Time perception user portrait modeling method based on sentiment analysis

Abstract

The invention discloses a time perception user portrait modeling method based on sentiment analysis, belongs to the field of user portraits and the field of recommendation systems, and constructs a new user portrait model and applies the new user portrait model to the recommendation systems for personalized recommendation. The construction process of the new user portrait model comprises the following steps: designing a data collection and preprocessing module comprising a user behavior data collector, a data cleaner and a feature extractor, wherein the data collection and preprocessing module is used for data collection and preprocessing; designing a time-aware user portrait construction module comprising a time sectionalizer, a sub-time-aware user portrait generator and a user portrait integrator, wherein the time-aware user portrait construction module is used for constructing a time-aware user portrait; a reviewer credibility and sentiment analysis module comprising a credibility calculator, a sentiment analyzer and a weighted sentiment evaluation calculator is designed and used for reviewer credibility and sentiment analysis; and the design recommendation generation and dynamic updating module is used for generating a personalized recommendation list and dynamically updating. According to the invention, the recommendation accuracy and the user experience are enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of user portraits and recommendation systems, and specifically relates to a time-aware user portrait modeling method based on sentiment analysis. Background Art

[0002] In the current information age, with the rapid development of Internet technology, personalized recommendation systems have become one of the core technologies of major e-commerce platforms, social media and content distribution platforms. User Profile, as the basis of personalized recommendation systems, collects and analyzes data such as user historical behavior, interest preferences, and demographic information to build a digital portrait of the user, so as to provide users with more accurate and personalized recommended content. Such a model can not only improve user experience, but also significantly improve the business efficiency and user stickiness of the platform. User profile types can be divided into static portraits and dynamic portraits. Static portraits describe the long-term stable characteristics of users, while dynamic portraits are further refined into long-term portraits and short-term portraits, reflecting the long-term interest tendencies and short-term interest changes of users, respectively. In the process of user profile construction, data is obtained through user information collection methods, and continuously improved through user profile update mechanisms (including explicit methods and implicit methods), ultimately achieving dynamic tracking and accurate description of user characteristics. However, with the diversification and complexity of user behavior data, traditional user profile models face many challenges.

[0003] First, the dynamic changes of user interests are not captured enough; most existing user portrait models assume that user interests and preferences are static or change slowly, and fail to fully consider the reality that user interests change dynamically over time. User interests often change due to time, events, and environment, such as seasonal shopping needs or interest fluctuations caused by emergencies. Due to the lack of analysis in the time dimension, these models find it difficult to accurately predict users' future behavior and preferences, resulting in poor performance of recommendation systems in terms of dynamics and personalization.

[0004] Second, the lack of combining sentiment analysis with the credibility of reviews; in existing recommendation systems, although sentiment analysis techniques have been introduced to understand users' attitudes towards products or services, and to infer users' satisfaction or dissatisfaction by analyzing the sentiment tendencies in user reviews, so as to better meet user needs, these techniques have significant shortcomings: they usually ignore the credibility of the reviewers. Due to the uneven quality of user-generated content, including possible false reviews, sentiment analysis often mistakenly regards these unreliable sentiment information as real feedback, causing the recommendation system to be misled when making decisions. For example, a large number of false reviews may raise the evaluation of a product or devalue its reputation, thereby affecting the accuracy of the recommendation results. The lack of analysis of the credibility of the reviewers may cause the system to recommend products that do not meet the actual needs of users, thereby reducing the effectiveness of the recommendation, damaging the user experience and trust in the system. Combining sentiment analysis with the credibility analysis of the reviewers can significantly improve the accuracy and reliability of the recommendation system.

[0005] In short, the existing technology has the following technical problems: the existing recommendation system models usually lack the ability to effectively capture and analyze the changes in user interests over time. This defect makes it difficult for the recommendation system to adapt to changes in user interests in real time, which in turn affects the accuracy and personalization of recommendations. When using sentiment analysis to make recommendations, current recommendation systems usually ignore the credibility of the reviewer, which may cause the system to generate recommendations based on false or unreliable reviews, affecting the reliability of the recommendation results. Summary of the invention

[0006] In order to solve the above problems, the present invention proposes a time-aware user portrait modeling method based on sentiment analysis. By introducing a time-aware user profile (TUP), the changes in user interests in different time periods are dynamically captured and characterized, thereby improving the responsiveness of the recommendation system in scenarios where user interests change rapidly. At the same time, by combining the credibility of the reviewer with the results of sentiment analysis, the ability to capture the real needs of users is improved, the negative impact of false reviews on the recommendation system is reduced, and the accuracy of recommendations and user experience are enhanced.

[0007] The technical solution of the present invention is as follows: A time-aware user portrait modeling method based on sentiment analysis builds a new user portrait model and applies it to the recommendation system for personalized recommendation. The construction process of the new user portrait model includes the following steps: Step 1: Design a data collection and preprocessing module for data collection and preprocessing; the data collection and preprocessing module includes a user behavior data collector, a data cleaner, and a feature extractor; Step 2: Design a time-aware user portrait construction module for constructing a time-aware user portrait; the time-aware user portrait construction module includes a time segmenter, a sub-time-aware user portrait generator, and a user portrait integrator; Step 3: Design a reviewer credibility and sentiment analysis module for reviewer credibility and sentiment analysis; the reviewer credibility and sentiment analysis module includes a credibility calculator, a sentiment analyzer, and a weighted sentiment evaluation calculator; Step 4: Design a recommendation generation and dynamic update module to generate personalized recommendation lists and dynamically update them.

[0008] Furthermore, in step 1, the working process of the data collection and preprocessing module is as follows: Step 1.1, collect user behavior data through a user behavior data collector; the user behavior data collector is a log analysis tool or a database query tool, which extracts the user's browsing, clicking, purchasing and commenting behaviors from the recommendation system log to construct a user behavior data set; the user behavior data includes the behavior type, the timestamp of the behavior, and the associated product or service identifier; the behavior type includes browsing, clicking or purchasing; (1) ; in, Indicates user Behavioral data collection; It is behavior Timestamp of occurrence; It is behavior The associated product or service logo; It is behavior Type of behavior; Total number of representative behaviors; Step 1.2: Use data cleaner to remove duplicate records, fill in missing data, and standardize data formats; Step 1.3: Extract features through the feature extractor. The specific process is as follows: Step 1.3.1, input user comments into the feature extractor, use the fastText tool to perform sentence-level fine-grained sentiment analysis, and add sentiment scores to user comment sentences; Step 1.3.2: Perform unigram and bigram analysis to extract unigram features and bigram features; unigram features consist of a single word; bigram features consist of two words.

[0009] Furthermore, in step 2, the specific working process of the time-aware user portrait construction module is as follows: Step 2.1: In the time segmenter, use the time series segmentation technology to segment the user behavior data into fixed time periods. The data in each time period is used to construct time period features: (2); in, Indicates time period; and They are The start and end time of a time period; Step 2.2, in the sub-time-aware user portrait generator, a sub-time-aware user portrait is generated based on the data of each time period, which is used to represent the user's interests and behaviors in a specific time period; A tree structure is used to represent the sub-time-aware user portrait, and the user's behavioral characteristics in a certain time period are used as tree nodes. By analyzing the user behavior data in each time period, a sub-time-aware user portrait reflecting the user's preferences in that time period is generated. A sub-time-aware user portrait consists of three parts, including the user's time-aware demographic information, user behavior characteristics, and user interest characteristics: (3); in, For users In the Sub-time-aware user portraits generated for each time period; For demographic information, , Represents different demographic information of users, The total number representing the user's demographic information; For user behavior characteristics, , Indicates the user's behavior on different services, Indicates the total number of services involved in user behavior; is the user interest feature, , Represents the weights of different interest relations, Represents user The number of related interest relationships; By calculating the differences in sub-time-aware user portraits in adjacent time periods, we can identify the changing trends of users’ short-term interests: (4); in, Indicates user In the Time period and Time period Changes in interests between For users In the Sub-time-aware user portraits generated for each time period; Identify long-term trends in user interests by analyzing time-aware user profiles over multiple time periods: (5); in, For users long-term trends in interest; Represents the number of historical time steps; is the weighting coefficient, indicating the number of steps in each historical time The degree of contribution to future predictions of time-aware user profiles; Determines the size of the historical data window, representing the length of the user's past behavior sequence; Representative User In the Time-aware user portraits for different time periods; Step 2.3: In the user portrait integrator, integrate the sub-time-aware user portraits in multiple time periods to form a global time-aware user portrait: (6); in, For users Global time-aware user portrait; For the The weight of each time period; Represents the number of time periods.

[0010] Furthermore, in step 3, the working process of the reviewer credibility and sentiment analysis module is as follows: Step 3.1: In the credibility calculator, firstly, the credibility score of the reviewer is quantified based on the number of reviews, deviation and review pass rate; then, the preliminary credibility score is calculated through statistical analysis and weighted average method; finally, the accuracy of the score is further optimized using the Bayesian method: (7); in, For commenters Credibility score of For commenters In the performance on each indicator; For the The weight of each indicator; is the total number of indicators; Step 3.2: In the sentiment analyzer, perform sentiment analysis on the user comment text and extract the sentiment score: (8); in, For comments Sentiment ratings; and Comments The probability of positive and negative emotions; Step 3.3. In the Weighted Sentiment Rating Calculator, generate a weighted sentiment score by multiplying the sentiment score by the reviewer’s credibility score: (9); in, For comments The weighted sentiment score of .

[0011] Furthermore, in step 4, the specific working process of the recommendation generation and dynamic update module is as follows: Step 4.1: Combine the user's time-aware user profile and weighted sentiment score to calculate the recommendation score of each candidate item through linear regression: (10); in, For users For candidate items Recommendation rating of and Different adjustment coefficients are used to control the influence of user portraits and sentiment scores respectively; Sort the recommendation scores of each candidate item in descending order, and select the top 10 recommendation scores to generate a personalized recommendation list; Step 4.2: The formula for dynamic update is: (11); in, Is a user In the Time-aware user portraits after updates for each time period; Is a user In the Time-aware user portraits for different time periods; Is a user In the Time-aware user portraits for different time periods; is the smoothing factor.

[0012] Beneficial technical effects brought about by the present invention: The present invention significantly improves the dynamic response capability of the recommendation system and the accuracy of the recommendation results by introducing a technology that combines time-aware user portraits and commentator credibility with sentiment analysis. By capturing changes in user interests in real time, the system can more accurately reflect user preferences in different time periods. At the same time, by quantifying the credibility of commentators, it reduces the interference of false comments on recommendation results and ensures the reliability of recommended content. In addition, the dynamic recommendation generation and update mechanism of the present invention enables the system to flexibly adapt to changes in user behavior and interests, and continuously optimize recommendation strategies, especially when dealing with new users or sudden changes in user interests. Compared with traditional methods, the present invention not only enhances the personalization capabilities of the recommendation system, but also significantly improves the overall user experience and satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 This is a flow chart of the time-aware user portrait modeling method based on sentiment analysis of the present invention.

[0014] Figure 2 A schematic diagram of a sub-time-aware user portrait of the present invention. DETAILED DESCRIPTION

[0015] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: A time-aware user portrait modeling method based on sentiment analysis constructs a new user portrait model and applies it to the recommendation system for personalized recommendation. The construction process of the new user portrait model includes the following steps: Step 1: Design a data collection and preprocessing module for data collection and preprocessing; the data collection and preprocessing module includes a user behavior data collector, a data cleaner, and a feature extractor. The specific process is as follows: Step 1.1: Build a user behavior data collector to collect user behavior data; Use log analysis tools or database query tools (such as SQL) to extract users' browsing, clicking, purchasing and commenting behaviors from the recommendation system logs and build a user behavior data set. The collected user behavior data includes the behavior type, the timestamp of the behavior, and the associated product or service identifier; the behavior type includes browsing, clicking or purchasing; (1) ; in, Indicates user Behavioral data collection; It is behavior Timestamp of occurrence; It is behavior The associated product or service logo; It is behavior Type of behavior; Represents the total number of behaviors.

[0016] Step 1.2: Build a data cleaner for data cleaning; Data cleaners are used to remove duplicate records, fill in missing data, and standardize data formats. For example, duplicate click behaviors within the same time period are removed, and only the initial behaviors are retained.

[0017] Step 1.3: Build a feature extractor for feature extraction; User reviews contain many words, but not all words are relevant to users' purchasing preferences. Product features refer to functions, capabilities, or visual features that play a role in product selection, such as "image quality", "battery life", "lens", "resolution" in "camera". Extract key features through text analysis to improve system efficiency and robustness. It takes user reviews as input and outputs context-driven features that help user preference modeling, review analysis, and sentiment mining. The feature extraction method consists of two parts: fine-grained sentiment analysis at the sentence level and unigram and bigram analysis. The specific working process of the feature extractor is as follows: Step 1.3.1: Input the user comments into the feature extractor, first perform sentence-level fine-grained sentiment analysis, and add sentiment scores to the user comment sentences; Sentiment analysis is a text classification task that determines the polarity of an opinion (such as positive, negative, or neutral). Aspect-based sentiment analysis automatically assigns sentiment to specific features or topics, while fine-grained sentiment analysis further breaks down sentiment intensity. In the present invention, fastText is used for fine-grained sentiment analysis, which is capable of handling complex sentiments and attaching sentiment scores to comment sentences without sacrificing efficiency. fastText is an efficient text classification tool developed by the Facebook AI research team. It performs fast and accurate text representation by representing words as bags of words and subwords, and is capable of handling complex sentiments and attaching sentiment scores to comment sentences without sacrificing efficiency.

[0018] Step 1.3.2, then perform univariate and bivariate analysis to extract univariate features and bivariate features; When selecting features from the dataset, we mainly focus on the following two types of product features: Unary features: consist of a single word, such as "lens" and "resolution".

[0019] Binary features: consist of two words, such as "image quality" and "red eye".

[0020] The present invention restricts the bigram features to noun and adjective combinations (noun + noun, adjective + noun), and does not include stop words or opinion dictionaries.

[0021] Step 2: Design a time-aware user portrait construction module for constructing a time-aware user portrait; the time-aware user portrait construction module includes a time segmenter, a sub-time-aware user portrait generator, and a user portrait integrator; In order to describe the evolution of user profiles, the present invention proposes a time-aware user profile (TUP). TUP covers multiple user profiles and their evolution in different time periods, and is obtained by learning the Personal Service Ecosystem (PSE) from user behavior logs. Since PSE is too complex to directly represent users, it is simplified into a sub-time-aware User Profile (sTUP) with a tree structure. TUP contains a set of sTUPs, each of which corresponds to a specific time period and can capture the user's changing preferences and interests. Unlike static user profiles, TUP is time-dependent.

[0022] Step 2.1: In the time segmenter, use time series segmentation technology (such as sliding window method) to segment user behavior data into fixed time periods (such as day, week, month), and the data in each time period is used to construct time period features: (2); in, Indicates time period; and They are The start and end time of a time period.

[0023] Step 2.2: In the sub-time-aware user portrait generator, a sub-time-aware user portrait is generated based on the data of each time period to represent the user's interests and behaviors in a specific time period.

[0024] like Figure 2 As shown in the figure, a tree structure is used to represent the sub-time-aware user portrait, and the user's behavior characteristics in a certain time period are used as tree nodes. By analyzing the user behavior data in each time period, a sub-time-aware user portrait reflecting the user's preferences in that time period is generated. A sub-time-aware user portrait mainly consists of three parts, including the user's time-aware demographic information, user behavior characteristics, and user interest characteristics: (3); in, For users In the Sub-time-aware user portraits generated for each time period; For demographic information, , Indicates different demographic information of users, such as age, gender, education, address, etc. The total number representing the user's demographic information; For user behavior characteristics, , Indicates user behaviors on different services, such as clicks, browsing, purchases, and reviews. Indicates the total number of services involved in user behavior; is the user interest feature, , Represents the weights of different interest relations, Represents user The number of related interest relationships.

[0025] By calculating the differences in sub-time-aware user portraits in adjacent time periods, we can identify the changing trends of users’ short-term interests: (4); in, Indicates user In the Time period and Time period Changes in interests between For users In the Sub-time-aware user portraits generated for each time period.

[0026] By analyzing time-aware user profiles over multiple time periods, we can identify long-term trends in user interests, such as cyclical changes or stability: (5); in, For users long-term trends in interest; Represents the number of historical time steps; is the weighting coefficient, indicating the number of steps in each historical time The degree of contribution to future predictions of time-aware user profiles; Determines the size of the historical data window, representing the length of the user's past behavior sequence; Representative User In the Time-aware user portraits for time periods, where is the current time period.

[0027] Step 2.3: In the user portrait integrator, integrate the sub-time-aware user portraits in multiple time periods to generate a complete time-aware user portrait to describe the evolution of user interests over time. Use weighted average or convolutional neural network to integrate the sub-time-aware user portraits in different time periods to form a global time-aware user portrait: (6); in, For users Global time-aware user portrait; For the The weight of each time period is used to adjust the impact of different time periods on the final portrait; Represents the number of time periods. This formula calculates the final time-aware user profile by weighted summing up the sub-time-aware user profiles of multiple time periods.

[0028] Step 3: Design a reviewer credibility and sentiment analysis module for reviewer credibility and sentiment analysis; the reviewer credibility and sentiment analysis module includes a credibility calculator, a sentiment analyzer, and a weighted sentiment evaluation calculator; Step 3.1: In the credibility calculator, first quantify the credibility score of the reviewer based on the number of reviews, deviation, and review pass rate. Then, calculate the preliminary credibility score through statistical analysis and weighted average method, and finally use the Bayesian method to further optimize the accuracy of the score: (7); in, For commenters Credibility score of For commenters In the performance on each indicator; For the The weight of each indicator; is the total number of indicators.

[0029] Step 3.2: In the sentiment analyzer, perform sentiment analysis on the user review text and extract the sentiment score to reflect the reviewer's emotional attitude towards the product, the probability difference between positive and negative emotions, and generate the sentiment score: (8); in, For comments Sentiment ratings; and Comments The probability of positive and negative emotions.

[0030] Step 3.3: In the weighted sentiment rating calculator, combine the reviewer's credibility with the sentiment analysis results to generate a weighted sentiment score to improve the accuracy of sentiment analysis. The weighted sentiment score is generated by multiplying the sentiment score with the reviewer's credibility score: (9); in, For comments The weighted sentiment score of In this way, the sentiment scores of reviewers with higher credibility have greater weight in the recommendation results.

[0031] Step 4: Design a recommendation generation and dynamic update module to generate a personalized recommendation list and dynamically update it. The specific working process of the recommendation generation and dynamic update module is as follows: Step 4.1: In the recommendation generation and dynamic update module, a personalized recommendation list is generated based on the user's time-aware user portrait and weighted sentiment score. The recommendation score of each candidate item is calculated by linear regression based on the user's time-aware user portrait and weighted sentiment score: (10); in, For users For candidate items Recommendation rating of and Different adjustment coefficients are used to control the influence of user portraits and sentiment scores respectively.

[0032] Sort the recommendation score of each candidate item in descending order, and select the top 10 recommendation scores to generate a personalized recommendation list.

[0033] Step 4.2: The formula for dynamic update is: (11); in, Is a user In the Time-aware user portraits after updates for each time period; Is a user In the Time-aware user portraits for different time periods; Is a user In the Time-aware user portraits for a time period (i.e., portraits for the previous time period); is a smoothing factor that controls the weight of the current portrait and the historical portrait. .

[0034] Through this dynamic update mechanism, the user profile will be continuously adjusted and optimized over time to reflect the changes in user interests in real time. It plays a key role in determining the weight of historical information and current information in the update process. Values ​​indicate greater reliance on historical data, while smaller A value of 0 means paying more attention to the changes in interest at the current moment. This design ensures that the model can make real-time adjustments based on the fluctuations in user interests, thereby providing more accurate and personalized recommendations.

[0035] The optimized user portrait model modeled by the present invention can dynamically capture and analyze changes in user interests through the collaborative work of various modules, and generate highly personalized recommendation content by combining the user's emotional tendency and the credibility of the commentator. The model not only performs well in real-time tracking of user interests, but also reduces the impact of false comments on the recommendation system by combining the credibility of the commentator with sentiment analysis, thereby improving the accuracy of recommendations and user experience.

[0036] In order to verify the effectiveness of the time-aware user portrait modeling method based on sentiment analysis proposed in this paper compared with the traditional recommendation method, a comparative experiment was conducted. The MovieLens 100K dataset was selected to evaluate the personalized recommendation task. The experimental goal is to evaluate the performance of the proposed method in key indicators such as recommendation precision, recall, and root mean square error (RMSE).

[0037] This experiment uses the MovieLens 100K dataset, which contains 100,000 user rating records involving 943 users and 1,682 movies. Each record includes information such as user ID, movie ID, rating, and timestamp. The MovieLens 100K dataset is a standard dataset widely used in the evaluation of personalized recommendation algorithms. This experiment divides the dataset into a training set and a test set at a ratio of 80%:20%, with 80% of the data used for training and the remaining 20% ​​for testing. In order to eliminate the bias in user ratings, this experiment normalized the user ratings and converted each user's rating into a standardized rating range (between 0 and 1) to ensure the comparability of ratings between different users.

[0038] In order to verify the effectiveness of the method of the present invention, two traditional methods were selected for comparison: User-based Collaborative Filtering (UB-CF): This algorithm recommends items to similar users based on their historical behaviors, assuming that users with similar interests will like similar items.

[0039] Item-based Collaborative Filtering (IB-CF): This algorithm recommends other items similar to the items rated by the user based on the user's rating history.

[0040] In order to comprehensively evaluate the effects of different recommendation methods, the following commonly used evaluation indicators are used: Precision: The ratio of items that users actually like to all recommended items in the recommendation list. The calculation formula is: (12); in, Indicates the number of items recommended by the recommendation system that users actually like (i.e., correctly recommended items); Represents the number of items recommended by the recommendation system that users actually dislike (i.e., incorrectly recommended items).

[0041] Recall: The ratio of items that users actually like in the recommendation list to all items that users like. The calculation formula is: (13); in, It indicates the number of items that the recommendation system did not recommend, that the user actually likes (i.e., missed items). These items are of interest to the user but are not recommended to the user.

[0042] Root Mean Square Error (RMSE): measures the difference between the recommendation results and the user's actual rating.

[0043] The experimental results are shown in Table 1: Table 1 Comparison results between the method of the present invention and two traditional methods .

[0044] As can be seen from Table 1, the method of the present invention is significantly superior to the user-based collaborative filtering algorithm in terms of precision and recall, indicating that it has obvious advantages in the accuracy and coverage of personalized recommendations. At the same time, the root mean square error of the method of the present invention is significantly lower than that of the traditional method, indicating that its recommendation results are more in line with the actual interests of the user. These results show that by introducing time perception and dynamic update mechanisms, the method of the present invention can more effectively adapt to changes in user interests, thereby providing more accurate and personalized recommendation services. In general, the present invention shows significant advantages in dealing with dynamic changes in user interests and improving the personalization and real-time performance of the recommendation system, proving its feasibility and superiority in practical applications.

[0045] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A time-aware user portrait modeling method based on sentiment analysis, characterized in that: Build a new user portrait model and apply it to the recommendation system for personalized recommendation. The process of building a new user portrait model includes the following steps: Step 1: Design a data collection and preprocessing module for data collection and preprocessing; the data collection and preprocessing module includes a user behavior data collector, a data cleaner, and a feature extractor; Step 2: Design a time-aware user portrait construction module for constructing a time-aware user portrait; the time-aware user portrait construction module includes a time segmenter, a sub-time-aware user portrait generator, and a user portrait integrator; Step 3: Design a reviewer credibility and sentiment analysis module for reviewer credibility and sentiment analysis; the reviewer credibility and sentiment analysis module includes a credibility calculator, a sentiment analyzer, and a weighted sentiment evaluation calculator; Step 4: Design a recommendation generation and dynamic update module to generate personalized recommendation lists and dynamically update them.

2. According to claim 1, the time-aware user portrait modeling method based on sentiment analysis is characterized in that: In step 1, the working process of the data collection and preprocessing module is as follows: Step 1.1, collect user behavior data through a user behavior data collector; the user behavior data collector is a log analysis tool or a database query tool, which extracts the user's browsing, clicking, purchasing and commenting behaviors from the recommendation system log to construct a user behavior data set; the user behavior data includes the behavior type, the timestamp of the behavior, and the associated product or service identifier; the behavior type includes browsing, clicking or purchasing; (1) ； in, Indicates user Behavioral data collection; It is behavior Timestamp of occurrence; It is behavior The associated product or service logo; It is behavior Type of behavior; Total number of representative behaviors; Step 1.2: Use data cleaner to remove duplicate records, fill in missing data, and standardize data formats; Step 1.3: Extract features through the feature extractor. The specific process is as follows: Step 1.3.1, input user comments into the feature extractor, use the fastText tool to perform sentence-level fine-grained sentiment analysis, and add sentiment scores to user comment sentences; Step 1.3.2: Perform unigram and bigram analysis to extract unigram features and bigram features; unigram features consist of a single word; bigram features consist of two words.

3. The time-aware user portrait modeling method based on sentiment analysis according to claim 2 is characterized in that: In step 2, the specific working process of the time-aware user portrait construction module is as follows: Step 2.1: In the time segmenter, use the time series segmentation technology to segment the user behavior data into fixed time periods. The data in each time period is used to construct time period features: (2)； in, Indicates time period; and They are The start and end time of a time period; Step 2.2, in the sub-time-aware user portrait generator, a sub-time-aware user portrait is generated based on the data of each time period, which is used to represent the user's interests and behaviors in a specific time period; A tree structure is used to represent the sub-time-aware user portrait, and the user's behavioral characteristics in a certain time period are used as tree nodes. By analyzing the user behavior data in each time period, a sub-time-aware user portrait reflecting the user's preferences in that time period is generated. A sub-time-aware user portrait consists of three parts, including the user's time-aware demographic information, user behavior characteristics, and user interest characteristics: (3)； in, For users In the Sub-time-aware user portraits generated for each time period; For demographic information, , Represents different demographic information of users, A total number representing user demographic information; For user behavior characteristics, , Indicates the user's behavior on different services, Indicates the total number of services involved in user behavior; is the user interest feature, , Represents the weights of different interest relations, Represents user The number of related interest relationships; By calculating the differences in sub-time-aware user portraits in adjacent time periods, we can identify the changing trends of users’ short-term interests: (4)； in, Indicates user In the Time period and Time period Changes in interests between For users In the Sub-time-aware user portraits generated for each time period; Identify long-term trends in user interests by analyzing time-aware user profiles over multiple time periods: (5)； in, For users long-term trends in interest; Represents the number of historical time steps; is the weighting coefficient, indicating the number of historical time steps The degree of contribution to future predictions of time-aware user profiles; Determines the size of the historical data window, representing the length of the user's past behavior sequence; Representative User In the Time-aware user portraits for different time periods; Step 2.3: In the user portrait integrator, integrate the sub-time-aware user portraits in multiple time periods to form a global time-aware user portrait: (6)； in, For users Global time-aware user portrait; For the The weight of each time period; Represents the number of time periods.

4. The time-aware user portrait modeling method based on sentiment analysis according to claim 3 is characterized in that: In step 3, the working process of the reviewer credibility and sentiment analysis module is as follows: Step 3.1: In the credibility calculator, firstly, the credibility score of the reviewer is quantified based on the number of reviews, deviation and review pass rate; then, the preliminary credibility score is calculated through statistical analysis and weighted average method; finally, the accuracy of the score is further optimized using the Bayesian method: (7)； in, For commenters Credibility score of For commenters In the performance on each indicator; For the The weight of each indicator; is the total number of indicators; Step 3.2: In the sentiment analyzer, perform sentiment analysis on the user comment text and extract the sentiment score: (8)； in, For comments Sentiment ratings; and Comments The probability of positive and negative emotions; Step 3.

3. In the Weighted Sentiment Rating Calculator, generate a weighted sentiment score by multiplying the sentiment score by the reviewer’s credibility score: (9)； in, For comments The weighted sentiment score of .

5. The time-aware user portrait modeling method based on sentiment analysis according to claim 4 is characterized in that: In step 4, the specific working process of the recommendation generation and dynamic update module is as follows: Step 4.1: Combine the user's time-aware user profile and weighted sentiment score to calculate the recommendation score of each candidate item through linear regression: (10)； in, For users For candidate items Recommendation rating of and Different adjustment coefficients are used to control the influence of user portraits and sentiment scores respectively; Sort the recommendation scores of each candidate item in descending order, and select the top 10 recommendation scores to generate a personalized recommendation list; Step 4.2: The formula for dynamic update is: (11)； in, Is a user In the Time-aware user portraits after updates for each time period; Is a user In the Time-aware user portraits for different time periods; Is a user In the Time-aware user portraits for different time periods; is the smoothing factor.

Images