Cross-platform user consumption preference real-time prediction and recommendation system

Abstract

The invention discloses a cross-platform user consumption preference real-time prediction and recommendation system, and relates to the technical field of e-commerce, and the system comprises a comprehensive recommendation management platform which is in communication connection with the following modules: a user portrait construction module which is used for collecting basic information of a new user, the entity relationship and the attribute information in the knowledge graph are utilized to construct a user portrait; and the commodity knowledge graph construction module is used for collecting and sorting cross-platform commodity information and constructing a commodity knowledge graph. According to the method, hidden demands of users are effectively mined and long-tail commodities are found and recommended by utilizing meta-path reasoning and cross-domain knowledge migration technologies of the knowledge graph, so that the exposure ratio of hot commodities to the long-tail commodities is balanced, excessive recommendation of the hot commodities is avoided, the long-tail commodities are ensured to obtain proper display opportunities, and the user experience is improved. The commodity diversity of the platform can be improved, sales of commodities of small people can be promoted, and long-term development of the platform is facilitated.

Description

Technical Field

[0001] This invention relates to the field of e-commerce technology, specifically to a cross-platform real-time prediction and recommendation system for user consumption preferences. Background Technology

[0002] With the rapid development of the internet and smart mobile devices, users are increasingly active across multiple platforms, including shopping, social networking, entertainment, and news reading. Traditional single-platform recommendation systems can no longer meet the diverse needs of users. Cross-platform recommendation systems can integrate user behavior data from different platforms to provide more comprehensive and accurate personalized services.

[0003] In existing technologies, for cold-start users (new users), it is difficult to accurately predict their consumption preferences due to a lack of sufficient behavioral data. For long-tail products, due to their low frequency of appearance in the data, recommendation systems are prone to the "Matthew effect," where popular products are over-recommended, leading to a poor recommendation experience for users, reduced platform stickiness, and impact on platform diversity. This is detrimental to the sales of niche products and the long-term development of the platform. Therefore, how to utilize meta-path reasoning and cross-domain knowledge transfer of knowledge graphs to reduce reliance on interaction data and discover users' implicit needs, thereby improving the accuracy of user consumption preference prediction and recommendation, is the problem that this invention aims to solve. To this end, a cross-platform real-time prediction and recommendation system for user consumption preferences is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a cross-platform real-time prediction and recommendation system for user consumption preferences, so as to solve the problems mentioned in the background art.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A cross-platform real-time prediction and recommendation system for user consumption preferences includes a comprehensive recommendation management platform, which is communicatively connected to the following modules: The user profile building module is used to collect basic information about new users and build user profiles using entity relationships and attribute information in the knowledge graph. The product knowledge graph construction module is used to collect and organize product information across platforms, including product attributes, categories, brands, user reviews, etc., to build a product knowledge graph. The meta-path reasoning module is used to define meta-paths in the product knowledge graph. It calculates the probability of association between users and products through path reasoning algorithms, uncovers implicit needs, discovers long-tail products, and avoids over-recommendation of popular products. The preference prediction and recommendation module is used to combine real-time user behavior with the reasoning results of the knowledge graph to generate a personalized recommendation list; The feedback optimization module is used to collect explicit and implicit user feedback and feed it back into the user profile and knowledge graph to update the knowledge graph.

[0006] A further improvement of the technical solution of the present invention is that the user profile construction module includes a basic information collection unit and an auxiliary profile unit; The basic information collection unit is used to collect basic information about new users and provide a basic profile framework for them. The auxiliary profiling unit is used to combine new user basic information and platform product categories to infer the user's potential areas of interest and construct a user profile.

[0007] A further improvement to the technical solution of the present invention is that the auxiliary portrait unit includes: Extract existing product categories from the platform, perform correlation analysis between new user basic information and platform product categories, and preliminarily filter out product categories that match new user basic information based on preset filtering rules in the correlation rule library; Based on the initially screened product categories, potential areas of interest for users are inferred, and these inferred potential areas of interest are integrated with the basic information of new users to form a complete user profile.

[0008] A further improvement to the technical solution of the present invention is that the auxiliary portrait unit includes: Extract existing product categories from the platform, perform correlation analysis between new user basic information and platform product categories, and preliminarily filter out product categories that match new user basic information based on preset filtering rules in the correlation rule library; Based on the initially screened product categories, potential areas of interest for users are inferred, and these inferred potential areas of interest are integrated with the basic information of new users to form a complete user profile.

[0009] A further improvement to the technical solution of the present invention is that the product knowledge graph construction module includes: Product information, including product attributes, categories, brands, and user reviews, is collected from multiple channels covering different e-commerce platforms, merchant-owned systems, and public data sources. The collected data is cleaned and standardized to remove duplicate and invalid information, ensuring data integrity and consistency. At the same time, data from different sources is integrated into a unified data format to form a structured preliminary dataset. Based on the integrated product information, product entities and their attributes are defined, the relationships between products are established, and the product information is transformed into a graph structure through entity recognition and relationship extraction technologies, forming a product knowledge graph containing nodes and edges. We continuously monitor changes in product information, regularly update the data in the product knowledge graph to ensure its timeliness and accuracy, and optimize the structure and content of the knowledge graph based on user feedback and market dynamics, adding new product entities and relationships, and adjusting the strength of existing associations.

[0010] A further improvement to the technical solution of the present invention is that the meta-path reasoning module includes: Based on the structure and business logic of the product knowledge graph, we determine the semantically meaningful meta-path patterns, preprocess the product knowledge graph, extract node features and edge information, and construct an index structure adapted to path reasoning. Using a path reasoning algorithm, the product knowledge graph is traversed and searched along the defined meta-path. Combining node features and edge weights, the probability of association between users and each product is calculated. By analyzing the probability distribution, potential latent needs of users are mined, and long-tail products that are not popular but match the potential interests of users are identified. Based on the calculated association probability, the exposure ratio of popular products and long-tail products in the recommendation results is adjusted to avoid over-recommending popular products and ensure that long-tail products receive appropriate display opportunities.

[0011] A further improvement of the technical solution of the present invention is that the preference prediction and recommendation module includes a real-time preference prediction unit and a dynamic recommendation generation unit; The real-time preference prediction unit is used to integrate the reasoning results of the product knowledge graph with the user's real-time behavior, and predict the user's immediate preferences through weighted fusion, thereby improving the accuracy of cold-start user recommendations and adapting to user preference drift. The dynamic recommendation generation unit is used to generate a candidate set based on user profiles, user real-time preferences, and the upper limit of the number of product categories to be recommended, and to dynamically adjust the recommendation strategy using reinforcement learning to balance popular and long-tail products, thereby generating a personalized recommendation list.

[0012] A further improvement to the technical solution of the present invention is that the real-time preference prediction unit includes: Data is acquired from multiple channels, and real-time user behavior data, including browsing, clicking, and dwell time, is collected. At the same time, product knowledge graph inference results are obtained, and the real-time user behavior data and knowledge graph inference results are integrated to form a comprehensive dataset. Based on pre-set weighting rules, the product knowledge graph reasoning results are weighted and fused with real-time user behavior data to generate a feature vector that comprehensively reflects the user's immediate preferences, thereby improving the accuracy of cold-start user recommendations. By using the fused feature vectors and combining them with a preference prediction model pre-trained based on machine learning, the system predicts users' real-time preferences, outputs the degree of user preference for various types of goods, and applies the prediction results to the recommendation system.

[0013] A further improvement to the technical solution of the present invention is that the dynamic recommendation generation unit includes: Based on user profiles, real-time user preferences, and a preset upper limit for the number of product categories to be recommended, a set of products that match user interests and recommendation rules is selected to form an initial candidate set. In this set, products that match user interests are selected from the product knowledge graph by combining basic information and real-time preference data from the user profile. The size of the candidate set is controlled according to the upper limit for the number of product categories to be recommended, so as to ensure the diversity and relevance of the recommended content. The recommendation strategy is dynamically adjusted using reinforcement learning algorithms to balance popular and long-tail products. By defining a reward mechanism, the recommendation strategy is optimized in real time based on user feedback, adjusting the recommendation ratio of popular and long-tail products to improve the accuracy of recommendations and user satisfaction. Based on the adjusted recommendation strategy, the recommendation results are sorted according to the degree of user preference, and then a final personalized recommendation list is generated from the candidate set to ensure that the recommended content not only matches the user's current interests, but also guides the user to discover new products.

[0014] A further improvement to the technical solution of the present invention is that the feedback optimization module includes: The system collects user feedback information, including explicit and implicit feedback, through front-end interaction and back-end logging. For explicit feedback, the platform interface is set up with function entry points such as rating, comment, like and favorite to obtain users' clear attitudes toward products or services. At the same time, data analysis technology is used to capture user behavior data such as browsing, clicking, dwell time and purchase conversion as implicit feedback. The collected feedback data is integrated into user profiles and knowledge graphs. Explicit feedback directly updates the preference information in user profiles, while implicit feedback adjusts the node attributes and edge weights in the knowledge graph by analyzing behavioral patterns, thereby enhancing the accuracy and dynamism of the knowledge graph. The knowledge graph is updated based on feedback data, the information of nodes and edges is optimized, the relationships between products are adjusted, new entities and relationships are added, or the strength of existing relationships is adjusted. Through continuous iteration, the quality of the knowledge graph is improved.

[0015] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows: 1. This invention provides a cross-platform user consumption preference real-time prediction and recommendation system. By utilizing meta-path reasoning and cross-domain knowledge transfer technology of knowledge graphs, it effectively mines users' implicit needs, discovers and recommends long-tail products, and balances the exposure ratio of popular products and long-tail products. This avoids over-recommendation of popular products and ensures that long-tail products also get appropriate display opportunities. This not only helps to improve the product diversity of the platform, but also promotes the sales of niche products, which is beneficial to the long-term development of the platform.

[0016] 2. This invention provides a cross-platform real-time prediction and recommendation system for user consumption preferences. By combining real-time user behavior data with the inference results of knowledge graphs, it comprehensively reflects the feature vectors of users' immediate preferences. It can fully utilize the advantages of multi-source data, improve the accuracy of cold-start user recommendations, and quickly adapt to user preference drift. At the same time, it optimizes the recommendation strategy based on real-time user feedback, further improving the accuracy and efficiency of recommendations. It can more effectively meet users' personalized needs and improve the overall performance of the recommendation system. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a schematic diagram of data flow in this invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings.

[0019] Example 1, such as Figure 1 , Figure 2 As shown, this invention provides a cross-platform user consumption preference real-time prediction and recommendation system, including a comprehensive recommendation management platform. The comprehensive recommendation management platform has the following modules connected for communication: The user profile building module is used to collect basic information of new users and build user profiles using entity relationships and attribute information in the knowledge graph. The user profile building module includes a basic information collection unit and an auxiliary profile unit. The basic information collection unit is used to collect basic information of new users, providing a basic profile framework for them. Based on system positioning and business needs, it clarifies the scope of basic information to be collected and collects basic information of new users through the user registration interface, including the user's age, gender, occupation, and geographical location. It ensures the legality of information collection, obtains user authorization, verifies the collected basic information of new users, checks the completeness of the information, and verifies the correctness of the format. For information that does not meet the requirements, it prompts the user to re-enter or supplement it. The verified basic information of new users is stored in the user profile database to form the basic profile framework of the user. The storage process is classified and stored according to a preset data structure. The specific tasks of the basic information collection unit are as follows: Based on the system's positioning and business needs, the scope of basic information to be collected from new users is defined, covering factors such as age, gender, occupation, and geographic location. User age is used to understand the user's lifecycle stage and predict spending power and potential demand tendencies. Gender information is used to grasp the differences in consumption preferences between different genders. Occupation information reflects the user's income level, consumption scenarios, and consumption habits. Geographic location information is used to assist in analyzing the consumption environment, cultural characteristics, and logistics and delivery factors of the user's region. The registration interface clearly displays the information fields that users need to fill in, adding clear prompts to each field explaining the filling requirements. A privacy policy and user authorization agreement are also provided. The system should clearly inform users of the purpose and scope of information use and obtain their explicit authorization. After users submit their registration information, the system should verify the information they fill in on the registration interface to ensure its completeness and correct format. Specific verification content includes completeness checks and format checks, verifying that each required field has been filled in to ensure that no key information is missing. If any required fields are found to be missing, the user should be notified and prompted to complete the relevant information. Based on the system's business requirements, the system should design the data structure of the user profile database in advance, classifying and planning according to different information categories. The basic information of verified new users should be stored in the user profile database according to the preset data structure to form the basic user profile framework. The auxiliary profiling unit is used to combine new user basic information and platform product categories to infer the user's potential interest areas, build a user profile, extract existing product categories on the platform, perform correlation analysis between new user basic information and platform product categories, and initially filter out product categories that match the new user basic information according to preset filtering rules in the association rule base. Based on the initially filtered product categories, it infers the user's potential interest areas and integrates the inferred potential interest areas with the new user basic information to form a complete user profile. The specific tasks of the auxiliary portrait unit are as follows: The platform's product database is queried, and all existing product categories are extracted. Then, the basic information of new users is correlated with the platform's product categories to classify them into age ranges. Based on these age ranges, the product categories of interest are inferred. For example, teenagers (13-18 years old) may be interested in electronic products (such as game consoles and headphones), trendy clothing, and books (teen literature); young adults (19-35 years old) may be interested in electronic products (such as smartphones and computers), fashion clothing, beauty and skincare products, and outdoor sports equipment; and middle-aged adults (36-55 years old) may be interested in household goods, health foods, and high-end electronic products (such as...). Users are interested in smart home devices and books (business management, health and wellness). Older users (55 years and older) may be interested in health foods, supplements, casual clothing, and cultural products (such as calligraphy supplies). Furthermore, based on the user's gender, their interests are inferred to be related to gender; for example, men may be interested in electronics, sporting goods, car parts, and tools, while women may be interested in clothing, beauty and skincare products, fashion accessories, and home furnishings. Based on the user's occupation, their interests are inferred to be related to their profession; for example, programmers may be interested in electronics (computers, keyboards, mice), office supplies, and professional books (programming books). Teachers might be interested in educational supplies, books (educational books), and office supplies; freelancers might be interested in creative products, office supplies, and home furnishings (related to a comfortable office environment). Based on the user's geographical location, the system infers the product categories related to their region of interest. For example, first-tier cities might be interested in high-end electronics, fashion brand clothing, and imported food; second-tier cities might be interested in mid-to-high-end electronics, local specialty foods, and home furnishings; and third-tier and lower-tier cities might be interested in cost-effective electronics, daily necessities, and local specialty foods. A filtering rule base is established, containing the mapping relationship between different basic information and product categories. Based on the preset filtering rules in the rule base, product categories matching the new user's basic information are initially filtered out. These initially filtered product categories are then expanded and refined to infer the user's potential areas of interest. For example, if the initial filtering of electronics reveals interest in smart home devices and audio equipment, or if the initial filtering of clothing reveals interest in fashion brands and sportswear, the system integrates the inferred potential areas of interest with the new user's basic information to form a complete user profile, which is then stored in a user profile database. The product knowledge graph construction module is used to collect and organize product information across platforms, including product attributes, categories, brands, user reviews, etc., to build a product knowledge graph. This provides a comprehensive product knowledge foundation for the recommendation system, ensuring the diversity and accuracy of recommendation results. Product information, including product attributes, categories, brands, and user reviews, is collected from multiple channels covering different e-commerce platforms, merchant-owned systems, and public data sources. The collected data is cleaned and standardized to remove duplicate and invalid information, ensuring data integrity and consistency. Simultaneously, data from different sources is integrated into a unified data format to form a structured preliminary dataset. Based on the integrated product information, product entities and their attributes are defined, and relationships between products are established. Through entity recognition and relationship extraction technologies, product information is transformed into a graph structure, forming a product knowledge graph containing nodes and edges. Changes in product information are continuously monitored, and the data in the product knowledge graph is updated regularly to ensure its timeliness and accuracy. Based on user feedback and market dynamics, the structure and content of the knowledge graph are optimized, adding new product entities and relationships, adjusting existing relationship strengths, and continuously iterating and optimizing to improve the quality of the knowledge graph. The specific tasks of the product knowledge graph construction module are as follows: Product information is collected from multiple channels, including different e-commerce platforms, merchant-owned systems, and public data sources. For e-commerce platforms, product information is obtained from mainstream platforms through API interfaces or web scraping technology. For merchant-owned systems, product data is directly obtained from their backend systems in cooperation with merchants. For public data sources, product-related information is obtained using open data platforms. This data includes product attributes, categories, brands, and user reviews. Duplicate product records are removed using unique identifiers, and data with missing key information or obvious errors is deleted. Data from different sources is converted to a unified format, and fields such as product category and brand are standardized in naming. The cleaned data is then merged into a unified database table. Based on the integrated product information, a knowledge graph building tool (Neo4) is used to define entities and attributes, determine unique identifiers for product entities, define basic attributes of product entities, determine the data type of each attribute, define attribute constraints, and establish relationships between products. Relationship types include same brand. The system identifies relationships, including same-category relationships, complementary relationships, and substitution relationships. Same-brand relationships represent different products under the same brand; same-category relationships represent different products within the same category; complementary relationships include mobile phones and phone cases; and substitution relationships include similar products from different brands. Entity recognition (NER) and relation extraction (RE) technologies are used to extract relationships between products from text data. A graph database is used to store product entities and their relationships. Product information is transformed into a graph structure, forming a product knowledge graph containing nodes and edges. Product entities are used as nodes in the graph, and relationships between products are used as edges. The system continuously monitors changes in product information, regularly updates the data in the product knowledge graph, optimizes its structure and content, sets scheduled tasks to periodically retrieve the latest product information from data sources, uses data comparison tools to detect changes in product information, updates node and edge information in the product knowledge graph, deletes outdated or invalid product entities and relationships, and adjusts the structure of the knowledge graph based on user feedback and market dynamics. The meta-path reasoning module defines meta-paths in the product knowledge graph. It calculates the association probability between users and products using path reasoning algorithms, uncovers latent needs, identifies long-tail products, avoids over-recommendation of popular products, and balances the exposure ratio of popular and long-tail products. Based on the structure and business logic of the product knowledge graph, it determines semantically meaningful meta-path patterns and preprocesses the product knowledge graph, extracting node features and edge information to construct an index structure adapted for path reasoning. Using the path reasoning algorithm, it traverses and searches the product knowledge graph along the defined meta-paths. Combining node features and edge weights, it calculates the association probability between users and each product. By analyzing the probability distribution, it uncovers potential latent needs of users, identifies long-tail products with low popularity but matching potential user interests, and adjusts the exposure ratio of popular and long-tail products in the recommendation results based on the calculated association probabilities. This avoids over-recommendation of popular products and ensures that long-tail products receive appropriate display opportunities. The specific tasks of the meta-path reasoning module are as follows: This process analyzes the knowledge graph structure, identifies node and edge types, recognizes semantic relationships between nodes and edges, and defines semantically meaningful meta-path patterns based on business requirements and the knowledge graph structure. For example, the direct purchase path from a user to a product is User-Purchase-Product; the indirect interest path is User-Purchase-Category-Contains-Product; the brand preference path is User-Purchase-Brand-Belongs to-Product; and the evaluation influence path is User-Evaluation-Product. These meta-path patterns capture the direct and indirect relationships between users and products. The product knowledge graph is preprocessed, extracting features from each node, including attribute values ​​and labels. Node features are vectorized, and edge weights are extracted. Weights are calculated based on edge type. An index structure adapted for path reasoning, i.e., an inverted index, is constructed, and the indexing functionality of the graph database is used for fast lookup and traversal of paths. The algorithm employs path reasoning to traverse the product knowledge graph along defined meta-paths. Combining node features and edge weights, a weighted path counting method is used to multiply the weights of all edges along a path to obtain its weight. These weights are then summed to determine the probability of user-product association. The algorithm analyzes the probability distribution of user-product association, identifying products with high probabilities but low popularity. Statistical analysis is used to visualize the probability distribution, and association and popularity thresholds are set. Products with association probabilities higher than the association threshold and popularity lower than the popularity threshold are identified as long-tail products. Based on the association probabilities, the recommendation results are sorted to ensure that both popular and long-tail products receive appropriate display opportunities. The exposure ratio of popular and long-tail products is adjusted according to business needs, with popular products accounting for 60% of the recommendation results and long-tail products accounting for 40%. The preference prediction and recommendation module is used to combine real-time user behavior with the reasoning results of the knowledge graph to generate a personalized recommendation list; The feedback optimization module is used to collect explicit and implicit user feedback and feed it back into the user profile and knowledge graph to update the knowledge graph.

[0020] Example 2, as Figure 1 , Figure 2 As shown, based on Embodiment 1, the present invention provides a technical solution: preferably, the preference prediction and recommendation module includes a real-time preference prediction unit and a dynamic recommendation generation unit; The real-time preference prediction unit integrates product knowledge graph inference results with real-time user behavior. Through weighted fusion, it predicts users' immediate preferences, improving the accuracy of recommendations for cold-start users and adapting to user preference drift. It acquires data from multiple channels, collecting real-time user behavior data, including browsing, clicking, and dwell time, while simultaneously acquiring product knowledge graph inference results. It integrates the real-time user behavior data with the knowledge graph inference results to form a comprehensive dataset. Based on pre-defined weighting rules, it weights and fuses the product knowledge graph inference results with the real-time user behavior data to generate a feature vector that comprehensively reflects users' immediate preferences, improving the accuracy of recommendations for cold-start users. Using the fused feature vector, combined with a preference prediction model pre-trained based on machine learning, it predicts users' immediate preferences, outputs the degree of user preference for various products, and applies the prediction results to the recommendation system. The specific tasks of the real-time preference prediction unit are as follows: We deploy event tracking code on the front-end interfaces of e-commerce platforms, such as web pages and apps, to collect real-time user behavior data, including browsing, clicks, and dwell time. For example, when a user browses a product details page, we record the product ID, start time, and end time. When a user clicks a product image or the "add to cart" button, we record the click event and related product information. We analyze server logs to obtain more information about user interactions with the platform, including login time, page paths visited, and search keywords. We also obtain inference results from the product knowledge graph inference module, including potential associations between users and products mined from the product knowledge graph. Finally, we clean the collected real-time user behavior data. This process removes duplicate, erroneous, and invalid data, and performs format conversion and standardization on the product knowledge graph inference results to ensure consistency with the user behavior data format. It then links real-time user behavior data with the product knowledge graph inference results, using user ID as the key field to integrate the behavior data and inference results of the same user, forming a comprehensive dataset. Based on business objectives and historical data analysis results, weighting rules are set for the product knowledge graph inference results and real-time user behavior data. For new users (cold start users), due to the lack of historical behavior data, the weight of the product knowledge graph inference results is appropriately increased to 0.7, while the weight of real-time user behavior data is set to 0.3. For existing users, the weighting is adjusted accordingly. Based on the richness and stability of historical behavioral data, weights are dynamically adjusted, setting the weight of real-time user behavior data at 0.6-0.8 and the weight of product knowledge graph inference results at 0.2-0.4. Then, relevant user and product features are extracted from the comprehensive dataset. User features include age, gender, region, and historical purchase categories; product features include product category, brand, price, and sales volume. Simultaneously, according to the set weighting rules, the extracted features are weighted. For user preference features for a specific product category, the number of times the user viewed or clicked on that product category in real-time behavior is multiplied by the corresponding weight, plus the weight of the product knowledge graph inference results. The user's association probability with a certain type of product is multiplied by its weight to obtain the weighted value of the feature. All weighted features are combined into a feature vector to comprehensively reflect the user's immediate preferences. A preference prediction model is constructed using a decision tree-based machine learning algorithm. Historical data is used to train the model. During training, the user's feature vector is used as input and the user's actual preferences are used as output. The model parameters are adjusted to make the model accurately predict the user's preferences, resulting in a trained preference prediction model. The fused feature vector is then input into the trained preference prediction model to output the user's preference level for each type of product. The preference level is represented by a probability value, with a higher value indicating a higher preference for that type of product. The dynamic recommendation generation unit generates a candidate set based on user profiles, real-time user preferences, and the upper limit of the number of product categories to recommend. It then uses reinforcement learning to dynamically adjust the recommendation strategy, balancing popular and long-tail products to generate a personalized recommendation list. Based on user profiles, real-time user preferences, and the preset upper limit of the number of product categories to recommend, it filters out a set of products that match user interests and recommendation rules, forming an initial candidate set. Specifically, it combines basic information and real-time preference data from the user profile to select products matching user interests from a product knowledge graph. The size of the candidate set is controlled according to the upper limit of the number of product categories to ensure the diversity and relevance of the recommended content. The recommendation strategy is dynamically adjusted using reinforcement learning algorithms to balance popular and long-tail products. A reward mechanism is defined to optimize the recommendation strategy in real time based on user feedback, adjusting the recommendation ratio of popular and long-tail products to improve recommendation accuracy and user satisfaction. Based on the adjusted recommendation strategy, the recommendation results are sorted according to user preference, and the final personalized recommendation list is generated from the candidate set, ensuring that the recommended content not only matches the user's current interests but also guides the user to discover new products. The specific tasks of the dynamic recommendation generation unit are as follows: This study analyzes basic information from user profiles and combines it with real-time user preference data to identify the user's current areas of interest. It then performs precise matching within a product knowledge graph, selecting products highly aligned with the user's interests. To ensure diversity and avoid over-concentration on any single category, a pre-set limit is set for the number of recommendations per product category. Among the selected products matching user interests, the recommendation limit is used to filter products, prioritizing those with high user preference and strong alignment with the user profile and real-time preferences, forming an initial candidate set. A reinforcement learning algorithm framework is used to construct the state space, incorporating user profiles, real-time preferences, and the current recommended product list. This comprehensively reflects the user's state and the recommendation environment. An action space is also defined to meet the needs and preferences of different users. This action space includes actions such as adjusting the recommendation ratio of popular and long-tail products, changing recommended product categories, and adjusting the recommendation order. A reward mechanism is then established, defining reward values ​​based on user feedback. For example, if a user clicks on a recommended product, a higher reward is given. The system employs a reward mechanism: a small negative reward is given if a user ignores a recommended product, while a larger positive reward is given if the user purchases it. During the recommendation process, user feedback is collected in real time, including behavioral data such as clicks, browsing, purchases, and favorites. Based on the collected user feedback data and the established reward mechanism, the parameters of the recommendation strategy are updated using a deep Q-network reinforcement learning algorithm. If a user's click-through rate for long-tail products is found to be high, the recommendation ratio of long-tail products is increased; if a user has a strong purchase intention for a certain category of products, the recommendation priority of that category is increased. Through continuous iteration and updates, the recommendation strategy gradually adapts to the changing needs and preferences of different users. Then, combining user profiles, real-time preferences, and the adjusted recommendation strategy, the user preference level is calculated for each product in the initial candidate set. Based on the calculated user preference level, the products in the initial candidate set are sorted. Products with higher preference levels rank higher and are more likely to be recommended to the user. The top-ranked products are selected from the sorted initial candidate set to form the final personalized recommendation list, which is then displayed to the user. The feedback optimization module includes: User feedback, including explicit and implicit feedback, is collected through a front-end interaction and back-end logging system. For explicit feedback, functional entry points such as rating, commenting, liking, and collecting are set up on the platform interface to obtain users' clear attitudes towards products or services. At the same time, data analysis technology is used to capture user behavior data such as browsing, clicking, dwell time, and purchase conversion as implicit feedback. The collected feedback data is integrated into user profiles and knowledge graphs. Explicit feedback directly updates the preference information in user profiles, while implicit feedback adjusts the node attributes and edge weights in the knowledge graph by analyzing behavioral patterns, enhancing the accuracy and dynamism of the knowledge graph. The knowledge graph is updated based on feedback data, optimizing node and edge information, adjusting the relationships between products, adding new entities and relationships, or adjusting the strength of existing relationships. Through continuous iteration, the quality of the knowledge graph is optimized. The specific tasks of the feedback optimization module are as follows: The front-end interfaces of e-commerce platforms, such as web pages and apps, are carefully designed with entry points for functions like ratings, reviews, likes, and favorites. When users perform these actions, the front-end transmits the relevant data to the back-end server in real time. The back-end server verifies and processes the data before storing it in a database. A comprehensive back-end logging system is also built to record user behavior on the platform in detail, including browsing history, click history, and purchase conversion history. By analyzing user browsing paths, click frequency, and dwell time, the system captures users' potential interest and attitudes towards products or services. If a user spends a long time on a product details page and clicks multiple times to view the product details, it indicates that the user has a high interest in that product. If a user browses multiple similar products and ultimately purchases one, the purchase conversion path can be analyzed to understand the user's purchase decision factors. Based on business needs and data analysis results, rules are formulated to explicitly update the preference information in the user profile. For rating data, different rating ranges are set to correspond to different levels of preference, with 4-5 stars indicating high preference. The system uses a preference scale, with 3 stars indicating moderate preference and 1-2 stars indicating low preference. For review data, natural language processing (NLP) is used to extract keywords and sentiment, updating user preferences for product attributes, brands, etc. When new explicit feedback data is collected, the system updates the user profile's preference information in real time according to established rules, extracts user behavior patterns, and adjusts node attributes and edge weights in the knowledge graph based on these patterns. The system also updates the knowledge graph based on feedback data, optimizing node and edge information and adjusting relationships between products. If new products or brands appear in user feedback, they are added as new entities to the knowledge graph, and existing node attributes are adjusted accordingly. If new relationships appear in user feedback, they are added as new relationships to the knowledge graph, and edge weights are adjusted based on user behavior data. Scheduled tasks are set to periodically extract information from user feedback data, update the knowledge graph, and dynamically adjust node and edge information based on real-time user feedback. The quality of the knowledge graph is periodically evaluated, and user satisfaction surveys are used to ensure its accuracy and dynamism.

[0021] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A cross-platform real-time prediction and recommendation system for user consumption preferences, comprising a comprehensive recommendation management platform, characterized in that, The comprehensive recommendation management platform has the following communication modules, among which: The user profile building module is used to collect basic information about new users and build user profiles using entity relationships and attribute information in the knowledge graph. The product knowledge graph construction module is used to collect and organize product information across platforms and build a product knowledge graph. The meta-path reasoning module is used to define meta-paths in the product knowledge graph, calculate the probability of association between users and products through path reasoning algorithms, and uncover implicit needs. The preference prediction and recommendation module is used to combine real-time user behavior with the reasoning results of the knowledge graph to generate a personalized recommendation list; The feedback optimization module is used to collect explicit and implicit user feedback and feed it back into the user profile and knowledge graph to update the knowledge graph.

2. The cross-platform user consumption preference real-time prediction and recommendation system according to claim 1, characterized in that: The user profile building module includes a basic information collection unit and an auxiliary profile unit; The basic information collection unit is used to collect basic information about new users and provide a basic profile framework for them. The auxiliary profiling unit is used to combine new user basic information and platform product categories to infer the user's potential areas of interest and construct a user profile.

3. The cross-platform user consumption preference real-time prediction and recommendation system according to claim 2, characterized in that: The basic information collection unit includes: Based on the system's positioning and business needs, the scope of basic information of new users to be collected is determined, and basic information of new users is collected through the user registration interface. Verify the basic information of new users collected, check the completeness of the information, verify the correctness of the format, and prompt the user to re-enter or supplement the information if it does not meet the requirements. The basic information of verified new users is stored in the user profile database to form the basic user profile framework.

4. The cross-platform user consumption preference real-time prediction and recommendation system according to claim 2, characterized in that: The auxiliary image unit includes: Extract existing product categories from the platform, perform correlation analysis between new user basic information and platform product categories, and preliminarily filter out product categories that match new user basic information based on preset filtering rules in the correlation rule library; Based on the initially screened product categories, potential areas of interest for users are inferred, and these inferred potential areas of interest are integrated with the basic information of new users to form a complete user profile.

5. The cross-platform user consumption preference real-time prediction and recommendation system according to claim 2, characterized in that: The product knowledge graph construction module includes: Product information, including product attributes, categories, brands, and user reviews, is collected from multiple channels covering different e-commerce platforms, merchant-owned systems, and public data sources. The collected data is then cleaned and standardized. At the same time, data from different sources is integrated into a unified data format to form a structured preliminary dataset. Based on the integrated product information, product entities and their attributes are defined, the relationships between products are established, and the product information is transformed into a graph structure through entity recognition and relationship extraction technologies, forming a product knowledge graph containing nodes and edges. We continuously monitor changes in product information, regularly update the data in the product knowledge graph, and optimize the structure and content of the knowledge graph based on user feedback and market dynamics.

6. The cross-platform user consumption preference real-time prediction and recommendation system according to claim 5, characterized in that: The meta-path reasoning module includes: Based on the structure and business logic of the product knowledge graph, we determine the semantically meaningful meta-path patterns, preprocess the product knowledge graph, extract node features and edge information, and construct an index structure adapted to path reasoning. Using a path reasoning algorithm, the system traverses and searches the product knowledge graph along the defined meta-path. Combining node features and edge weights, it calculates the probability of association between users and each product. By analyzing the probability distribution, it uncovers users' potential latent needs. Based on the calculated association probability, the exposure ratio of popular products and long-tail products in the recommendation results is adjusted.

7. A cross-platform user consumption preference real-time prediction and recommendation system according to claim 6, characterized in that: The preference prediction and recommendation module includes a real-time preference prediction unit and a dynamic recommendation generation unit; The real-time preference prediction unit is used to integrate the reasoning results of the product knowledge graph with the user's real-time behavior, and predict the user's immediate preferences through weighted fusion. The dynamic recommendation generation unit is used to generate a candidate set based on user profiles, user real-time preferences, and the upper limit of the number of product categories to be recommended, and to dynamically adjust the recommendation strategy using reinforcement learning to balance popular and long-tail products, thereby generating a personalized recommendation list.

8. A cross-platform user consumption preference real-time prediction and recommendation system according to claim 7, characterized in that: The real-time preference prediction unit includes: Data is acquired from multiple channels, and real-time user behavior data, including browsing, clicking and dwell time, is collected. At the same time, the product knowledge graph inference results are obtained. The real-time user behavior data and the knowledge graph inference results are integrated to form a comprehensive dataset. Based on pre-set weighting rules, the product knowledge graph reasoning results are weighted and fused with real-time user behavior data to generate a feature vector that comprehensively reflects the user's immediate preferences. By using the fused feature vectors and combining them with a preference prediction model pre-trained based on machine learning, the system predicts users' real-time preferences, outputs the degree of user preference for various types of goods, and applies the prediction results to the recommendation system.

9. A cross-platform user consumption preference real-time prediction and recommendation system according to claim 7, characterized in that: The dynamic recommendation generation unit includes: Based on user profiles, real-time user preferences, and a preset upper limit for the number of product categories recommended, a set of products that match user interests and recommendation rules is selected to form an initial candidate set. Among them, products that match user interests are selected from the product knowledge graph by combining basic information and real-time preference data in the user profile, and the size of the candidate set is controlled according to the upper limit for the number of product categories recommended. The recommendation strategy is dynamically adjusted using reinforcement learning algorithms. By defining a reward mechanism, the recommendation strategy is optimized in real time based on user feedback, and the recommendation ratio of popular products and long-tail products is adjusted. Based on the adjusted recommendation strategy, the recommendation results are sorted according to the degree of user preference, and then a final personalized recommendation list is generated from the candidate set.

10. A cross-platform user consumption preference real-time prediction and recommendation system according to claim 9, characterized in that: The feedback optimization module includes: The system collects user feedback information, including explicit and implicit feedback, through front-end interaction and back-end logging. For explicit feedback, the platform interface is set up with entry points for rating, commenting, liking, and collecting to obtain users' clear attitudes toward products or services. At the same time, data analysis technology is used to capture user browsing, clicking, dwell time, and purchase conversion behavior data as implicit feedback. The collected feedback data is integrated into user profiles and knowledge graphs. Explicit feedback directly updates the preference information in user profiles, while implicit feedback adjusts the node attributes and edge weights in the knowledge graph by analyzing behavioral patterns. The knowledge graph is updated based on feedback data, the information of nodes and edges is optimized, the relationships between products are adjusted, new entities and relationships are added, or the strength of existing relationships is adjusted. Through continuous iteration, the quality of the knowledge graph is improved.

Images