intelligent agents and hotspot information as described in any one of claims 1-6, the system includes: an intelligent agent scheduling strategy module, a hotspot information collection tool management module, a product information collection tool management module, a user information collection tool management module, a recommendation demand information fusion and generation module, a recommendation module, a prompt template management module, a product information management module, and a user interaction module; The intelligent agent scheduling strategy module is responsible for presetting the startup order, execution priority, and data transmission path of each module according to different recommendation scenarios and business needs; monitoring the load of each module in real time, dynamically allocating computing resources, and establishing a module fault detection mechanism to automatically trigger a backup scheduling scheme when a module is abnormal; recording the scheduling log of the entire system and generating scheduling efficiency reports periodically. The hotspot information collection tool management module accesses and manages the hotspot data sources, configuring their respective crawling rules and frequencies. It cleans, deduplicates, and classifies the crawled raw hotspot data, extracts structured hotspot feature words, and synchronizes the processed high-quality hotspot information to other modules in real time. The product information collection tool management module connects with various product supply platforms to obtain the static attributes and dynamic information of products, verify and clean the product data, mark abnormal data, set attribute classification tags for products, and establish a product data index library for the recommendation module to quickly retrieve. The user information collection tool management module collects user static data, dynamic behavior, historical preferences and feedback data, cleans the user static data, desensitizes sensitive information, and continuously updates user profiles. The recommendation demand information fusion generation module uses user ID and timestamp as a basis to align multi-source information in time and space. Based on LLM fusion analysis, it uses a large language model to analyze the correlation between hotspots and users, and combines product constraints to perform logical reasoning to output structured recommendation demands that include target user attributes, hotspot association direction, product constraints, user preference matching points, and recommendation priorities. The recommendation module understands the various instructions and constraints in the recommendation requirements, retrieves candidate products from the product library, and uses the semantic understanding capabilities of the large language model to deeply filter products that meet the requirements. It also sorts the products comprehensively according to multiple dimensions, generates recommendation reasons for each recommended product, and obtains the final product recommendation list. The prompt template management module designs prompt templates for different tasks of the large language model, and continuously iterates and optimizes the template content based on the output effect of the large language model and user feedback. The product information management module establishes and maintains a structured product database, provides product query interfaces for other modules, and performs product data statistical analysis. The user interaction module displays the final product recommendation list and the reasons for the recommendation to the user through visualization methods, collects user feedback data, and allows users to actively input natural language requirements for customized recommendations.

Images