Commodity pushing method and system based on user preference analysis

By combining multi-source data analysis and sentiment models, user profiles and product knowledge graphs are constructed, and context-aware similarity is calculated. This solves the problem of existing recommendation systems ignoring implicit feedback, enabling personalized and real-time product recommendations and improving user satisfaction.

CN118941365AActive Publication Date: 2024-11-12WUHAN YUEDONG WUXIAN NETWORK TECHNOLOGY CO LTD

Patent Information

Application Number
CN202411121634.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-15
Publication Date
2024-11-12
Estimated Expiration
2044-08-15

AI Technical Summary

Technical Problem

Existing recommendation systems rely on explicit user feedback, such as ratings and reviews, while neglecting the potential value of implicit feedback. They fail to fully utilize user sentiment and contextual information, resulting in recommendations that lack personalization and relevance, thus affecting user satisfaction.

Method used

By collecting multi-source data to perform sentiment analysis and generate user profiles, constructing a product knowledge graph, using a fuzzy Kano model to evaluate feature importance, combining a BERT model for sentiment analysis, constructing a user-product interaction graph, calculating context-aware similarity between products, using a GraphSAGE model to learn product embedding representations, and using a GBDT model for personalized ranking and recommendation.

Benefits of technology

It significantly improves the personalization and user satisfaction of the recommendation system, and can dynamically adjust the recommended content, taking into account the user's real-time location and environmental changes, to provide highly relevant product recommendations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941365A_ABST
    Figure CN118941365A_ABST
Patent Text Reader

Abstract

The invention discloses a commodity pushing method and system based on user preference analysis, and relates to the technical field of e-commerce, and the method comprises the steps: collecting multi-source data, carrying out the emotion tendency analysis of user comments, generating a user portrait, constructing a commodity knowledge graph, evaluating the feature importance through a fuzzy Carnot model, and generating a feature-emotion pair; and constructing a user-commodity interaction graph based on the user portraits and the feature-emotion pairs, and calculating context perception similarity among commodities according to spatio-temporal information in the user-commodity interaction graph. According to the method, the sentiment analysis technology and the fuzzy Carnot model are used, so that the accuracy of a user portrait is improved, importance evaluation of commodity features is refined through feature-sentiment pairs, a user-commodity interaction graph is constructed, context perception similarity calculation is combined, the real-time position and environment change of the user are considered, and the user experience is improved. And the recommendation content can be dynamically adjusted, and the individuation degree and the user satisfaction degree of the recommendation system are remarkably improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of e-commerce, and in particular to a commodity push method and system based on user preference analysis. Background Art

[0002] In the context of the rapid development of digitalization and e-commerce, personalized recommendation systems have become a key technology to attract users and improve user experience. The core goal of recommendation systems is to accurately predict users' interests and preferences in order to recommend the most suitable goods or services to users. Early recommendation systems mainly relied on simple collaborative filtering technology, which predicts new goods or services that users may be interested in by analyzing users' historical behavior data. With the advancement of technology, recommendation systems have gradually incorporated more complex algorithms, such as matrix decomposition, deep learning, and natural language processing, making recommendation results more accurate and personalized.

[0003] However, existing recommendation technologies still have some shortcomings. Many systems still rely on explicit user feedback, such as ratings and comments, but ignore the potential value of implicit feedback and fail to fully utilize user sentiment tendencies and contextual information, resulting in a lack of personalization and relevance in recommendation results in some cases, affecting the effectiveness of the recommendation system and the ultimate satisfaction of users. Summary of the invention

[0004] In view of the problems existing in the above-mentioned existing commodity push method and system based on user preference analysis, the present invention is proposed.

[0005] Therefore, the problem to be solved by the present invention is that many systems still rely on explicit user feedback, such as ratings and comments, but ignore the potential value of implicit feedback and fail to fully utilize user emotional tendencies and contextual information, resulting in a lack of personalization and relevance in recommendation results in some cases, affecting the effectiveness of the recommendation system and the ultimate satisfaction of users.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a product push method based on user preference analysis, which includes collecting multi-source data, performing sentiment tendency analysis on user comments to generate user portraits and constructing a product knowledge graph, using a fuzzy Kano model to evaluate feature importance and generate feature-emotion pairs; constructing a user-product interaction graph based on user portraits and feature-emotion pairs, calculating the context-aware similarity between products based on the spatiotemporal information in the user-product interaction graph; and generating a recommendation list based on the context-aware similarity between products.

[0007] As a preferred solution of the product push method based on user preference analysis described in the present invention, wherein: the collection of multi-source data, the sentiment analysis of user comments to generate user portraits and the construction of product knowledge graphs refers to collecting user behavior data, product data and context data including geographic location, timestamp and device information from websites, mobile applications and social media, cleaning the collected data and using NLP tools to segment user comments, remove stop words and perform part-of-speech tagging;

[0008] Select key features of user behavior, including visit frequency, purchase frequency, and social interaction frequency. Use one-hot encoding and TE-IDF methods to convert behavior features into vector representations and identify characteristic words in comments. Integrate behavior data with geographic location, timestamp, and device information, and attach context labels to each user behavior record.

[0009] Use the BERT model for sentiment analysis. Input the feature words into the BERT model to obtain the feature vector of the text. Input the feature vector output by the BERT model into the classification layer composed of a fully connected layer and a Softmax activation function to output the probability of each sentiment category.

[0010] In another fully connected layer, a linear function is used to calculate the sentiment intensity:

[0011] I=σ(V i T x+b i ),

[0012] Where I is the emotion intensity, σ is the Sigmoid function, V i and b i are the weights and biases of the sentiment intensity layer, and T is the transposition operation;

[0013] Calculate the sentiment score for each review based on the sentiment category probability value and strength:

[0014] S l =P(r l )×I(r l ),

[0015] Where S l is the sentiment score of the lth user comment, P(r l ) is the probability value of the sentiment category of the lth user comment, I(r l ) is the sentiment intensity of the lth user comment;

[0016] Calculate the overall sentiment score for each user, combining the sentiment information of all comments:

[0017]

[0018] In the formula, S total is the user's overall sentiment score, and N is the number of comments by the user;

[0019] Assign sentiment labels to users based on their total sentiment scores:

[0020] If S total ≥0.75, the sentiment label is “positive”;

[0021] If 0.75>S total >0.25, the sentiment label is “neutral”;

[0022] If S total ≤0.25, the sentiment label is “negative”;

[0023] Combine the behavioral feature vector with the sentiment score to form a user portrait and use the K-means clustering algorithm to divide users into different groups;

[0024] Extract features from the collected product data, standardize them, and convert the attribute information into feature vector A. Use one-hot encoding to represent categories and brands, and standardize prices and ratings.

[0025] The relationship between products is identified through co-occurrence analysis and collaborative filtering technology, and the relationship strength PMI (x, y) between products is calculated using point mutual information:

[0026]

[0027] Where P(x,y) is the probability of goods x and y appearing at the same time, P(x) and P(y) are the probabilities of goods x and y appearing separately, respectively;

[0028] Create a node for each product. The node attributes include the product feature vector A and the user sentiment label. Define the edges between products based on the PMI value and the user's behavior pattern.

[0029] When new products are added or product information changes, updates are performed, the PMI values ​​of related products are recalculated, and the corresponding edge weights and node information are updated in the product knowledge graph.

[0030] As a preferred solution of the product push method based on user preference analysis described in the present invention, wherein: identifying product-related features in user comments through product knowledge graph and using fuzzy Kano model to evaluate feature importance and generate feature-sentiment pairs refers to designing a questionnaire containing the extracted product features, using Likert scale to ask users about their satisfaction and expectation level for each product feature, using online survey tools to collect user responses, encoding and converting user responses, and converting Likert scale scores into numerical inputs required for fuzzy sets;

[0031] For the satisfaction and expectation levels, three fuzzy sets are defined, including a high fuzzy set, a medium fuzzy set, and a low fuzzy set, and each fuzzy set is represented using a triangular membership function:

[0032]

[0033] In the formula, δ(q) is the membership function, a', b', c' are the cut-off points for high, medium and low satisfaction, and q is the score value;

[0034] Using fuzzy Kano logic, the features are classified according to the fuzzy set membership of user satisfaction and expectation level, and the importance index of each product feature is calculated using fuzzy AND:

[0035] G t =min(μ sat (f),μ exp (f)),

[0036] In the formula, G t is the importance index of product feature t, μ sat (f) and μ exp (f) are the fuzzy membership of satisfaction and expectation respectively;

[0037] Generate a comprehensive importance report for each product feature based on feature importance and sentiment score;

[0038] The sentiment score and importance score of the feature are combined to generate a feature-sentiment pair list for each product. Each feature-sentiment pair includes the feature name, sentiment score, and importance score.

[0039] As a preferred solution of the product push method based on user preference analysis of the present invention, wherein: the construction of the user-product interaction graph based on the user portrait and feature-emotion pair refers to combining the user's behavior data with the feature-emotion pair, and marking the relevant product features and the user's emotional score and importance of the product features for each layer of interaction between the user and the product;

[0040] A node is created for each independent user, and the node attributes include the user's identification information. A node is created for each independent product, and the node attributes include the product ID, category, and brand. Each set of user behaviors is mapped to the edge of the user-product interaction graph and the context information is used as the edge attribute. The user-product interaction graph is stored in a graph database.

[0041] As a preferred solution of the product push method based on user preference analysis of the present invention, the calculation of the context-aware similarity between products based on the spatiotemporal information in the user-product interaction graph refers to learning the embedded representation of the product from the interaction graph using the GraphSAGE model, and the embedding generation formula is:

[0042]

[0043] In the formula, is the embedding vector of node v in layer k, is the embedding vector of node u at the k-1th layer, Q(v) is the set of neighbor nodes of node v, A is the aggregation function, and W (k) is the weight matrix, B (k) is the bias term of the kth layer, and μ is the ReLU activation function;

[0044] Combining the embedding vectors and contextual attributes of the items, we can calculate the context-aware similarity between the items:

[0045]

[0046] In the formula, Sim og is the similarity between product o and product g, h o and h g is the embedding vector of the product, γ is the weight factor of the context, and O(o,g) is the context similarity;

[0047] Based on the calculated similarity, similar context-related products are recommended to users.

[0048] As a preferred solution of the product push method based on user preference analysis of the present invention, wherein: the generating of the recommendation list according to the context-aware similarity between products means that for each user, M products are selected from the products with the highest similarity to the user's historical interaction products as the candidate recommendation set, and the products that do not meet the current context are filtered according to the user's current context, and the candidate set is optimized;

[0049] Construct a feature vector for each candidate product, use the GBDT model to predict the score of each user's recommendation candidate set, sort the candidate set in descending order according to the predicted score, select the top M products based on the sorting results to form the final recommendation list, and push the product content that the user is interested in to the user according to the recommendation list.

[0050] As a preferred solution of the product push method based on user preference analysis described in the present invention, wherein: pushing the product content that the user is interested in to the user according to the recommendation list refers to defining the push frequency and time to perform instant push according to the user's online status and active time, real-time monitoring and recording of user reactions to the pushed products, integrating the collected user feedback into the user portrait, and updating user preferences and behavior patterns in real time.

[0051] Another object of the present invention is to provide a product push system based on user preference analysis, which comprises:

[0052] A data collection module, which is used to collect user behavior data, product data, and context data from multiple channels and pre-process the data;

[0053] A construction module is used to extract key features from preprocessed data to build user portraits and to build a product knowledge graph based on the analysis of product data;

[0054] Evaluation module, which is used to calculate the importance index of each product feature and generate feature-sentiment pairs by combining the sentiment score and importance score of the product feature;

[0055] The analysis module is used to construct a user-item interaction graph based on user profiles and feature-sentiment pairs, learn the embedded representation of items from the interaction graph using the GraphSAGE model, and calculate the context-aware similarity between items;

[0056] The recommendation module is used to generate a preliminary recommendation list based on similarity and use the GBDT model to personalize the sorting of the list to finally generate a personalized recommendation list, push products based on the list and collect user feedback.

[0057] A computer device comprises: a memory and a processor; the memory stores a computer program, and the processor implements the steps of a commodity push method based on user preference analysis when executing the computer program.

[0058] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a commodity push method based on user preference analysis.

[0059] The beneficial effects of the present invention are as follows: the present invention not only improves the accuracy of user portraits by using sentiment analysis technology and fuzzy Kano model, but also refines the importance evaluation of product features through feature-sentiment pairs, constructs a user-product interaction graph and combines it with context-aware similarity calculation, which not only takes into account the user's real-time location and environmental changes, but also dynamically adjusts the recommended content, significantly improving the personalization of the recommendation system and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0061] Figure 1 The flowchart of the product push method based on user preference analysis is shown in FIG.

[0062] Figure 2 Schematic diagram of the implementation of feature-sentiment pair generation.

[0063] Figure 3 This is a structural diagram of a product push system based on user preference analysis. DETAILED DESCRIPTION

[0064] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings.

[0065] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0066] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or selective embodiment that is mutually exclusive with other embodiments.

[0067] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, and which provides a commodity push method based on user preference analysis. The commodity push method based on user preference analysis includes:

[0068] S1. Collect multi-source data, analyze user comments for sentiment tendency to generate user portraits and build product knowledge graphs, use fuzzy Kano model to evaluate feature importance and generate feature-sentiment pairs;

[0069] Specifically, collecting multi-source data, analyzing user comments for sentiment tendency to generate user portraits and constructing product knowledge graphs means collecting user behavior data, product data and contextual data including geographic location, timestamp and device information from websites, mobile applications and social media, cleaning the collected data and using NLP tools to segment user comments, remove stop words and perform part-of-speech tagging;

[0070] Select key features of user behavior, including visit frequency, purchase frequency, and social interaction frequency, use one-hot encoding and TE-IDF methods to convert behavioral features into vector representations and identify characteristic words in reviews, including battery life and screen resolution;

[0071] Integrate behavioral data with geolocation, timestamp, and device information to attach contextual tags to each user behavior record;

[0072] Use the BERT model for sentiment analysis, input the feature words into the BERT model to obtain the feature vector of the text, input the feature vector output by the BERT model into the classification layer composed of the fully connected layer and the Softmax activation function, and output the probability of each sentiment category:

[0073]

[0074] Where P(y=c|x) is the probability of predicting category c given the feature vector x, and W c and b c is the weight and bias of category c, C is the total number of sentiment categories, W j is the weight matrix associated with category j, b j is the bias associated with class j, T is the transpose operation, and the weights and biases are calculated through initialization, definition of the loss function, backpropagation, and iterative application of the optimization algorithm;

[0075] In another fully connected layer, a linear function is used to calculate the sentiment intensity:

[0076] I=σ(V i T x+b i ),

[0077] Where I is the emotion intensity, ranging from [0,1], σ is the Sigmoid function, and V i and b i are the weights and biases of the sentiment intensity layer, and T is the transposition operation;

[0078] Calculate the sentiment score for each review based on the sentiment category probability value and strength:

[0079] S l =P(r l)×I(r l ),

[0080] Where S l is the sentiment score of the lth user comment, P(r l ) is the probability value of the sentiment category of the lth user comment, I(r l ) is the sentiment intensity of the lth user comment (a real number between 0 and 1);

[0081] Calculate the overall sentiment score for each user, combining the sentiment information of all comments:

[0082]

[0083] In the formula, S total is the user's overall sentiment score, ranging from [0,1], and N is the number of comments by the user;

[0084] Assign sentiment labels to users based on their total sentiment scores:

[0085] If S total ≥0.75, the sentiment label is “positive”;

[0086] If 0.75>S total >0.25, the sentiment label is “neutral”;

[0087] If S total ≤0.25, the sentiment label is “negative”;

[0088] The selection of 0.75 and 0.25 as thresholds is based on the idea of ​​quartiles, where the data is divided into four equal parts, each containing 25% of the data points. Positive sentiment is usually considered to be a comment with a higher score, so selecting data points above the 75th percentile as positive sentiment can ensure that only truly positive comments are classified as positive. The threshold for negative sentiment is set below the 25th percentile to ensure that comments classified as negative do contain more negative emotions. This data-based quartile method can adapt to different data sets and sentiment distributions, making sentiment classification closer to the actual data distribution. Such a division can better balance the number of categories and avoid too many or too few categories, thereby improving the generalization ability and accuracy of the model;

[0089] Combine the behavioral feature vector with the sentiment score to form a user portrait and use the K-means clustering algorithm to divide users into different groups;

[0090] Extract features including category, brand, price and rating from the collected product data, convert the attribute information into feature vector A after standardization, use one-hot encoding to represent category and brand, and standardize price and rating;

[0091] The relationship between products is identified through co-occurrence analysis and collaborative filtering technology, and the relationship strength PMI (x, y) between products is calculated using point mutual information:

[0092]

[0093] Where P(x,y) is the probability of goods x and y appearing at the same time, P(x) and P(y) are the probabilities of goods x and y appearing separately, respectively;

[0094] Create a node for each product. The node attributes include the product feature vector A and the user sentiment label. Define the edges between products based on the PMI value and the user's behavior pattern. The edge attributes reflect the type and strength of the relationship between products.

[0095] When new products are added or product information changes, updates are performed, the PMI values ​​of related products are recalculated, and the corresponding edge weights and node information are updated in the product knowledge graph.

[0096] By collecting user behavior data and contextual information from multiple channels such as websites, mobile applications, and social media, the platform can obtain rich and diverse data sets that reflect users' real interests and behavior patterns. Using NLP tools to analyze the sentiment tendency of user comments, the BERT model can accurately capture the complex emotions in user comments. When users browse products on the platform and leave comments, the system can analyze their sentiment tendencies in real time. If a user expresses strong positive emotions towards a certain type of product (such as electronic products) in the evaluation, the system will automatically adjust the user portrait, record the preference for this type of product, and provide support for subsequent product recommendations. By combining the user's behavioral characteristics (such as visit frequency, purchase frequency) with the results of sentiment analysis, the system builds an accurate user portrait. At the same time, the product knowledge graph updates product relationships and user preference changes in real time to ensure the timeliness of the recommendation system. The accuracy of the user portrait directly affects the construction of the product knowledge graph, and the real-time updated knowledge graph in turn optimizes the user portrait, so that a positive feedback is formed between the two. When users continue to browse different categories of products (such as electronic devices, household items), the system will update the user portrait in real time based on the browsing frequency and the emotional tendency in the comments, identify the user's special preference for electronic devices, and strengthen the relationship weight between the electronic device node and the user node in the knowledge graph, enhance the accuracy of subsequent recommendations, and use the K-means clustering algorithm to divide users into groups, and divide users into different interest groups based on emotional scores and behavioral characteristics. In this way, the recommendation system can identify and push personalized content that is more in line with user preferences, significantly improving the relevance and accuracy of recommendations.

[0097] User group segmentation provides a more detailed user group image, which enables the recommendation system to make differentiated recommendations between different groups, thereby improving the success rate of recommendations. The platform can identify user groups that like technology products and recommend the latest technology products and related accessories to them, or push home furnishing promotion information to users who like home products. Group segmentation allows each user to receive product recommendations that match their interests, thereby increasing the user's purchase intention and the platform's sales. The dynamic update mechanism of the product knowledge graph enables the system to flexibly respond to market changes and user demand fluctuations. Real-time data updates ensure that the system's recommendation results are always based on the latest information, providing strong market decision-making support for enterprises. Dynamic adaptability ensures the real-time nature of the recommendation system, and real-time data provides an accurate basis for market decisions. When new products are launched or the sales of a certain product surge, the system will automatically update the weights of relevant nodes and edges in the product knowledge graph, adjust the recommendation strategy, and give priority to recommending new products and popular products. Enterprises can use this information to adjust inventory, formulate promotion strategies, and improve market response speed and profitability.

[0098] Furthermore, the product knowledge graph is used to identify product-related features in user reviews and the fuzzy Kano model is used to evaluate feature importance and generate feature-sentiment pairs. A questionnaire containing the extracted product features is designed, and the Likert scale (1-5 points) is used to ask users about their satisfaction and expectation level for each product feature. An online survey tool is used to collect user responses, and the user responses are encoded and converted to convert the Likert scale scores into the numerical inputs required by the fuzzy set.

[0099] For the satisfaction and expectation levels, three fuzzy sets are defined, including a high fuzzy set, a medium fuzzy set, and a low fuzzy set, and each fuzzy set is represented using a triangular membership function:

[0100]

[0101] In the formula, δ(q) is the membership function, with a value range of [0,1], a', b', c' are the cutoff points for high, medium, and low satisfaction, and q is the score value;

[0102] Using fuzzy Kano logic, the features are classified according to the fuzzy set membership of user satisfaction and expectation level, and the importance index of each product feature is calculated using fuzzy AND:

[0103] G t =min(μ sat (f),μ exp (f));

[0104] In the formula, G t is the importance index of product feature t, μ sat (f) and μexp (f) are the fuzzy membership of satisfaction and expectation respectively;

[0105] Generate a comprehensive importance report for each product feature based on feature importance and sentiment score;

[0106] The sentiment score and importance score of the feature are combined to generate a feature-sentiment pair list for each product. Each feature-sentiment pair includes the feature name, sentiment score, and importance score.

[0107] By designing a questionnaire that includes extracting product features and using a Likert scale to collect user satisfaction and expectation levels, the fuzzy Kano model can carefully evaluate the importance of each product feature. This process combines the user's subjective perception and objective data analysis to generate a more comprehensive feature importance index. Fuzzy sets and triangular membership functions are used to represent user satisfaction and expectations. Fuzzy logic is introduced into feature evaluation, which significantly improves the recognition of feature importance. The membership function calculates the fuzzy membership under different satisfaction levels through a piecewise linear function, accurately capturing the user's fuzzy cognition of the feature. This analysis can not only identify the features that users consider to be critical, but also adjust the direction of product development and improvement based on user feedback. For example, if "battery life" shows a high importance index in both the fuzzy membership of satisfaction and expectation, then this feature should become the focus of product improvement. Combining the feature's sentiment score and importance score, feature-sentiment pairs are generated, further deepening the understanding of user preferences. Each feature-sentiment pair includes the feature name, sentiment score, and importance score, providing a detailed overview of user sentiment and feature preferences. The calculated feature importance index G t It integrates user satisfaction and expectations to ensure that the importance of each product feature can be quantified and compared. This result is combined with the sentiment score to generate a feature-sentiment pair list that provides accurate sentiment and performance analysis for each product. Through the analysis of feature-sentiment pairs, the system can generate a comprehensive importance report for each product feature. This report not only provides strong support for the company's market strategy, but also provides a clear direction for product optimization. For example, by analyzing the feature-sentiment pairs of a certain smartphone, it can be found that users have positive emotions and high importance assessments for "screen resolution", thus suggesting that this feature should be emphasized in marketing promotion.

[0108] S2, construct a user-product interaction graph based on user portraits and feature-emotion pairs, and calculate the context-aware similarity between products based on the spatiotemporal information in the user-product interaction graph;

[0109] Specifically, constructing a user-product interaction graph based on user portraits and feature-emotion pairs means combining user behavior data with feature-emotion pairs, annotating relevant product features and user emotions for each piece of user behavior data, and annotating relevant product features and user emotion scores and importance for each layer of user-product interaction;

[0110] A node is created for each independent user, and the node attributes include the user's identification information. A node is created for each independent product, and the node attributes include the product ID, category, and brand. Each set of user behaviors is mapped to the edge of the user-product interaction graph and the context information is used as the edge attribute. The user-product interaction graph is stored in a graph database.

[0111] A node is created for each independent user. The node attributes include the user's identification information, behavioral characteristics, interest tags, etc. These attributes can fully describe the user's profile and behavior pattern. A node is created for each independent product. The node attributes include the product's ID, category, brand, and the sentiment score and feature importance obtained through the feature-sentiment pair. This information helps the system identify the value of the product from the user's perspective. Through the creation of user and product nodes, the system can accurately identify and describe each user and product. This detailed node information lays a solid foundation for subsequent data analysis and recommendation. The feature-sentiment pair information contained in the product node enables the recommendation system to consider the user's emotional tendency and feature preference, thereby improving the personalization and accuracy of the recommendation. By converting user behavior into edges in the graph, the system can fully capture the complex interactive relationship between users and products. This multi-dimensional interaction modeling helps to reveal the user's deep needs and potential interests. By converting user behavior into edges in the graph, the system can fully capture the complex interactive relationship between users and products. This multi-dimensional interaction modeling helps to reveal the user's deep needs and potential interests. A complete user-product interaction graph is constructed using user and product nodes and the edge relationships between them. The graph reflects all interactive relationships between users and products, including direct interactions and indirect associations.

[0112] Furthermore, calculating the context-aware similarity between products based on the spatiotemporal information in the user-product interaction graph refers to using the GraphSAGE model to learn the embedded representation of products from the interaction graph. The embedding generation formula is:

[0113]

[0114] In the formula, is the embedding vector of node v in layer k, is the embedding vector of node u at the k-1th layer, Q(v) is the set of neighbor nodes of node v, A is an aggregation function, common ones include mean, maximum, sum, etc. In this embodiment, mean aggregation is preferred, W (k) is the weight matrix, B (k) is the bias term of the kth layer, and μ is the ReLU activation function;

[0115] Combining the embedding vectors and contextual attributes of the items, we can calculate the context-aware similarity between the items:

[0116]

[0117] In the formula, Sim og is the similarity between product o and product g, h o and h g is the embedding vector of the product, γ is the context weight factor, by determining the context factors that affect the user's purchase decision (such as time, location, device type, etc.), use statistical analysis methods (such as variance analysis, chi-square test) to evaluate the degree of influence of different context factors on user behavior, determine the influence of each context factor, and provide a basis for setting the weight factor γ. Based on the analysis results, set a preliminary γ value. For example, if it is found that the time factor has a significant impact on user purchase behavior, a higher γ value can be set accordingly. Design A / B tests or multivariate tests, randomly assign users to different experimental groups, and use different γ values ​​for each group. Compare the performance of the recommendation system under different γ values, focusing on indicators including click-through rate, conversion rate, user satisfaction, etc. Use the verification error of the machine learning model, the precision and recall rate of information retrieval, and other indicators to evaluate the effects of different γ values. Adjust the value of γ according to the experimental results to find the optimal solution, and select the γ value that performs best in the experiment as the final value. C(o,g) is the context similarity;

[0118] Based on the calculated similarity, similar context-related products are recommended to users.

[0119] The GraphSAGE model generates an embedding vector that can capture the complex relationships and features between products by sampling and aggregating neighbor node information layer by layer. This high-dimensional embedding representation provides accurate data support for similarity calculation and improves the representation ability of the model. GraphSAGE uses sampling technology to avoid the complexity of full-graph calculation in traditional graph neural networks and improves the computational efficiency of large-scale data sets. This feature enables the system to process a large amount of user and product data and maintain efficient real-time response. The embedding vector is not only used for the calculation of product similarity, but also provides a basis for the personalized recommendation system, which directly affects the final recommendation accuracy. Context-aware similarity combines product embedding and context attributes, and can be dynamically adjusted according to the user's real-time needs and preferences to ensure that the recommended products are highly relevant. This personalized recommendation improves user satisfaction and engagement. By combining contextual information (such as user's time preference and geographic location), the recommendation system can identify and match the user's specific needs, provide more targeted product recommendations, and increase purchase intention. The calculation of context-aware similarity depends on real-time context information, which forms a closed loop with the dynamic update of the interaction graph and the embedding representation learning of the GraphSAGE model, enabling the system to respond quickly and adjust the recommendation strategy.

[0120] S3, generating a recommendation list based on the context-aware similarity between products;

[0121] Specifically, generating a recommendation list based on the context-aware similarity between products means that for each user, N products are selected from the products with the highest similarity to the user's historical interaction products as a candidate recommendation set, and products that do not match the current context are filtered out according to the user's current context to optimize the candidate set;

[0122] Construct a feature vector for each candidate product, including product similarity, user historical behavior, contextual information, product attributes, and sentiment analysis scores. Use the GBDT model for ranking prediction. Use historical user interaction data to train the GBDT model. Apply cross-validation and grid search techniques to optimize model parameters. Define the optimization goal as maximizing click-through rate and user satisfaction.

[0123] Use the trained GBDT model to predict the score of each user's recommendation candidate set, sort the candidate set in descending order according to the predicted score, and select the top N products based on the sorting results to form the final recommendation list;

[0124] Push product content that the user is interested in to the user according to the recommendation list.

[0125] By analyzing the user's real-time context information (such as geographic location, browsing time, device type, etc.), the user is provided with product recommendations that are highly relevant to the current situation. This context-aware capability makes the recommendation not only based on the user's historical behavior, but also responds to its immediate environmental and emotional changes, greatly enhancing the relevance and attractiveness of the recommendation. By providing time-space related product recommendations, the system can significantly improve the user's shopping experience, reduce information overload, and increase the user's trust and satisfaction with the recommendation system. For example, when a user searches for gifts on the eve of a holiday, the system can recommend products that match the festive atmosphere and enhance the willingness to buy. By using cross-validation and grid search techniques, the present invention ensures the consistency and superiority of the GBDT model on different data subsets and systematically finds the optimal model parameters. This method improves the generalization ability of the model and reduces the risk of overfitting. The optimization of the GBDT model makes it faster to process user requests, improves the system's responsiveness to real-time data, and makes the user feel almost no waiting, thereby improving the user's overall satisfaction. The candidate set is dynamically adjusted according to the user's current situation to ensure that the recommended products are not only consistent with the user's personal preferences, but also adapt to their current needs. This flexible candidate set adjustment strategy improves the acceptance rate and click-through rate of recommendations. It uses the GBDT model to personalize the sorting of candidate products, taking into account the multi-dimensional characteristics of products, such as similarity, user historical behavior scores, and contextual information, making the recommendation results more accurate and targeted. By analyzing the recommendation effect and user feedback, the present invention can provide companies with insights into user preferences and behavior trends, assisting companies in making more informed decisions in marketing, inventory management, and new product development. The system can capture and analyze market dynamics in real time, respond to changes in user behavior, and help companies quickly adjust market strategies and seize market opportunities.

[0126] Furthermore, pushing product content that the user is interested in to the user according to the recommendation list means defining the push frequency and time to perform instant push according to the user's online status and active time, monitoring and recording the user's response to the pushed products in real time, including click-through rate, purchase conversion rate and page dwell time, and providing a user feedback channel, integrating the collected user feedback into the user portrait, and updating the user preferences and behavior patterns in real time.

[0127] Through accurate online status detection, push content can be sent when users are most active, increasing the exposure and click-through rate of push content. Dynamic adjustment of push frequency makes the push rhythm more in line with user habits, avoids excessive disturbance of users, and improves user experience. The system records the user's behavior after receiving push content in real time, including click-through rate, purchase conversion rate, and page dwell time. Using data analysis technology, the effect of push content is evaluated, and a user feedback channel is provided. Users can submit opinions and suggestions on push content through this channel to help the system understand the real needs and satisfaction of users. By continuously integrating user feedback, the system can keep the user portrait updated in real time, making the user portrait more in line with real user behavior.

[0128] Example 2, reference Figure 3 , which is the second embodiment of the present invention. This embodiment is different from the previous embodiment and provides a product push system based on user preference analysis, which includes:

[0129] A data collection module, which is used to collect user behavior data, product data, and context data from multiple channels and pre-process the data;

[0130] A construction module is used to extract key features from preprocessed data to build user portraits and to build a product knowledge graph based on the analysis of product data;

[0131] Evaluation module, which is used to calculate the importance index of each product feature and generate feature-sentiment pairs by combining the sentiment score and importance score of the product feature;

[0132] The analysis module is used to construct a user-item interaction graph based on user profiles and feature-sentiment pairs, learn the embedded representation of items from the interaction graph using the GraphSAGE model, and calculate the context-aware similarity between items;

[0133] The recommendation module is used to generate a preliminary recommendation list based on similarity and use the GBDT model to personalize the sorting of the list to finally generate a personalized recommendation list, push products based on the list and collect user feedback.

[0134] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program codes.

[0135] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0136] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0137] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used to implement: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

Claims

1. A method for pushing products based on user preference analysis, characterized in that: include, Collect multi-source data, perform sentiment analysis on user comments to generate user portraits and build product knowledge graphs, use the fuzzy Kano model to evaluate feature importance and generate feature-sentiment pairs; Construct a user-product interaction graph based on user portraits and feature-sentiment pairs, and calculate the context-aware similarity between products based on the spatiotemporal information in the user-product interaction graph; Generate a recommendation list based on context-aware similarities between items.

2. The method for pushing products based on user preference analysis according to claim 1, characterized in that: The collecting of multi-source data, performing sentiment analysis on user comments to generate user portraits and constructing a product knowledge graph refers to collecting user behavior data, product data and context data including geographic location, timestamp and device information from websites, mobile applications and social media, cleaning the collected data and using NLP tools to segment user comments, remove stop words and perform part-of-speech tagging; Select key features of user behavior, including visit frequency, purchase frequency, and social interaction frequency. Use one-hot encoding and TE-IDF methods to convert behavior features into vector representations and identify characteristic words in comments. Integrate behavior data with geographic location, timestamp, and device information, and attach context labels to each user behavior record. Use the BERT model for sentiment analysis. Input the feature words into the BERT model to obtain the feature vector of the text. Input the feature vector output by the BERT model into the classification layer composed of a fully connected layer and a Softmax activation function to output the probability of each sentiment category. In another fully connected layer, a linear function is used to calculate the sentiment intensity: Where I is the emotion intensity, σ is the Sigmoid function, V i and b i are the weights and biases of the sentiment intensity layer, and T is the transposition operation; Calculate the sentiment score for each review based on the sentiment category probability value and strength: S l =P(r l )×I(r l ), Where S l is the sentiment score of the lth user comment, P(r l ) is the probability value of the sentiment category of the lth user comment, I(r l ) is the sentiment intensity of the lth user comment; Calculate the overall sentiment score for each user, combining the sentiment information of all comments: In the formula, S total is the user's overall sentiment score, and N is the number of comments by the user; Assign sentiment labels to users based on their total sentiment scores: If S total ≥0.75, the sentiment label is "positive"; If 0.75>S total >0.25, the sentiment label is "neutral"; If S total ≤0.25, the sentiment label is "negative"; Combine the behavioral feature vector with the sentiment score to form a user portrait and use the K-means clustering algorithm to divide users into different groups; Extract features from the collected product data, standardize them, and convert the attribute information into feature vector A. Use one-hot encoding to represent categories and brands, and standardize prices and ratings. The relationship between products is identified through co-occurrence analysis and collaborative filtering technology, and the relationship strength PMI (x, y) between products is calculated using point mutual information: Where P(x,y) is the probability of goods x and y appearing at the same time, P(x) and P(y) are the probabilities of goods x and y appearing separately, respectively; Create a node for each product. The node attributes include the product feature vector A and the user sentiment label. Define the edges between products based on the PMI value and the user's behavior pattern. When new products are added or product information changes, updates are performed, the PMI values ​​of related products are recalculated, and the corresponding edge weights and node information are updated in the product knowledge graph.

3. The method for pushing products based on user preference analysis according to claim 2, characterized in that: The method includes identifying product-related features in user reviews through a product knowledge graph and using a fuzzy Kano model to evaluate feature importance and generate feature-sentiment pairs, designing a questionnaire containing the extracted product features, using a Likert scale to ask users about their satisfaction and expectation level for each product feature, using an online survey tool to collect user responses, encoding and converting user responses, and converting Likert scale scores into numerical inputs required for fuzzy sets; For the satisfaction and expectation levels, three fuzzy sets are defined, including a high fuzzy set, a medium fuzzy set, and a low fuzzy set, and each fuzzy set is represented using a triangular membership function: In the formula, δ(q) is the membership function, a', b', c' are the cut-off points for high, medium and low satisfaction, and q is the score value; Using fuzzy Kano logic, the features are classified according to the fuzzy set membership of user satisfaction and expectation level, and the importance index of each product feature is calculated using fuzzy AND: G t =min(μ sat (f),m exp (f)), In the formula, G t is the importance index of product feature t, μ sat (f) and μ exp (f) are the fuzzy membership of satisfaction and expectation respectively; Generate a comprehensive importance report for each product feature based on feature importance and sentiment score; The sentiment score and importance score of the feature are combined to generate a feature-sentiment pair list for each product. Each feature-sentiment pair includes the feature name, sentiment score, and importance score.

4. The method for pushing products based on user preference analysis according to claim 3, characterized in that: The construction of the user-product interaction graph based on the user portrait and feature-emotion pair refers to combining the user's behavior data with the feature-emotion pair, and marking the relevant product features and the user's emotional score and importance of the product features for each layer of user-product interaction; A node is created for each independent user, and the node attributes include the user's identification information. A node is created for each independent product, and the node attributes include the product ID, category, and brand. Each set of user behaviors is mapped to the edge of the user-product interaction graph and the context information is used as the edge attribute. The user-product interaction graph is stored in a graph database.

5. The method for pushing products based on user preference analysis according to claim 4, characterized in that: The calculation of the context-aware similarity between products based on the spatiotemporal information in the user-product interaction graph refers to using the GraphSAGE model to learn the embedded representation of the products from the interaction graph. The embedding generation formula is: In the formula, is the embedding vector of node v in layer k, is the embedding vector of node u at the k-1th layer, Q(v) is the set of neighbor nodes of node v, A is the aggregation function, and W (k) is the weight matrix, B (k) is the bias term of the kth layer, and μ is the ReLU activation function; Combining the embedding vectors and contextual attributes of the items, we can calculate the context-aware similarity between the items: In the formula, Sim og is the similarity between product o and product g, h o and h g is the embedding vector of the product, γ is the weight factor of the context, and O(o,g) is the context similarity; Based on the calculated similarity, similar context-related products are recommended to users.

6. The method for pushing products based on user preference analysis according to claim 5, characterized in that: Generating a recommendation list based on context-aware similarity between products means, for each user, selecting M products with the highest similarity to the user's historical interaction products as a candidate recommendation set, filtering products that do not match the current context based on the user's current context, and optimizing the candidate set; Construct a feature vector for each candidate product, use the GBDT model to predict the score of each user's recommendation candidate set, sort the candidate set in descending order according to the predicted score, select the top M products based on the sorting results to form the final recommendation list, and push the product content that the user is interested in to the user according to the recommendation list.

7. The method for pushing products based on user preference analysis according to claim 6, characterized in that: Pushing the content of products that the user is interested in to the user according to the recommendation list refers to defining the push frequency and time to perform instant push according to the user's online status and active time, real-time monitoring and recording of the user's reaction to the pushed products, integrating the collected user feedback into the user portrait, and updating the user's preferences and behavior patterns in real time.

8. A product push system based on user preference analysis based on the product push method based on user preference analysis according to any one of claims 1 to 7, characterized in that: include, A data collection module, which is used to collect user behavior data, product data, and context data from multiple channels and pre-process the data; A construction module is used to extract key features from preprocessed data to build user portraits and to build a product knowledge graph based on the analysis of product data; Evaluation module, which is used to calculate the importance index of each product feature and generate feature-sentiment pairs by combining the sentiment score and importance score of the product feature; The analysis module is used to construct a user-item interaction graph based on user profiles and feature-sentiment pairs, learn the embedded representation of items from the interaction graph using the GraphSAGE model, and calculate the context-aware similarity between items; The recommendation module is used to generate a preliminary recommendation list based on similarity and use the GBDT model to personalize the sorting of the list to finally generate a personalized recommendation list, push products based on the list and collect user feedback.

9. A computer device comprising: Memory and processor; The memory stores a computer program, characterized in that: when the processor executes the computer program, the steps of the product push method based on user preference analysis described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the product push method based on user preference analysis described in any one of claims 1 to 7 are implemented.

Citation Information

Patent Citations

  • Expression input method based on BERT technology and device thereof

    CN110543242A

  • User comment text emotion mining model for automatically extracting fine-grained attributes

    CN114564956A

  • Commodity recommendation method integrating comment emotion and score based on knowledge graph

    CN115170216A

  • Multi-behavior recommendation method and system based on multi-graph semantic fusion

    CN115618125A

  • Comment information integrated e-commerce network commodity recommendation method and system

    CN116579824A

Cited By

  • Time perception user portrait modeling method based on sentiment analysis

    CN119537704A

  • Personalized content recommendation and ROI (Region of Interest) improvement method and system based on multi-dimensional user portraits

    CN120125294A

  • Personalized content recommendation and ROI improvement method and system based on multi-dimensional user portrait

    CN120125294B

  • Commodity recommendation method and system based on big data analysis

    CN120125324A

  • Knowledge graph-based identification analysis system

    CN120337931A