An e-commerce intelligent recommendation system based on user behavior portrait

Abstract

This invention discloses an intelligent e-commerce recommendation system based on user behavior profiles, comprising a multimodal data processing module for fusing user behavior data into a unified user profile; a reinforcement learning recommendation module for continuously optimizing recommendation strategies based on user interactions with the recommendation system; a metric learning module for recommending the most relevant products and content to users by measuring the similarity between user preferences and product features; and a user sentiment analysis module for providing more personalized recommendations that better meet users' emotional needs by incorporating their emotional state. This invention achieves more accurate, personalized, and emotion-driven product recommendations through multimodal data fusion, reinforcement learning, and sentiment analysis technologies, and is applicable to scenarios such as social e-commerce platforms, online retail, and personalized recommendation systems.

Description

Technical Field

[0001] This invention relates to the fields of multimodal learning and data processing technology, and in particular to an intelligent recommendation system for e-commerce based on user behavior profiles. Background Technology

[0002] With the widespread adoption of the internet and social media platforms, e-commerce has gradually incorporated social elements, making social e-commerce a new trend. In traditional e-commerce recommendation systems, recommendation algorithms primarily rely on users' historical behavioral data (such as purchase and browsing history) and product attribute information. While these methods can provide some personalized recommendations, they have limitations in handling user interaction behavior, social circle influence, and emotional changes in the social e-commerce environment.

[0003] Most existing e-commerce recommendation systems employ collaborative filtering and content-based recommendation algorithms. Collaborative filtering methods make recommendations based on historical interaction data between users and products, but they perform poorly for new users and new products, and often fail to address the "cold start" problem. While content-based recommendation methods can utilize product features, they only consider the product's attributes and lack a comprehensive understanding of user interests. Furthermore, traditional recommendation systems often ignore user interactions on social media platforms (such as likes, comments, and shares), which effectively reflect user interests and emotional attitudes.

[0004] Social platforms provide e-commerce with more user behavior data, especially through user interactions, comments, and emotional expressions on social networks, enabling platforms to more accurately capture users' potential interests. However, existing social e-commerce recommendation systems still face many problems, mainly in three aspects: First, the modeling of social behavior data is not comprehensive enough, failing to fully explore the interaction relationships and dissemination paths within users' social circles; second, multimodal data (such as images, videos, and text) are not effectively combined, resulting in singular and limited recommendation effects; and finally, sentiment analysis technology is not fully utilized, and recommendation results lack emotional drive, failing to truly meet users' personalized needs.

[0005] Therefore, existing social e-commerce recommendation systems cannot effectively process the large amounts of behavioral data from social platforms, nor can they accurately uncover users' interests and emotional needs, resulting in low relevance and personalization levels in recommendations. These problems limit the intelligence and accuracy of recommendation systems, impacting user experience and the platform's commercial benefits. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent e-commerce recommendation system based on user behavior profiles. This invention fully utilizes multimodal data processing, reinforcement learning, metric learning, and sentiment analysis techniques, detailing how to provide users with accurate and personalized product or content recommendations by analyzing user behavior data on social platforms (such as likes, comments, and shares), as well as users' social circles and emotional states. Specifically, this invention integrates various user behavior data (such as text, images, and videos) through a deep learning model, combining social behavior and user sentiment information to dynamically optimize recommendation strategies, thereby improving recommendation accuracy and personalization. It possesses the following advantages: accurately capturing users' multi-dimensional interests and needs, achieving emotion-driven personalized recommendations, improving user satisfaction and engagement, and dynamically adapting to changes in user behavior, significantly enhancing the intelligence and commercial value of the recommendation system.

[0007] An intelligent recommendation system for e-commerce based on user behavior profiles according to an embodiment of the present invention includes: Multimodal data processing module: Used to process user behavior data from different data sources and integrate the user behavior data into a unified user profile; Reinforcement learning recommendation module: used to continuously optimize recommendation strategies based on user interactions with the recommendation system; The metric learning module is used to learn the similarity between users and products. By measuring the similarity between user preferences and product features, it recommends the most relevant products and content to users. User sentiment analysis module: This module analyzes users' emotional expressions on social platforms and combines these with users' emotional states to provide personalized recommendations that better meet users' emotional needs.

[0008] Optionally, modules can be integrated using the following methods: S1. Collect and clean user behavior data from different data sources; S2. Use multimodal learning methods to process user behavior data, which includes text, images, and videos. Convert text, images, and videos into a unified user profile to form a multi-dimensional user information model. S3. Based on the interaction feedback between users and the e-commerce intelligent recommendation system, continuously optimize the recommendation strategy. Use the deep Q network reinforcement learning algorithm to update the recommendation strategy after each user feedback to improve the personalization of recommended content and user satisfaction. S4. Using metric learning methods, by calculating the similarity between users and products, and based on the embedded representation of user historical behavior and product features, the Siamese network is used for optimization, thereby achieving accurate product recommendations. S5. Analyze the emotional information in users' comments, likes and shares on social platforms, and combine the emotion-behavior mapping model to integrate users' emotional states into the e-commerce intelligent recommendation system.

[0009] Optionally, the user data collection module collects user behavior data from social media and e-commerce platforms, specifically including browsing and click records, search and query keywords, adding to cart and wishlist actions, and social interaction information such as user following and friend relationships. Preprocessing includes deduplication, standardization, and missing value imputation. First, deduplication eliminates duplicate behavior records, ensuring that each user behavior data is counted only once. Standardization aims to transform different types of data to the same scale for effective comparison in subsequent recommendation algorithms. The standardization process uses the following formula: ; in, Represents the original data. and These represent the mean and standard deviation of the data, respectively. This is the standardized data.

[0010] A certain behavioral data For missing values, use the mean to fill in: ; in, This is the average of behavioral data from other users similar to the current user. In some cases, it can be filled using a weighted average based on similarity. ; in, This indicates the number of other users similar to the current user. User The weight, User Behavioral data.

[0011] Through the above data preprocessing steps, the cleaned data will have higher quality, providing reliable input for subsequent modules such as multimodal learning, reinforcement learning, and metric learning. This ensures that the e-commerce intelligent recommendation system can make personalized recommendations based on accurate user behavior profiles, thereby improving the relevance of the recommendation results and user satisfaction.

[0012] Optionally, step S2 includes the following specific steps: S21. Extract text data, image data, and video data from the cleaned user behavior data. After segmenting and removing stop words from the text data, use a word embedding model to map each word into a vector. , of which The paragraph text contains If there are 100 words, then the feature vector of the text segment is... Represented as: ; S22, Regarding image data Input a pre-trained convolutional neural network to extract visual features. Let the CNN feature mapping function be... Then the image feature vector for: ; S23, Regarding video data A 3D convolutional network is used to extract spatiotemporal features. Let the 3D-CNN mapping function be... Then the video feature vector for: ; S24. Transfer the text feature vector Image feature vectors With video feature vectors Parallel input by A multimodal fusion network composed of multi-layer cross-modal Transformers, the network in the first layer... The layer generates a query matrix through linear mapping. Key matrix Value matrix And calculate the first Head attention: ; Subsequently, through multi-head splicing And through residual connections and layer normalization iterations to the th... Layer, let the fusion mapping function be Generate a unified user profile vector ; S25. Unify user profile vectors A multi-dimensional user information model is constructed, including information such as user interests, purchasing power, and activity level. This profile vector is then input into the subsequent recommendation module for personalized and emotion-driven recommendations, thereby improving recommendation accuracy and user experience.

[0013] Optionally, step S3 includes the following specific steps: S31. Initialize the master network parameters of the deep Q-network as follows: The target network parameters are Let the discount factor be... The learning rate is The experience replay pool capacity is S32. Before each recommendation is executed, the current user profile vector is... As an environmental state, the set of products to be recommended is also obtained from the product management module. Each item is represented by a feature vector. This means that the feature vector is composed of product category label embeddings, text description word embeddings, main image CNN features, and historical interaction statistical features, and an ε-greedy strategy is used to select actions from this vector set. ; S33, Execution Action Rewards for receiving user feedback And update the user profile vector through the multimodal data processing module. As the next state ; S34, Transform the quadruple Stored in the experience replay pool; when the number of samples in the pool exceeds The earliest sample is deleted according to the first-in, first-out (FIFO) principle. S35, Randomly sample batch conversion from the experience playback pool Calculate the target value Compared with the estimated value, and with loss For the main network parameters Perform gradient descent updates: ; ; ; S36, every fixed number of steps Synchronize the target network parameters with the master network parameters, i.e. ; S37. Repeat steps S32 to S36 until the convergence condition is met or the maximum number of training rounds is reached. This allows for continuous optimization of recommendation strategies in e-commerce intelligent recommendation systems based on user behavior profiles, enabling adaptive updates to real-time changes in user preferences and improving the personalization and satisfaction of recommendations.

[0014] Optionally, step S4 includes the following specific steps: S41. Embedding Space Learning: Learning from Multimodal Fusion of User Profile Vectors The feature vector set of all products to be recommended A Siamese network with parallel input and shared weights, where the mapping function is... Generate user embeddings respectively Product embedding ; S42. Similarity Measurement: Based on User Embedding Product embedding Calculate the cosine similarity and sort all items: ; S43. Similarity Optimization and Adjustment: Constructing Triples ,in For products that users have already interacted with, For non-interactive products, define a distance metric:

[0015] With fixed margin Triple loss Network weights Perform gradient descent optimization: ; After training, based on similarity Output from high to low Each item is used as a personalized recommendation list.

[0016] Optionally, step S5 includes the following specific steps: S51. Extract user comment text sets, likes, and shares from social platform behavioral data.

[0017] S52. Apply the sentiment analysis model sequentially to the collection of comment texts to obtain the sentiment polarity score of each comment and calculate the average sentiment score of all comments. The sentiment analysis model is a pre-trained BERT text classification model based on Transformer. First, each comment text is segmented and embedded. The BERT model generates an embedding vector for each word or subword based on the parameters learned during training. After obtaining the word embeddings of the comment, BERT merges these embedding vectors into a global vector representation through a self-attention mechanism. Then, the output of BERT is passed to a classification layer, which classifies the sentiment polarity of the comment as "positive," "negative," or "neutral" and outputs the corresponding sentiment score.

[0018] Specifically, the emotional polarity score The prediction result corresponding to each comment is usually represented by a probability distribution as follows: ; in, and These are the weights and biases learned during training. Comments obtained through BERT The vector representation of , This indicates the probability that the comment belongs to a certain sentiment category (such as positive, negative, or neutral). Then, by calculating the average sentiment polarity score of all comments, the overall sentiment tendency of the user or a group of users is obtained. ; in, It's the number of comments. It is the first The sentiment polarity score of each comment. It is the average sentiment score of all comments.

[0019] S53. The number of likes and shares is normalized based on the total interaction volume, and combined with the average sentiment score, a comprehensive sentiment state of the user is generated according to preset weights. In this process, the number of likes... and sharing number The total interaction volume is normalized and calculated as follows: ; ; in, and These represent the number of likes and shares received by the user, respectively. and This is the normalized value. Then, it is combined with the average sentiment score of the comments. According to the preset weighting coefficients , , Calculate the user's overall emotional state : ; in, , , These are preset weighting coefficients. This comprehensive emotional state It reflects users' emotional attitudes toward products and provides an emotionally driven basis for subsequent recommendation systems.

[0020] S54. Assess the user's overall emotional state. The product similarity scores calculated by the metric learning module are weighted and combined using a fusion coefficient to generate a sentiment-driven product recommendation score. This score considers the overall sentiment state. Similarity score with the product Combined, through a preset fusion coefficient Weighted merging is then performed. Product similarity. Typically calculated via a metric learning module, it is derived from embedded representations of user behavior and product features. The final sentiment-driven product recommendation score... : ; in, It is based on metric learning to calculate the similarity between users and products. It is the user's emotional state. It is the fusion coefficient, which is usually adjusted during training to ensure a balance between emotion-driven recommendations and product similarity.

[0021] S55, Recommendation score based on emotion-driven factors Sort all products to be recommended and output the top-ranked products as personalized recommendations. The recommendation scores for all products are obtained through the sorting operation. The system identifies 100 products and returns them as the final recommendation to the user. The sorting operation can use common sorting algorithms, such as quicksort or mergesort, to ensure that the recommendation is based on the optimal product weighted by sentiment and similarity.

[0022] S56. After users subsequently comment, like, and share, the system continuously collects new user feedback and adjusts the sentiment analysis weights and recommendation fusion coefficients online to continuously optimize recommendation performance. Whenever a user's new social behavior is recorded by the system, the sentiment analysis model and recommendation fusion coefficients are fine-tuned based on the new data to ensure the recommendation system can adapt to changes in user sentiment in real time. This process of fine-tuning the sentiment analysis model and updating the fusion coefficients is a dynamic optimization process aimed at improving the personalization and accuracy of recommendations, thereby enhancing user satisfaction and the platform's commercial benefits.

[0023] The beneficial effects of this invention are: This invention proposes an innovative intelligent recommendation system for e-commerce by combining user behavior profiling, social behavior data, and multimodal learning technology. First, this invention utilizes a multimodal data processing module to comprehensively analyze and integrate user behavior data from e-commerce and social platforms, including various data formats such as text, images, and videos. This fully captures users' behavioral characteristics, interests, and social interactions, avoiding the limitations of traditional recommendation systems that rely on only a single data source. Through multimodal learning, this invention can build more complete and detailed user profiles, thereby providing more accurate data support for personalized recommendations.

[0024] This invention utilizes a reinforcement learning recommendation module to continuously optimize the recommendation strategy, enabling the recommendation system to dynamically adjust based on real-time user feedback. Compared to traditional static recommendation methods, this invention's system can adaptively respond to changes in user behavior and interests, providing more personalized and timely product recommendations. The introduction of a deep Q-network reinforcement learning algorithm effectively improves recommendation accuracy and optimizes the user's interaction with the recommendation system.

[0025] This invention introduces a Siamese network through a metric learning module, further optimizing the accuracy of product recommendations. By calculating the similarity between users and products, and based on the embedded representations of user historical behavior and product features, the system can effectively identify the most relevant products and make accurate recommendations. Compared with traditional collaborative filtering algorithms, the metric learning method can better handle the problem of data sparsity, especially in the cases of new products and cold start problems, exhibiting better adaptability and accuracy.

[0026] This invention utilizes a user sentiment analysis module to deeply analyze users' emotional expressions on social media platforms and integrates these sentiment-behavior mapping models into the recommendation system. Through sentiment analysis, the system can not only identify users' interests but also capture their emotional needs, thereby achieving emotion-driven personalized recommendations. Sentiment analysis provides traditional recommendation systems with more dimensions of support, ensuring that recommendations not only align with users' behavioral habits but also better meet their emotional needs, thus improving user satisfaction and the platform's commercial conversion rate.

[0027] In summary, this invention, through the combination of multimodal learning, reinforcement learning, metric learning, and sentiment analysis, not only breaks through the bottlenecks of traditional recommendation systems in data processing and model training, but also effectively improves the accuracy, personalization, and emotional resonance of recommendations. It provides users with more accurate product recommendations that better meet their emotional needs, greatly enhancing user experience and the commercial value of the platform. Attached Figure Description

[0028] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for an intelligent e-commerce recommendation system based on user behavior profiles proposed in this invention. Figure 2 This is a system flowchart of an e-commerce intelligent recommendation system based on user behavior profiles proposed in this invention. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0030] refer to Figure 1-2 A system for intelligent recommendation in e-commerce based on user behavior profiles, comprising: Multimodal data processing module: Used to process user behavior data from different data sources and integrate the user behavior data into a unified user profile; Reinforcement learning recommendation module: used to continuously optimize recommendation strategies based on user interactions with the recommendation system; The metric learning module is used to learn the similarity between users and products. By measuring the similarity between user preferences and product features, it recommends the most relevant products and content to users. User sentiment analysis module: This module analyzes users' emotional expressions on social platforms and combines these with users' emotional states to provide personalized recommendations that better meet users' emotional needs.

[0031] In this embodiment, the modules are interconnected using the following method: S1. Collect and clean user behavior data from different data sources; S2. Use multimodal learning methods to process user behavior data, which includes text, images, and videos. Convert text, images, and videos into a unified user profile to form a multi-dimensional user information model. S3. Based on the interaction feedback between users and the e-commerce intelligent recommendation system, continuously optimize the recommendation strategy. Use the deep Q network reinforcement learning algorithm to update the recommendation strategy after each user feedback to improve the personalization of recommended content and user satisfaction. S4. Using metric learning methods, by calculating the similarity between users and products, and based on the embedded representation of user historical behavior and product features, the Siamese network is used for optimization, thereby achieving accurate product recommendations. S5. Analyze the emotional information in users' comments, likes and shares on social platforms, and combine the emotion-behavior mapping model to integrate users' emotional states into the e-commerce intelligent recommendation system.

[0032] In this embodiment, the user data collection module collects user behavior data from social media and e-commerce platforms, specifically including browsing and click records, search and query keywords, adding to cart and wishlist operations, and social interaction information such as user following and friend relationships. Preprocessing includes deduplication, standardization, and missing value imputation. First, deduplication eliminates duplicate behavior records, ensuring that each user behavior data is counted only once. Standardization aims to transform different types of data to the same scale for effective comparison in subsequent recommendation algorithms. The standardization process uses the following formula: ; in, Represents the original data. and These represent the mean and standard deviation of the data, respectively. This is the standardized data.

[0033] A certain behavioral data For missing values, use the mean to fill in: ; in, This is the average of behavioral data from other users similar to the current user. In some cases, it can be filled using a weighted average based on similarity. ; in, This indicates the number of other users similar to the current user. User The weight, User Behavioral data.

[0034] Through the above data preprocessing steps, the cleaned data will have higher quality, providing reliable input for subsequent modules such as multimodal learning, reinforcement learning, and metric learning. This ensures that the e-commerce intelligent recommendation system can make personalized recommendations based on accurate user behavior profiles, thereby improving the relevance of the recommendation results and user satisfaction.

[0035] In this embodiment, step S2 includes the following specific steps: S21. Extract text data, image data, and video data from the cleaned user behavior data. After segmenting and removing stop words from the text data, use a word embedding model to map each word into a vector. , of which The paragraph text contains If there are 100 words, then the feature vector of the text segment is... Represented as: ; S22, Regarding image data Input a pre-trained convolutional neural network to extract visual features. Let the CNN feature mapping function be... Then the image feature vector for: ; S23, Regarding video data A 3D convolutional network is used to extract spatiotemporal features. Let the 3D-CNN mapping function be... Then the video feature vector for: ; S24. Transfer the text feature vector Image feature vectors With video feature vectors Parallel input by A multimodal fusion network composed of multi-layer cross-modal Transformers, the network in the first layer... The layer generates a query matrix through linear mapping. Key matrix Value matrix And calculate the first Head attention: ; Subsequently, through multi-head splicing And through residual connections and layer normalization iterations to the th... Layer, let the fusion mapping function be Generate a unified user profile vector ; S25. Unify user profile vectors A multi-dimensional user information model is constructed, including information such as user interests, purchasing power, and activity level. This profile vector is then input into the subsequent recommendation module for personalized and emotion-driven recommendations, thereby improving recommendation accuracy and user experience.

[0036] In this implementation, the cleaned user behavior data is first processed by extracting text, image, and video data. The text data undergoes word segmentation and stop word removal, and a word embedding model is used to map each word into a vector, generating a text feature vector. Image data is input into a pre-trained convolutional neural network to extract visual features, generating an image feature vector. Video data is processed using a 3D convolutional network to extract spatiotemporal features, generating a video feature vector. Then, the text, image, and video feature vectors are input in parallel into a multimodal fusion network composed of multiple cross-modal Transformers. A query matrix, key matrix, and value matrix are generated through linear mapping, attention weights are calculated, and a unified user profile vector is finally generated through multi-head concatenation, residual connections, and layer normalization. This user profile vector, combined with user interests, purchasing power, and activity levels, constructs a multi-dimensional user information model. This model is then input into the subsequent recommendation module as the basis for personalized and emotion-driven recommendations, improving recommendation accuracy and user experience.

[0037] In this embodiment, step S3 includes the following specific steps: S31. Initialize the master network parameters of the deep Q-network as follows: The target network parameters are Let the discount factor be... The learning rate is The experience replay pool capacity is S32. Before each recommendation is executed, the current user profile vector is... As an environmental state, the set of products to be recommended is also obtained from the product management module. Each item is represented by a feature vector. This means that the feature vector is composed of product category label embeddings, text description word embeddings, main image CNN features, and historical interaction statistical features, and an ε-greedy strategy is used to select actions from this vector set. ; S33, Execution Action Rewards for receiving user feedback And update the user profile vector through the multimodal data processing module. As the next state ; S34, Transform the quadruple Stored in the experience replay pool; when the number of samples in the pool exceeds The earliest sample is deleted according to the first-in, first-out (FIFO) principle. S35, Randomly sample batch conversion from the experience playback pool Calculate the target value Compared with the estimated value, and with loss For the main network parameters Perform gradient descent updates: ; ; ; S36, every fixed number of steps Synchronize the target network parameters with the master network parameters, i.e. ; S37. Repeat steps S32 to S36 until the convergence condition is met or the maximum number of training rounds is reached. This allows for continuous optimization of recommendation strategies in e-commerce intelligent recommendation systems based on user behavior profiles, enabling adaptive updates to real-time changes in user preferences and improving the personalization and satisfaction of recommendations.

[0038] In this implementation, the main network parameters and target network parameters of the deep Q-network are first initialized, and the discount factor, learning rate, and experience replay pool capacity are set. Before each recommendation execution, the current user profile vector is used as the environment state, and a set of products to be recommended is obtained from the product management module. The product feature vector is composed of product category label embedding, text description word embedding, main image CNN features, and historical interaction statistical features, and an ε-greedy strategy is used to select the action. After the action is executed, the system receives the user's feedback reward and updates the user profile vector through the multimodal data processing module as the next state. Each transformation quadruple is stored in the experience replay pool. If the number of samples in the pool exceeds the set capacity, the earliest sample is deleted in a first-in-first-out manner. Then, batch transformations are randomly sampled from the experience replay pool, the target value and the estimated value are calculated, and the loss is calculated based on this. The main network parameters are updated using the gradient descent method. Every fixed number of steps, the target network parameters are synchronized with the main network parameters. Finally, by continuously repeating the above steps until the convergence condition is met or the maximum number of training rounds is reached, the recommendation strategy is adaptively updated, thereby improving the personalization and satisfaction of the recommendation.

[0039] In this embodiment, step S4 includes the following specific steps: S41. Embedding Space Learning: Learning from Multimodal Fusion of User Profile Vectors The feature vector set of all products to be recommended A Siamese network with parallel input and shared weights, where the mapping function is... Generate user embeddings respectively Product embedding ; S42. Similarity Measurement: Based on User Embedding Product embedding Calculate the cosine similarity and sort all items: ; S43. Similarity Optimization and Adjustment: Constructing Triples ,in For products that users have already interacted with, For non-interactive products, define a distance metric:

[0040] With fixed margin Triple loss Network weights Perform gradient descent optimization: ; After training, based on similarity Output from high to low Each item is used as a personalized recommendation list.

[0041] In this implementation, firstly, through embedding space learning, the user profile vector fused from multimodal analysis and the feature vector sets of all products to be recommended are input in parallel into a Siamese network with shared weights to generate user embeddings and product embeddings. Then, based on the similarity between user embeddings and product embeddings, cosine similarity is used to calculate the similarity between products and users, and all products are ranked. Next, a user-positive-negative triplet is constructed, where positive samples are products the user has interacted with and negative samples are products the user has not interacted with. A distance metric and fixed margin are set, and gradient descent optimization of the network weights is performed by minimizing the triplet loss function. After training, based on the optimized similarity, the top few products are output from high to low, generating a personalized recommendation list.

[0042] In this embodiment, step S5 includes the following specific steps: S51. Extract user comment text sets, likes, and shares from social platform behavioral data.

[0043] S52. Apply the sentiment analysis model sequentially to the collection of comment texts to obtain the sentiment polarity score of each comment and calculate the average sentiment score of all comments. The sentiment analysis model is a pre-trained BERT text classification model based on Transformer. First, each comment text is segmented and embedded. The BERT model generates an embedding vector for each word or subword based on the parameters learned during training. After obtaining the word embeddings of the comment, BERT merges these embedding vectors into a global vector representation through a self-attention mechanism. Then, the output of BERT is passed to a classification layer, which classifies the sentiment polarity of the comment as "positive," "negative," or "neutral" and outputs the corresponding sentiment score.

[0044] Specifically, the emotional polarity score The prediction result corresponding to each comment is usually represented by a probability distribution as follows: ; in, and These are the weights and biases learned during training. Comments obtained through BERT The vector representation of , This indicates the probability that the comment belongs to a certain sentiment category (such as positive, negative, or neutral). Then, by calculating the average sentiment polarity score of all comments, the overall sentiment tendency of the user or a group of users is obtained. ; in, It's the number of comments. It is the first The sentiment polarity score of each comment. It is the average sentiment score of all comments.

[0045] S53. The number of likes and shares is normalized based on the total interaction volume, and combined with the average sentiment score, a comprehensive sentiment state of the user is generated according to preset weights. In this process, the number of likes... and sharing number The total interaction volume is normalized and calculated as follows: ; ; in, and These represent the number of likes and shares received by the user, respectively. and This is the normalized value. Then, it is combined with the average sentiment score of the comments. According to the preset weighting coefficients , , Calculate the user's overall emotional state : ; in, , , These are preset weighting coefficients. This comprehensive emotional state It reflects users' emotional attitudes toward products and provides an emotionally driven basis for subsequent recommendation systems.

[0046] S54. Assess the user's overall emotional state. The product similarity scores calculated by the metric learning module are weighted and combined using a fusion coefficient to generate a sentiment-driven product recommendation score. This score considers the overall sentiment state. Similarity score with the product Combined, through a preset fusion coefficient Weighted merging is then performed. Product similarity. Typically calculated via a metric learning module, it is derived from embedded representations of user behavior and product features. The final sentiment-driven product recommendation score... : ; in, It is based on metric learning to calculate the similarity between users and products. It is the user's emotional state. It is the fusion coefficient, which is usually adjusted during training to ensure a balance between emotion-driven recommendations and product similarity.

[0047] S55, Recommendation score based on emotion-driven factors Sort all products to be recommended and output the top-ranked products as personalized recommendations. The recommendation scores for all products are obtained through the sorting operation. The system identifies 100 products and returns them as the final recommendation to the user. The sorting operation can use common sorting algorithms, such as quicksort or mergesort, to ensure that the recommendation is based on the optimal product weighted by sentiment and similarity.

[0048] S56. After users subsequently comment, like, and share, the system continuously collects new user feedback and adjusts the sentiment analysis weights and recommendation fusion coefficients online to continuously optimize recommendation performance. Whenever a user's new social behavior is recorded by the system, the sentiment analysis model and recommendation fusion coefficients are fine-tuned based on the new data to ensure the recommendation system can adapt to changes in user sentiment in real time. This process of fine-tuning the sentiment analysis model and updating the fusion coefficients is a dynamic optimization process aimed at improving the personalization and accuracy of recommendations, thereby enhancing user satisfaction and the platform's commercial benefits.

[0049] In this implementation, the system first extracts user comment text sets, likes, and shares from social media platform behavioral data. For each comment text, sentiment analysis is performed using a pre-trained BERT model based on Transformer to obtain a sentiment polarity score, and the average sentiment score for all comments is calculated. Then, likes and shares are normalized based on total interaction volume and combined with the average sentiment score to generate a user's overall sentiment state according to preset weights. Next, the user's overall sentiment state is weighted and merged with the product similarity score calculated by the metric learning module using a fusion coefficient to generate a sentiment-driven product recommendation score. All products to be recommended are ranked according to the sentiment-driven recommendation score, and the top few products are output as personalized recommendations. After subsequent user comments, likes, and shares, the system continuously collects new user feedback and adjusts the sentiment analysis weights and recommendation fusion coefficients online to ensure the recommendation system can adapt to changes in user sentiment in real time, ultimately improving the personalization, accuracy, and user satisfaction of recommendations, thereby achieving better platform business benefits.

[0050] Example 1: To verify the feasibility of this invention in practice, this embodiment applies it to the recommendation system of a well-known e-commerce platform. The platform aims to improve the accuracy of personalized recommendations and user satisfaction, particularly in its ability to combine user social behavior and sentiment analysis, by improving its existing recommendation system.

[0051] Traditional recommendation systems primarily rely on users' purchase history and browsing records to generate recommendations, employing collaborative filtering and content-based recommendation algorithms. While this approach can provide personalized recommendations to some extent, it neglects users' social behavior and emotional expression, resulting in a lack of emotion-driven approaches and deep social network analysis. This is particularly problematic for recommending new users and products during the initial launch phase.

[0052] The platform acquires and cleans user behavior data from social media platforms (such as user comments, likes, and shares) and e-commerce platforms (such as browsing and clicks, search keywords, and shopping cart actions) through its data collection system. After preprocessing (such as deduplication, standardization, and missing value imputation), this data is used to construct a multi-dimensional user behavior profile. This data includes, but is not limited to: user comment content, number of likes, number of shares, product browsing and click history, search terms, and shopping cart additions. The platform performs multimodal fusion of all social behavior data (such as comment text, number of likes, and number of shares) with user shopping behavior data. Specifically, comment text is fed into a pre-trained BERT model for sentiment analysis, which yields a sentiment polarity score (positive, negative, or neutral) for each comment. Simultaneously, the number of likes and shares is normalized and combined with the sentiment analysis score to generate a comprehensive user sentiment state based on preset weights.

[0053] The platform calculates the similarity between users and products through a metric learning module, and combines user social behavior and product characteristics to generate accurate product recommendation scores. The recommendation system uses a deep Q-network algorithm based on reinforcement learning to update the recommendation strategy after each user feedback, and dynamically adjusts the recommended content by combining emotion-driven recommendation scores to improve personalization and user satisfaction.

[0054] During implementation, the platform collected user behavior data over a week and combined user emotional states with product similarity for recommendations. The recommendations not only considered users' historical behavior but also incorporated emotional drivers, making the recommendations more aligned with users' personalized needs and emotional preferences.

[0055] By applying this invention to the platform's recommendation system, the platform can optimize its recommendation strategy in real time to adapt to changes in user behavior and interactions within social circles, thereby enhancing recommendation accuracy and user engagement. To verify the effectiveness of this invention, the platform conducted A / B testing for one week before and after applying the recommendation system of this invention. The test results are shown in Table 1, a comparison table of A / B test results: Table 1 Comparison of AB Test Results

[0056] In the experiment, the experimental group used an intelligent recommendation system based on user behavior profiles and emotion-driven approaches, while the control group used a traditional recommendation system based on historical behavior and product content. A / B testing data showed that the experimental group achieved significant improvements in recommendation click-through rate, user engagement, purchase conversion rate, and user satisfaction, validating the effectiveness of this invention in improving the accuracy of personalized recommendations and user experience.

[0057] In the implementation of recommendation systems, by combining sentiment analysis and social circle interaction, the system not only relies on users' historical behavior but can also adjust recommended content based on users' emotional expressions on social platforms. For example, when a user expresses strong feelings about a certain type of product in a review, the system can dynamically adjust its recommendation strategy, prioritizing products that the user may be interested in, thereby improving the relevance and personalization of the recommendations.

[0058] Furthermore, by continuously collecting and updating user behavior data in real time, the platform can continuously optimize its recommendation algorithm, ensuring that recommended content evolves in response to changes in user behavior. Ultimately, the intelligent recommendation system of this invention improves user engagement and the platform's business benefits, while enhancing user satisfaction and loyalty.

[0059] Through the above embodiments and data verification, this invention demonstrates its enormous application potential in social e-commerce platforms and proves that the system can significantly improve the accuracy, personalization, and user experience of product recommendations, thereby promoting technological progress in personalized and emotion-driven recommendations on e-commerce platforms.

[0060] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent recommendation system for e-commerce based on user behavior profiles, characterized in that, include: Multimodal data processing module: Used to process user behavior data from different data sources and integrate the user behavior data into a unified user profile; Reinforcement learning recommendation module: used for user interaction with the recommendation system, optimizing recommendation strategies through deep Q-networks; The metric learning module is used to learn the similarity between users and products through the Siamese network, measure the similarity between user preferences and product features, and recommend the most relevant products and content to users. User sentiment analysis module: This module analyzes users' emotional expressions on social platforms and combines these with users' emotional states to provide personalized recommendations that better meet users' emotional needs.

2. The e-commerce intelligent recommendation system based on user behavior profiles according to claim 1, characterized in that, The modules are connected in the following way: S1. Collect and clean user behavior data from different data sources; S2. Use multimodal learning methods to process user behavior data, which includes text, images, and videos, and transform the user behavior data into a unified user profile. S3. Based on the user's interaction feedback with the e-commerce intelligent recommendation system, use the deep Q network reinforcement learning algorithm to update the recommendation strategy after each user feedback. S4. Using metric learning methods, the similarity between users and items is calculated, and the embedded representations of users and items are learned and optimized through the Siamese network. S5. Analyze the emotional information in users' comments, likes and shares on social platforms, and combine the emotion-behavior mapping model to integrate users' emotional states into the e-commerce intelligent recommendation system.

3. The e-commerce intelligent recommendation system based on user behavior profiles according to claim 2, characterized in that, The data source collects user behavior data through social and e-commerce platforms, including browsing and click records, search and query keywords, adding to shopping cart and wishlist operations, and social interaction information on user following and friend relationships. The user behavior data is preprocessed, including deduplication, standardization, and missing value imputation.

4. The e-commerce intelligent recommendation system based on user behavior profiles according to claim 3, characterized in that, S2 includes the following specific steps: S21. Extract text data, image data, and video data from the cleaned user behavior data. After segmenting and removing stop words from the text data, use a word embedding model to map each word into a vector. , of which The paragraph text contains One word, Represented as text feature vectors; S22, Regarding image data Input a pre-trained convolutional neural network to extract visual features. Let the CNN feature mapping function be... , Image feature vector: S23, Regarding video data A 3D convolutional network is used to extract spatiotemporal features. Let the 3D-CNN mapping function be... , For video feature vectors; S24. Transfer the text feature vector Image feature vectors With video feature vectors Parallel input by A multimodal fusion network composed of layers of cross-modal Transformers, the network in the first layer... The layer generates a query matrix through linear mapping. Key matrix Value matrix And calculate the first Head attention; S25. Through multi-head splicing, and iteratively through residual connection and layer normalization to the [number]th [level]... Layer, let the fusion mapping function be Generate a unified user profile vector ; S26. Unify user profile vectors A multi-dimensional user information model is constructed, including information such as user interests, purchasing power, and activity level. This profile vector is then input into the subsequent recommendation module for personalized and emotion-driven recommendations, thereby improving recommendation accuracy and user experience.

5. The e-commerce intelligent recommendation system based on user behavior profiles according to claim 4, characterized in that, S3 includes the following specific steps: S31. Initialize the main network parameters and target network parameters of the deep Q network, and configure the discount factor, learning rate and experience replay pool capacity; S32. Before each recommendation is executed, the current user profile vector is... As an environmental state, obtain the set of products to be recommended. Each item is represented by a feature vector. In other words, the feature vector It is composed of product category label embedding, text description word embedding, main image CNN features and historical interaction statistical features, and uses an ε-greedy strategy to select actions in this vector set; S33. After pushing the selected products to the user, the e-commerce intelligent recommendation system receives the user's click, browsing, and purchase feedback, and updates the user profile vector based on the user's click, browsing, and purchase feedback. , as the next state; S34. Store the transition record consisting of the current state, action, reward and next state into the experience replay pool. If the capacity is exceeded, the oldest record will be discarded. S35. Randomly select a small batch of transformation samples from the experience replay pool, calculate the maximum estimate of the next state based on the target network, and at the same time estimate the action pair of the current state based on the main network. Calculate the error between the two and use the error between the two as the loss to update the parameters of the main network. S36. Copy the main network parameters to the target network every preset number of steps; S37. Repeat the training of the deep Q network until the training process of the deep Q network meets the convergence condition.

6. The e-commerce intelligent recommendation system based on user behavior profiles according to claim 5, characterized in that, S4 includes the following specific steps: S41. Convert the current user profile vector With feature vectors A parallel input Siamese network with shared weights is used to learn low-dimensional embedding representations of users and items. S42. In the embedding space, cosine similarity is used to calculate the similarity between each product embedding and the user embedding to obtain the product similarity score. And based on similarity, all products are initially sorted. S43. Construct user-positive-negative sample triples, where positive samples are products that the user has interacted with and negative samples are products that the user has not yet interacted with. By minimizing the distance difference between triples and maintaining the preset margin, continuously adjust and optimize the parameters of the Siamese network. S44. After completing network training, based on the optimized embedding and similarity measurement results, sort all products to be recommended from high to low similarity, and select the top few products to generate the final personalized recommendation list.

7. The e-commerce intelligent recommendation system based on user behavior profiles according to claim 6, characterized in that, S5 includes the following specific steps: S51. Extract user comment text sets, likes, and shares from social platform behavioral data; S52. Apply the sentiment analysis model to the collection of comment texts in sequence to obtain the sentiment polarity score of the currently processed comment text, and calculate the average sentiment score of all comments; The sentiment analysis model is a pre-trained BERT text classification model based on Transformer, which has been fine-tuned with a set of comment text data to map comment texts to sentiment polarity scores. S53. Normalize the number of likes and shares based on the total number of interactions, and generate the user's comprehensive emotional state according to the average emotional score and preset weights. S54. Combine the comprehensive emotional state with the product similarity score calculated by the measurement learning module. The products are weighted and merged according to the fusion coefficient to generate an emotion-driven product recommendation score; S55. Sort all products to be recommended according to the emotion-driven recommendation score, and output the top-ranked products as personalized recommendation results. S56. After users subsequently comment, like, and share, new feedback is continuously collected, and the sentiment analysis weight and recommendation fusion coefficient are adjusted online to continuously optimize the recommendation effect.

Images