A personalized clothing recommendation method and system for e-commerce platforms based on artificial intelligence
By building multi-level user portraits and using hybrid recommendation algorithms, combined with style tag library, data sparseness and cold start problems in the recommendation system of e-commerce platform are solved, the accuracy and transparency of personalized clothing recommendations are achieved, and users are improved Satisfaction and commercial operational efficiency.
Patent Information
- Application Number
- CN202411104217.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-13
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-08-13
AI Technical Summary
The existing e-commerce platform recommendation system has data sparsity, cold start problems, and optimization problems that are difficult to dynamically capture user preferences, resulting in poor recommendation results and lack of personalization and explanation.
By collecting basic user information and behavioral data, building multi-level user portraits, combining multi-dimensional features of the product, using a hybrid recommendation algorithm to generate personalized recommendation results, establishing a style tag library for classification and update, and providing personalized clothing recommendations.
It improves the accuracy and personalization of recommendations, meets users' personalized needs, enhances user shopping experience and satisfaction, and provides e-commerce platforms with user preference analysis and purchase prediction support, and optimizes inventory management and marketing strategies.
Smart Images

Figure CN118628214B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of e-commerce, and in particular to a method and system for recommending personalized clothing on an e-commerce platform based on artificial intelligence. Background Art
[0002] In current e-commerce platforms, recommendation systems are widely used to improve user shopping experience and platform sales. Traditional recommendation systems mainly rely on collaborative filtering and content recommendation technologies. The collaborative filtering-based method analyzes the user's historical behavior data to find other users with similar behaviors to the current user and recommend products that these similar users like. However, this method faces data sparsity and cold start problems, resulting in poor recommendation results. The content recommendation-based method recommends products similar to those that the user liked in the past by analyzing product features, but it is difficult to capture the user's potential interests and preference changes. Existing systems usually lack the interpretability of recommendation results, making it difficult for users to understand the reasons for the recommendation, which reduces user satisfaction and trust. In addition, traditional methods fail to make full use of users' basic information and behavior data, resulting in recommendation results that are difficult to meet users' personalized needs. Summary of the invention
[0003] In view of the above-mentioned problems, the present invention is proposed.
[0004] Therefore, the technical problem solved by the present invention is: the existing recommendation methods have data sparsity, cold start problems, and the optimization problem of how to dynamically capture user preferences.
[0005] In order to solve the above technical problems, the present invention provides the following technical solutions: a personalized clothing recommendation method for an e-commerce platform based on artificial intelligence, comprising:
[0006] Collect basic user information and user behavior data to build user portraits;
[0007] The user portrait construction includes constructing a first user portrait based on basic user information and constructing a second user portrait based on user behavior data; extracting features of products on the e-commerce platform and matching them with the first user portrait, calculating the matching degree between the user portrait and the comprehensive feature vector of the product, and establishing a style tag library for classification based on the matching results;
[0008] Based on the second user portrait, a hybrid recommendation algorithm is used to generate personalized recommendation results, and the personalized recommendation results are updated through the style tag library to perform personalized clothing recommendations.
[0009] As a preferred solution of the personalized clothing recommendation method for an e-commerce platform based on artificial intelligence described in the present invention, wherein: the basic information of the user includes age, gender, geographical location, occupation, and income level;
[0010] The user behavior data includes a list of recently browsed product IDs, browsing frequency, a list of recently purchased product IDs, purchase frequency, average ratings, evaluation sentiment analysis results, a list of recently searched keywords, search frequency, a list of recently clicked advertisement IDs, click frequency, a list of shared product IDs, and comment sentiment analysis results.
[0011] As a preferred solution of the personalized clothing recommendation method for an e-commerce platform based on artificial intelligence described in the present invention, wherein: the establishment of a style tag library for classification includes obtaining product images from a product database of the e-commerce platform, resizing, denoising and standardizing the images; using a ResNet50 model to perform preliminary feature extraction on the preprocessed product images, and extracting visual features of color, material and style; using a visual transformation network to further process the preliminary extracted visual features, and extract high-level features of the image, the high-level features including global structural features and local pattern features of the image; extracting text data from the product description, the text data including brand, style, material and price, cleaning and standardizing the text data, and using a BERT model to extract text features to generate a text feature vector representing the semantics of the product description; the high-level features and text features of the image are standardized respectively, and the high-level features and text features of the standardized image are merged to construct a comprehensive feature vector of the product.
[0012] The comprehensive feature vector of the product is preliminarily matched with the first user portrait through the matching degree between the user portrait and the comprehensive feature vector of the product, and a preliminary matching product list is generated for each user; based on the preliminary matching, the label propagation algorithm and graph convolutional network are used to refine the classification of the products according to the preferences of the user portrait and the matching degree of the comprehensive feature vector of the product, and the products are subdivided into specific subcategories; based on the results of the refined classification, a style label is generated for each product, and the style label includes color, material, style, brand and price, and the style label is stored to generate a style label library.
[0013] As a preferred solution of the personalized clothing recommendation method for an e-commerce platform based on artificial intelligence described in the present invention, the matching degree between the calculated user portrait and the comprehensive feature vector of the commodity is expressed as:
[0014]
[0015]
[0016]
[0017]
[0018] in, represents cosine similarity, A represents user portrait vector, B represents product comprehensive feature vector, represents the norm of the user portrait vector, represents the norm of the comprehensive feature vector of the product, represents the adjusted comprehensive feature vector of the product, represents the mean of the comprehensive feature vector of the commodity, S represents the covariance matrix, Indicates the number of preliminary matching products. Indicates The adjusted comprehensive feature vector of the initially matched products, Indicates The transpose of the difference between the comprehensive feature vector of the initially matched products and the mean after adjustment, represents the Mahalanobis distance, represents the transpose of the difference between the user portrait vector and the adjusted product comprehensive feature vector, represents the inverse matrix of the covariance matrix;
[0019] The generating of the preliminary matching commodity list includes setting a matching degree threshold and screening out commodities whose Mahalanobis distance is less than the matching degree threshold.
[0020] As a preferred solution of the personalized clothing recommendation method for an e-commerce platform based on artificial intelligence described in the present invention, wherein: the use of a hybrid recommendation algorithm to generate personalized recommendation results includes analyzing the similarity between users and commodities through a collaborative filtering algorithm, calculating a similarity matrix using the user's rating data, calculating the similarity between the user and the commodity, and obtaining a similarity matrix between the user and the commodity;
[0021] According to the similarity matrix, filter out a list of products similar to the current user's preferences, and select the products with the highest similarity as candidate products;
[0022] Using the content recommendation algorithm, extract keywords and tag features from product descriptions, use natural language processing technology to perform text analysis, build candidate product feature vectors, and generate initial user behavior feature vectors based on the user's browsing frequency, purchase frequency, recently browsed product ID list, recently purchased product ID list, recently searched keyword list, recently clicked ad ID list, shared product ID list, and evaluation sentiment analysis results. Then, perform weighted average of the candidate product feature vectors that the user has interacted with in the past to form the final user behavior feature vector.
[0023] Calculate the similarity between the user behavior feature vector and the candidate product feature vector, recommend products similar to those that the user liked in the past, and generate a list of recommended products;
[0024] Use matrix decomposition technology to decompose the user-item rating matrix, extract potential features, and use the user's historical rating data to construct a user-item rating matrix, where rows represent users, columns represent items, and values are users' ratings of items.
[0025] Decompose the rating matrix into a user feature matrix and a product feature matrix, extract the potential features of users and products, and use the singular value decomposition method to obtain the user feature vector and the candidate product feature vector;
[0026] Using the decomposed feature matrix, predict the possible ratings of users for unrated products, calculate the inner product of the user feature vector and the candidate product feature vector, and get the predicted rating;
[0027] The similarity matrix, similarity and predicted score are combined. For each product, weights are assigned according to the results of collaborative filtering, content recommendation and matrix decomposition, and a comprehensive score is calculated. According to the comprehensive score, the products are sorted from high to low to generate a preliminary recommendation list.
[0028] As a preferred solution of the personalized clothing recommendation method for an e-commerce platform based on artificial intelligence described in the present invention, wherein: the updating of the personalized recommendation results through the style tag library includes checking the tag matching of the preliminary recommended products, if there is no corresponding tag, marking the corresponding product as a missing tag, comparing the product with the missing tag with other products in the tag library, and using the tags of similar products to complete the missing tags;
[0029] If the features of the initially recommended product have corresponding tags in the style tag library, extract the color, material, style, brand and price feature tags related to the initially recommended product from the style tag library;
[0030] Calculate the label matching score of each preliminary recommended product, calculate the behavior feature score for each product based on the user's second user profile, combine the label matching score with the behavior feature score, and calculate the comprehensive matching score of each product through the weighted average method. According to the comprehensive matching score, sort the preliminary recommended products from high to low to generate the final recommendation list.
[0031] As a preferred solution of the method for personalized clothing recommendation on an e-commerce platform based on artificial intelligence described in the present invention, the personalized clothing recommendation includes sorting the recommended products according to the comprehensive matching scores in the user interface of the e-commerce platform and displaying them on the user's recommendation page;
[0032] On the recommended product display page, display the style label of each product. On the recommended product page, provide the recommendation reason for each recommended product and generate a recommendation explanation. Add a user feedback mechanism so that users can rate each recommended product or mark whether they like it.
[0033] Dynamically adjust the recommendation list based on real-time user feedback, use real-time data and instant user feedback to update user portraits and adjust recommendation results.
[0034] Another object of the present invention is to provide an artificial intelligence-based personalized clothing recommendation system for an e-commerce platform, which can solve the data sparsity, cold start problem, and lack of personalization problems in existing recommendation systems by constructing an artificial intelligence-based personalized clothing recommendation system for an e-commerce platform.
[0035] To solve the above technical problems, the present invention provides the following technical solutions: a personalized clothing recommendation system for an e-commerce platform based on artificial intelligence, comprising: a data acquisition module, a feature extraction module and a clothing recommendation module; the data acquisition module is used to collect basic user information and user behavior data to construct a user portrait; the construction of the user portrait includes constructing a first user portrait based on the user's basic information and constructing a second user portrait based on the user's behavior data; the feature extraction module is used to extract features of products on the e-commerce platform, and match them with the first user portrait, calculate the matching degree between the user portrait and the comprehensive feature vector of the product, and establish a style tag library for classification based on the matching results; the clothing recommendation module is used to generate personalized recommendation results based on the second user portrait using a hybrid recommendation algorithm, update the personalized recommendation results through the style tag library, and perform personalized clothing recommendations.
[0036] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method for recommending personalized clothing on an e-commerce platform based on artificial intelligence are implemented as described above.
[0037] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for recommending personalized clothing on an e-commerce platform based on artificial intelligence as described above.
[0038] Beneficial effects of the invention: The artificial intelligence-based personalized clothing recommendation method for an e-commerce platform provided by the invention constructs a multi-level user portrait by comprehensively utilizing the user's basic information and behavioral data, combines the multi-dimensional characteristics of the goods, and uses a hybrid recommendation algorithm to generate personalized recommendation results, thereby improving the accuracy and personalization of the recommendation.
[0039] It can meet the personalized needs of users, enhance their shopping experience and satisfaction, and at the same time provide user preference analysis and purchase forecast support to the management of e-commerce platforms, optimize inventory management and marketing strategies, and improve business operation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] 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.
[0041] Figure 1 An overall flow chart of a method for personalized clothing recommendation on an e-commerce platform based on artificial intelligence provided by one embodiment of the present invention.
[0042] Figure 2 An overall structural diagram of an artificial intelligence-based personalized clothing recommendation system for an e-commerce platform provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0043] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.
[0044] 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.
[0045] Example 1, reference Figure 1 , which is an embodiment of the present invention, provides a personalized clothing recommendation method for an e-commerce platform based on artificial intelligence, comprising:
[0046] Collect basic user information and user behavior data to build user portraits;
[0047] The constructing of the user portrait includes constructing a first user portrait based on the user's basic information and constructing a second user portrait based on the user's behavior data;
[0048] Extract features of products on the e-commerce platform and match them with the first user portrait, calculate the matching degree between the user portrait and the comprehensive feature vector of the product, and establish a style tag library for classification based on the matching results;
[0049] Based on the second user portrait, a hybrid recommendation algorithm is used to generate personalized recommendation results, and the personalized recommendation results are updated through the style tag library to perform personalized clothing recommendations.
[0050] The user's basic information includes age, gender, geographic location, occupation, and income level;
[0051] The user behavior data includes a list of recently browsed product IDs, browsing frequency, a list of recently purchased product IDs, purchase frequency, average ratings, evaluation sentiment analysis results, a list of recently searched keywords, search frequency, a list of recently clicked advertisement IDs, click frequency, a list of shared product IDs, and comment sentiment analysis results.
[0052] Basic information, including age, gender, geographic location, occupation, income level, etc., is extracted from user registration forms and profile settings, and database query and form processing techniques are used to ensure data integrity and accuracy. Users are required to fill in their age, gender, geographic location, occupation, and income level when registering or setting up their personal profile. In order to ensure data accuracy, database query and form processing techniques are used to verify and standardize the data. Geographic location information is obtained through the user's IP address or GPS data, and the API interface is called to obtain detailed address information and encode it. Occupational information is encoded using predefined occupational classification standards to ensure data consistency and availability. Next, this basic information is processed through data aggregation and analysis techniques, using data cleaning tools (Pandas) to remove invalid or duplicate data to ensure data quality, and using database management (MySQL) to store the sorted user basic information to provide data support for subsequent user portrait construction. Finally, through a real-time synchronization mechanism, the update and accuracy of user basic information is ensured so that this information can be accessed and processed quickly and efficiently when needed.
[0053] The user behavior data is collected in real time through website logs, transaction records, user reviews, search history, ad click logs, and social media interactions. These data include the list of recently browsed product IDs, browsing frequency, the list of recently purchased product IDs, purchase frequency, average ratings, evaluation sentiment analysis results, the list of recently searched keywords, search frequency, the list of recently clicked ad IDs, click frequency, the list of shared product IDs, and comment sentiment analysis results. Using logging and data mining technology, the user behavior data is captured in real time through the event stream processing engine, and the data is preprocessed using data cleaning tools (such as Pandas) to ensure data consistency and integrity. Data cleaning includes steps such as removing outliers, filling missing data, and data format conversion. Next, these preprocessed data are stored in the big data platform through data warehouse technology for subsequent analysis and processing. Stream processing technology (Apache Kafka) is also used for real-time data processing to ensure the timely update of user behavior data and provide accurate and up-to-date behavior data for the construction of the second-layer user portrait.
[0054] Constructing a user portrait includes constructing a first user portrait based on the user's basic information. The user's basic information includes age, gender, geographic location, occupation, and income level. This basic information is extracted from the user registration form and personal profile settings, and the database query and form processing technology is used to ensure the integrity and accuracy of the data. The user's detailed address information is obtained through the API interface and coded; the occupational information is coded using the predefined occupational classification standard. Then, this basic information is summarized using data aggregation and analysis technology, and the user's basic information is clustered using a data clustering algorithm (K-Means) to generate the first user portrait. The first user portrait is based on the user's age, gender, geographic location, occupation, and income level. The relationship between the user's basic information is represented by a graph model (KnowledgeGraph) to ensure the accuracy and completeness of the user portrait. Finally, the generated first user portrait is stored in the database to ensure efficient query and update.
[0055] Building a user profile also includes building a second user profile based on user behavior data. User behavior data includes a list of recently browsed product IDs, browsing frequency, a list of recently purchased product IDs, purchase frequency, average ratings, evaluation sentiment analysis results, a list of recently searched keywords, search frequency, a list of recently clicked ad IDs, click frequency, a list of shared product IDs, and comment sentiment analysis results. These behavior data are collected in real time through website logs, transaction records, user reviews, search history, ad click logs, and social media interactions. Using log records and data mining techniques, user behavior data is captured in real time through an event stream processing engine, and data cleaning tools (Pandas) are used to pre-process the data to ensure data consistency and integrity. Next, data mining and machine learning techniques are used to analyze user behavior data, and clustering algorithms (K-Means) and classification algorithms are used to analyze and model user behavior data. A second user profile is generated based on the user's browsing habits, purchasing habits, evaluation tendencies, search habits, and advertising interactions. The second user portrait uses a graph model to represent the relationship between user behavior characteristics, ensuring that the user portrait can be dynamically updated to reflect the latest behavior data. The generated second user portrait is finally stored in the database (MongoDB) to ensure efficient query and update, and the second user portrait is updated in real time to reflect the user's latest behavior data, ensuring the accuracy and personalization of the recommendation results.
[0056] The establishment of the style tag library for classification includes obtaining product images from the product database of the e-commerce platform, resizing, denoising and standardizing the images; using the ResNet50 model to perform preliminary feature extraction on the preprocessed product images, and extracting visual features of color, material and style; using a visual transformation network (Transformer network) to further process the initially extracted visual features, and extract high-level features of the image, and the high-level features include global structural features and local pattern features of the image; extracting text data from the product description, the text data includes brand, style, material and price, cleaning and standardizing the text data, and using the BERT model to extract text features to generate a text feature vector representing the semantics of the product description; standardizing the high-level features and text features of the image respectively, fusing the standardized high-level features and text features of the image, and constructing a comprehensive feature vector of the product;
[0057] Global structural features are the overall layout, shape, and structural information of the image; local pattern features are the local details and texture patterns in the image;
[0058] The comprehensive feature vector of the product is preliminarily matched with the first user portrait through the matching degree between the user portrait and the comprehensive feature vector of the product, and a preliminarily matching product list is generated for each user. Based on the preliminary matching, the label propagation algorithm and graph convolutional network are used to refine the classification of products according to the preferences of the user portrait and the matching degree of the comprehensive feature vector of the product, and the products are subdivided into specific subcategories, such as dresses, jackets and pants, and the label propagation algorithm is used to perform label propagation on the preliminarily matched products.
[0059] Use graph convolutional networks to refine the classification of products and generate specific subcategories;
[0060] Based on the results of the refined classification, a style label is generated for each product. The style label includes color, material, style, brand and price. The style label is generated by combining the results of image and text feature extraction through predefined labeling rules. A relational database is used to store the generated style label library, and the database table structure of the label library is designed, including product ID, style label and comprehensive feature vector. The refined classification results and labels are stored in the database.
[0061] The matching degree between the calculated user portrait and the comprehensive feature vector of the product is expressed as:
[0062]
[0063]
[0064]
[0065]
[0066] in, represents cosine similarity, A represents user portrait vector, B represents product comprehensive feature vector, represents the norm of the user portrait vector, represents the norm of the comprehensive feature vector of the product, represents the adjusted comprehensive feature vector of the product, represents the mean of the comprehensive feature vector of the commodity, S represents the covariance matrix, Indicates the number of preliminary matching products. Indicates The adjusted comprehensive feature vector of the initially matched products, Indicates The transpose of the difference between the comprehensive feature vector of the initially matched products and the mean after adjustment, represents the Mahalanobis distance, represents the transpose of the difference between the user portrait vector and the adjusted product comprehensive feature vector, represents the inverse matrix of the covariance matrix;
[0067] In the personalized clothing recommendation method of e-commerce platform based on artificial intelligence, the existing technology uses cosine similarity and Mahalanobis distance separately, each with its own advantages and disadvantages. Cosine similarity can effectively measure the angular similarity between the comprehensive feature vectors of users and commodities, focusing on its directionality while ignoring the size of the vector, and performs well in capturing the user's preference direction. However, this method does not take into account the correlation and distribution between features, resulting in insufficient matching accuracy and failure to fully reflect the complex relationship between users and commodities. In the personalized clothing recommendation of e-commerce platforms, this deficiency may lead to inaccurate and indivisible recommendation results. On the other hand, Mahalanobis distance can consider the correlation between features, is suitable for multi-dimensional data, and adjusts the influence of different features through the covariance matrix to make the matching more accurate. However, the computational complexity is high, especially when the feature dimension is high, the calculation of the covariance matrix and the solution of the inverse matrix are very time-consuming, and it is sensitive to the quality of the covariance matrix estimation and is easily affected by noise data, which may lead to low computational efficiency and unsatisfactory recommendation results in the actual clothing recommendation system of e-commerce platforms.
[0068] Our invention calculates cosine similarity and uses its result as a weight to adjust the comprehensive feature vector of the product, so that the products with high similarity occupy a larger proportion in the subsequent Mahalanobis distance calculation. This adjustment method not only effectively combines cosine similarity with Mahalanobis distance, but also more accurately reflects the relationship between user preferences and product features through weighting. The cosine similarity is used to adjust the comprehensive feature vector of the product to make it more consistent with the user portrait vector. Then, the adjusted comprehensive feature vector of the product and the covariance matrix are used to calculate the Mahalanobis distance, which can better consider the correlation and distribution between the features, thereby further improving the accuracy of the match. In the personalized clothing recommendation of the e-commerce platform, this method can more accurately capture the real preferences of users and provide more accurate recommendation results. Finally, the matching threshold is set according to the calculated Mahalanobis distance, and the products with a distance less than the threshold are screened out to generate a preliminary matching product list. This mechanism ensures that the selected products have a higher matching degree with the user portrait, thereby improving the relevance of the recommendation and user satisfaction.
[0069] By combining cosine similarity and Mahalanobis distance, our invention has achieved unexpected results in improving matching accuracy, reducing computational complexity and enhancing robustness in the personalized clothing recommendation method for e-commerce platforms based on artificial intelligence. Cosine similarity does not directly perform preliminary screening, but adjusts the comprehensive feature vector of the products to enhance the weight of high-similar products in the Mahalanobis distance calculation; Mahalanobis distance provides a more accurate matching metric by considering the correlation between features. In addition, the comprehensive feature vector of the products adjusted by cosine similarity makes the calculation of Mahalanobis distance more efficient and reduces the computational complexity. Combining these two methods makes the algorithm more robust to noise and abnormal data, and can still maintain efficient and accurate matching performance in complex actual data. This innovative combination not only avoids the drawbacks of simple technology superposition, but also achieves a recommendation effect beyond expectations through scientific design and optimization. In practical applications, this method significantly improves the accuracy and efficiency of personalized recommendations, provides users with a more satisfactory shopping experience, and thus has important practical application value in e-commerce platforms.
[0070] Although cosine similarity and Mahalanobis distance have unique advantages in their respective fields, in this field, the use of these algorithms alone can already meet basic matching needs, so there is no need to simply combine the two algorithms. Traditional methods usually use a single algorithm to generate a preliminary matching product list, such as using only cosine similarity for directional matching or only using Mahalanobis distance for multi-dimensional feature matching. However, simply superimposing the two cannot give full play to their respective advantages, but may introduce problems of computational complexity and accuracy. Through the method invented by us, the result of cosine similarity is used as the weight to adjust the comprehensive feature vector of the product, and then combined with Mahalanobis distance for precise matching. This combination is not a simple technical superposition, but a scientifically designed weighting and adjustment to solve the limitations of each algorithm.
[0071] The generating of the preliminary matching commodity list includes setting a matching degree threshold and screening out commodities whose Mahalanobis distance is less than the matching degree threshold.
[0072] The method of generating personalized recommendation results by using a hybrid recommendation algorithm includes analyzing the similarity between users and products by a collaborative filtering algorithm, calculating a similarity matrix using the user's rating data, calculating the similarity between users and products, and obtaining a similarity matrix between users and products;
[0073] According to the similarity matrix, filter out a list of products similar to the current user's preferences, and select the products with the highest similarity as candidate products;
[0074] Using the content recommendation algorithm, extract keywords and tag features from product descriptions, use natural language processing technology to perform text analysis, build candidate product feature vectors, and generate initial user behavior feature vectors based on the user's browsing frequency, purchase frequency, recently browsed product ID list, recently purchased product ID list, recently searched keyword list, recently clicked ad ID list, shared product ID list, and evaluation sentiment analysis results. Then, perform weighted average of the candidate product feature vectors that the user has interacted with in the past to form the final user behavior feature vector.
[0075] Calculate the similarity between the user behavior feature vector and the candidate product feature vector, recommend products similar to those that the user liked in the past, and generate a list of recommended products;
[0076] Use matrix decomposition technology to decompose the user-item rating matrix, extract potential features, and use the user's historical rating data to construct a user-item rating matrix, where rows represent users, columns represent items, and values are users' ratings of items.
[0077] Decompose the rating matrix into a user feature matrix and a product feature matrix, extract the potential features of users and products, and use the singular value decomposition method to obtain the user feature vector and the candidate product feature vector;
[0078] Using the decomposed feature matrix, predict the possible ratings of users for unrated products, calculate the inner product of the user feature vector and the candidate product feature vector, and get the predicted rating;
[0079] The similarity matrix, similarity and predicted score are combined. For each product, weights are assigned according to the results of collaborative filtering, content recommendation and matrix decomposition, and a comprehensive score is calculated. According to the comprehensive score, the products are sorted from high to low to generate a preliminary recommendation list.
[0080] Personalized recommendations are achieved through the combination of collaborative filtering, content recommendation, and matrix decomposition technology. The collaborative filtering algorithm is based on user rating data, the content recommendation algorithm is based on product features, and the matrix decomposition technology extracts the potential features of users and products. The results of these algorithms are combined to generate a personalized preliminary recommendation list. The fusion recommendation algorithm uses a comprehensive scoring mechanism to weight the results of different algorithms to ensure the reliability of the recommendation results. Compared with the traditional single recommendation algorithm, this multi-algorithm fusion method improves the accuracy of the recommendation results.
[0081] The updating of personalized recommendation results through the style tag library includes checking the tag matching of the initially recommended products. If there is no corresponding tag, the corresponding product is marked as missing a tag, the product with the missing tag is compared with other products in the tag library, and the missing tag is completed with the tags of similar products;
[0082] If the features of the initially recommended product have corresponding tags in the style tag library, extract the color, material, style, brand and price feature tags related to the initially recommended product from the style tag library;
[0083] Calculate the label matching score of each preliminary recommended product, calculate the behavior feature score for each product based on the user's second user profile, combine the label matching score with the behavior feature score, and calculate the comprehensive matching score of each product through the weighted average method. According to the comprehensive matching score, sort the preliminary recommended products from high to low to generate the final recommendation list.
[0084] By checking the label matching of the initially recommended products, the problem of incomplete product label information is solved. The missing labels are completed with the labels of similar products to ensure the integrity of the recommendation results. The label matching score and the behavior feature score are combined, and the comprehensive matching score is calculated by weighted average to optimize the recommendation results. The label completion mechanism and comprehensive score optimization ensure multi-dimensional consideration and optimization of the recommendation results. Compared with traditional methods, it can handle the problem of missing product label information and improve the integrity of the recommendation results.
[0085] The making of personalized clothing recommendations includes sorting the recommended products according to the comprehensive matching scores in the user interface of the e-commerce platform and displaying them on the user's recommendation page;
[0086] On the recommended product display page, display the style label of each product. On the recommended product page, provide the recommendation reason for each recommended product and generate a recommendation explanation. Add a user feedback mechanism so that users can rate each recommended product or mark whether they like it.
[0087] Dynamically adjust the recommendation list based on real-time user feedback, use real-time data and instant user feedback to update user portraits and adjust recommendation results.
[0088] By displaying recommended products in the user interface of the e-commerce platform and providing reasons for recommendation and user feedback mechanism, the transparency and interactivity of personalized clothing recommendation are achieved. Users can rate and provide feedback on recommended products, and the recommendation list is dynamically adjusted according to real-time feedback to ensure the timeliness and accuracy of recommendation results. Through the user feedback mechanism, dynamic adjustment and optimization of the recommendation system are achieved. Compared with the traditional static recommendation system, the method of the present invention can learn and optimize the recommendation results in real time, improving user satisfaction.
[0089] Example 2, reference Figure 2 , which is an embodiment of the present invention, provides an e-commerce platform personalized clothing recommendation system based on artificial intelligence, including:
[0090] Data collection module, feature extraction module and clothing recommendation module;
[0091] The data collection module is used to collect basic user information and user behavior data and build user portraits;
[0092] Building a user portrait includes building a first user portrait based on basic user information and building a second user portrait based on user behavior data;
[0093] The feature extraction module is used to extract features of products on the e-commerce platform and match them with the first user portrait, calculate the matching degree between the user portrait and the comprehensive feature vector of the product, and establish a style tag library for classification based on the matching results;
[0094] The clothing recommendation module is used to generate personalized recommendation results based on the second user portrait using a hybrid recommendation algorithm, update the personalized recommendation results through the style tag library, and perform personalized clothing recommendations.
[0095] Embodiment 3, an embodiment of the present invention, is different from the first two embodiments in that:
[0096] 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 to enable a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments 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.
[0097] 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.
[0098] 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.
[0099] 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-mentioned embodiments, multiple 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, it can be implemented by any one of the following technologies known in the art or their combination: 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.
[0100] Example 4 is an embodiment of the present invention, which provides a personalized clothing recommendation method for an e-commerce platform based on artificial intelligence. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through simulation experiments.
[0101] Under the same hardware and software environment, we selected an e-commerce platform dataset containing rich basic user information and behavior data, including user age, gender, geographic location, occupation, income level, recently browsed product ID list, browsing frequency, recently purchased product ID list, purchase frequency, average rating, evaluation sentiment analysis results, recently searched keyword list, search frequency, recently clicked ad ID list, click frequency, shared product ID list, comment sentiment analysis results, etc. We selected 1,000 users as experimental subjects and used 10,000 products for recommendation testing.
[0102] Traditional method 1 builds a simple user profile based on the user's basic information (age, gender, geographic location), uses a recommendation method based on collaborative filtering algorithm, mainly calculates the similarity matrix based on user rating data, and recommends products similar to the current user's preferences. After the recommendation list is generated, it is displayed to the user, and the user rates the recommendation results. The system does not make dynamic adjustments.
[0103] Traditional method 2 builds a user profile based on the user's basic information and some behavior data (browsing frequency and purchase frequency), uses a content recommendation algorithm to extract keywords and tag features from product descriptions, and makes recommendations based on user behavior data. After generating a recommendation list, it is displayed to the user, who rates the recommendation results, and the system makes limited dynamic adjustments.
[0104] The method invented by us first constructs the first-layer user portrait through the basic information of the user, and constructs the second-layer user portrait through the user behavior data. The product features are extracted and classified. The visual features of the product image are extracted using the ResNet50 model. The high-level features of the product image are further extracted using the visual transformation network. The feature vector of the product text description is extracted using the BERT model. The image and text features are integrated to construct a comprehensive feature vector of the product, and a style tag library is established. The hybrid recommendation algorithm is used to generate personalized recommendation results. The similarity between users and products is analyzed by the collaborative filtering algorithm. The keywords and tag features are extracted from the product description using the content recommendation algorithm. The products that the user liked in the past are recommended by combining the similarity between the user behavior feature vector and the candidate product feature vector. The user-item rating matrix is decomposed by matrix decomposition technology to extract potential features and generate a preliminary recommendation list. The personalized recommendation results are updated through the style tag library, the label matching of the preliminary recommended products is checked, the missing labels are completed, the style labels of the preliminary recommended products are extracted, the label matching scores are calculated, the comprehensive matching scores are calculated by combining the behavior feature scores, the preliminary recommended products are sorted, and the final recommendation list is generated. The experimental results are shown in Table 1.
[0105] Table 1 Comparison of experimental results
[0106]
[0107] Traditional recommendation systems fail to fully utilize users' basic information and behavior data, resulting in recommendation results that are difficult to meet users' personalized needs. The present invention constructs a multi-level user portrait by comprehensively utilizing users' basic information and behavior data, thereby improving the personalization of recommendation results. Existing systems lack the ability to explain recommendation results, making it difficult for users to understand the reasons for recommendations. The present invention provides recommendation reasons and user feedback mechanisms on the recommended product page, which increases the transparency of recommendation results and improves user satisfaction and trust. Traditional recommendation systems find it difficult to dynamically capture changes in user preferences. The present invention dynamically adjusts recommendation results by updating user portraits and recommendation results in real time to maintain the accuracy and timeliness of recommendation results.
[0108] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.
Claims
1. A personalized clothing recommendation method for an e-commerce platform based on artificial intelligence, characterized in that: include: Collect basic user information and user behavior data to build user portraits; The constructing of the user portrait includes constructing a first user portrait based on the user's basic information and constructing a second user portrait based on the user's behavior data; Extract features of products on the e-commerce platform and match them with the first user portrait, calculate the matching degree between the user portrait and the comprehensive feature vector of the product, and establish a style tag library for classification based on the matching results; Generate personalized recommendation results based on the second user portrait using a hybrid recommendation algorithm, update the personalized recommendation results through the style tag library, and make personalized clothing recommendations; The matching degree between the calculated user portrait and the comprehensive feature vector of the product is expressed as: in, represents cosine similarity, A represents user portrait vector, B represents product comprehensive feature vector, represents the norm of the user portrait vector, represents the norm of the comprehensive feature vector of the product, represents the adjusted comprehensive feature vector of the product, represents the mean of the comprehensive feature vector of the commodity, S represents the covariance matrix, Indicates the number of preliminary matching products. Indicates The adjusted comprehensive feature vector of the initially matched products, Indicates The transpose of the difference between the comprehensive feature vector of the initially matched products and the mean after adjustment, represents the Mahalanobis distance, represents the transpose of the difference between the user portrait vector and the adjusted product comprehensive feature vector, represents the inverse matrix of the covariance matrix; The user's basic information includes age, gender, geographic location, occupation, and income level; The user behavior data includes a list of recently browsed product IDs, browsing frequency, a list of recently purchased product IDs, purchase frequency, average rating, evaluation sentiment analysis results, a list of recently searched keywords, search frequency, a list of recently clicked advertisement IDs, click frequency, a list of shared product IDs, and comment sentiment analysis results; The establishment of the style tag library for classification includes obtaining product images from the product database of the e-commerce platform, resizing, denoising and standardizing the images; using the ResNet50 model to perform preliminary feature extraction on the preprocessed product images, and extracting visual features of color, material and style; using the visual transformation network to further process the preliminary extracted visual features, and extract high-level features of the image, and the high-level features include global structural features and local pattern features of the image; extracting text data from the product description, the text data includes brand, style, material and price, cleaning and standardizing the text data, and using the BERT model to extract text features to generate a text feature vector representing the semantics of the product description; standardizing the high-level features and text features of the image respectively, fusing the standardized high-level features and text features of the image, and constructing a comprehensive feature vector of the product; The comprehensive feature vector of the product is preliminarily matched with the first user portrait through the matching degree between the user portrait and the comprehensive feature vector of the product, and a preliminary matching product list is generated for each user; based on the preliminary matching, the label propagation algorithm and graph convolutional network are used to refine the classification of the products according to the matching degree between the user portrait preference and the comprehensive feature vector of the product, and the products are subdivided into specific subcategories; Based on the results of the refined classification, a style tag is generated for each product. The style tag includes color, material, style, brand and price, and the style tags are stored to generate a style tag library; Generating a preliminary matching product list includes setting a matching threshold and screening out products whose Mahalanobis distance is less than the matching threshold; The method of generating personalized recommendation results by using a hybrid recommendation algorithm includes analyzing the similarity between users and products by a collaborative filtering algorithm, calculating a similarity matrix using the user's rating data, calculating the similarity between users and products, and obtaining a similarity matrix between users and products; According to the similarity matrix, filter out a list of products similar to the current user's preferences, and select the products with the highest similarity as candidate products; Using the content recommendation algorithm, extract keywords and tag features from product descriptions, use natural language processing technology to perform text analysis, build candidate product feature vectors, and generate initial user behavior feature vectors based on the user's browsing frequency, purchase frequency, recently browsed product ID list, recently purchased product ID list, recently searched keyword list, recently clicked ad ID list, shared product ID list, and evaluation sentiment analysis results. Then, perform weighted average of the candidate product feature vectors that the user has interacted with in the past to form the final user behavior feature vector. Calculate the similarity between the user behavior feature vector and the candidate product feature vector, recommend products similar to those that the user liked in the past, and generate a list of recommended products; Use matrix decomposition technology to decompose the user-item rating matrix, extract potential features, and use the user's historical rating data to construct a user-item rating matrix, where rows represent users, columns represent items, and values are users' ratings of items. Decompose the rating matrix into a user feature matrix and a product feature matrix, extract the potential features of users and products, and use the singular value decomposition method to obtain the user feature vector and the candidate product feature vector; Using the decomposed feature matrix, predict the possible ratings of users for unrated products, calculate the inner product of the user feature vector and the candidate product feature vector, and get the predicted rating; Combine the similarity matrix, similarity and predicted score, assign weights to each product based on the results of collaborative filtering, content recommendation and matrix decomposition, calculate the comprehensive score, sort from high to low based on the comprehensive score, and generate a preliminary recommendation list; The updating of personalized recommendation results through the style tag library includes checking the tag matching of the initially recommended products. If there is no corresponding tag, the corresponding product is marked as missing a tag, the product with the missing tag is compared with other products in the tag library, and the missing tag is completed with the tags of similar products; If the features of the initially recommended product have corresponding tags in the style tag library, extract the color, material, style, brand and price feature tags related to the initially recommended product from the style tag library; Calculate the label matching score of each preliminary recommended product, calculate the behavior feature score for each product based on the user's second user profile, combine the label matching score with the behavior feature score, and calculate the comprehensive matching score of each product through the weighted average method. According to the comprehensive matching score, sort the preliminary recommended products from high to low to generate the final recommendation list; The making of personalized clothing recommendations includes sorting the recommended products according to the comprehensive matching scores in the user interface of the e-commerce platform and displaying them on the user's recommendation page; On the recommended product display page, display the style label of each product. On the recommended product page, provide the recommendation reason for each recommended product and generate a recommendation explanation. Add a user feedback mechanism so that users can rate each recommended product or mark whether they like it. Dynamically adjust the recommendation list based on real-time user feedback, use real-time data and instant user feedback to update user portraits and adjust recommendation results.
2. A system using the method for personalized clothing recommendation on an e-commerce platform based on artificial intelligence as claimed in claim 1, characterized in that: include: Data collection module, feature extraction module and clothing recommendation module; The data collection module is used to collect basic user information and user behavior data and build a user portrait; The constructing of the user portrait includes constructing a first user portrait based on the user's basic information and constructing a second user portrait based on the user's behavior data; The feature extraction module is used to extract features of products on the e-commerce platform, match them with the first user portrait, calculate the matching degree between the user portrait and the comprehensive feature vector of the product, and establish a style tag library for classification based on the matching results; The clothing recommendation module is used to generate personalized recommendation results based on the second user portrait using a hybrid recommendation algorithm, update the personalized recommendation results through a style tag library, and perform personalized clothing recommendations.
3. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method for recommending personalized clothing on an e-commerce platform based on artificial intelligence described in claim 1 are implemented.
4. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the steps of the method for recommending personalized clothing on an e-commerce platform based on artificial intelligence described in claim 1 are implemented.
