An electronic book intelligent recommendation method and system based on multi-source data

Abstract

The application relates to the technical field of data processing, and discloses an electronic book intelligent recommendation method and system based on multi-source data, which comprises the following steps: collecting user data, score data, time periods, device types and environment labels, and extracting a user feature vector; performing similarity calculation on the behavior features and score change vectors of a chapter to be recommended and the user feature vector to obtain a first similarity; constructing a current context vector based on a user current position and real-time weather information, and comparing the current context vector with an optimal reading context to obtain a context matching degree; comparing current user historical weather information with historical weather information when a feedback user participates in a score behavior to obtain a historical weather matching degree; calculating a second similarity according to the context matching degree and the historical weather matching degree; and sorting candidate chapters based on the product of the first similarity and the second similarity to generate a recommendation list. The application realizes personalized recommendation of electronic book chapters by constructing a user interest graph.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to an intelligent e-book recommendation method and system based on multi-source data. Background Technology

[0002] With the rapid development of mobile reading devices and the digital publishing industry, e-book platforms have gradually replaced traditional paper books, becoming an important channel for users to acquire knowledge and engage in leisure reading. To enhance the user reading experience, many e-book platforms have introduced recommendation systems, using users' historical behavior data to push books of interest, thereby improving content distribution efficiency and platform stickiness.

[0003] Most existing ebook recommendation methods are based on collaborative filtering, content tag matching, or simple user rating clustering, primarily considering users' past click behavior, rating records, or the categories of ebooks they have read. Some advanced methods introduce machine learning models to generate personalized recommendations by calculating the similarity between user and book features. However, in practical applications, the following shortcomings still exist:

[0004] Narrow data dimensions and lack of recommendation context: Current methods generally ignore changes in user reading behavior under different environmental conditions (such as reading time period, reading device, reading location, and weather conditions), resulting in recommendations lacking context awareness and failing to adapt to users' reading preferences in different states; A single feedback mechanism and lack of proactive adjustment: Most recommendation systems rely primarily on explicit user ratings as feedback, failing to effectively utilize fine-grained behavioral characteristics of users during the reading process (such as page dwell time, swipe frequency, chapter reading order, etc.), leading to insufficient system response to users' true preferences and limited recommendation accuracy; Coarse selection paths for similar users, making precise modeling difficult: Traditional collaborative filtering methods often use the overlap of users' overall reading records as the basis for similarity judgment, failing to comprehensively consider the distribution characteristics of users' behavior in specific chapters, rating fluctuation trends, and contextual similarity, thus limiting the accurate construction of similar user profiles.

[0005] Therefore, it is necessary to propose an intelligent e-book recommendation method to solve the problems of insufficient recommendation accuracy, single feedback path, and incomplete user profile construction in existing technologies. Summary of the Invention

[0006] In view of this, the present invention proposes an intelligent e-book recommendation method and system based on multi-source data to solve the above problems.

[0007] On the one hand, the present invention proposes an intelligent e-book recommendation method based on multi-source data, comprising:

[0008] The system collects reading feedback for each chapter of the e-book. After a user finishes reading a chapter, it collects recommended reading scenarios for the current chapter through an interactive window. These recommended scenarios include the best weather and the best reading location for the current chapter.

[0009] When the number of valid reading feedbacks for the current chapter reaches a set threshold, the recommended reading context for the current chapter is statistically analyzed based on the responses of all users who participated in the reading feedback, and the current chapter is labeled in the form of tags, which are denoted as chapter context tags; user behavior data and historical reading data of the feedback users are obtained to construct a set of feedback user profiles for the current chapter;

[0010] Based on current user behavior data and historical reading data, construct a set of user profiles to be recommended, and determine the set of e-book categories suitable for the current user; each set of e-book categories corresponds to a set of e-book chapters containing chapter context tags;

[0011] In the e-book chapter set, the behavioral similarity and historical data similarity between the current user and the feedback user profile set are calculated respectively, and the first similarity is determined based on the behavioral similarity and historical data similarity.

[0012] The current user's real-time location and real-time weather are compared with the best reading location and best reading weather to determine the context matching degree; the current user's historical weather information is obtained and compared with the historical weather when the user participated in the reading feedback to determine the historical weather similarity; and a second similarity degree is determined based on the context matching degree and the historical weather matching degree.

[0013] Based on the product of the first similarity and the second similarity, the candidate chapters are sorted to generate a chapter recommendation list for the current user.

[0014] Furthermore, the generation of chapter context tags includes:

[0015] Once the number of valid reading feedbacks for the current chapter reaches a set feedback threshold, a user feedback state graph is constructed based on all users who provided reading feedback for the current chapter. Each node of the user feedback state graph is a combination of the reading context selected by a user in the current chapter and the corresponding feedback action. The combination of feedback actions is a four-element feedback vector composed of the user's page dwell time, swipe frequency, rating behavior and tag selection behavior recorded in the current chapter.

[0016] The edge weight between any two nodes in the user feedback state graph is the similarity weight between the combination of feedback behaviors of the two users. The similarity weight of the combination of feedback behaviors is obtained by calculating the Euclidean distance of the quaternary feedback vector and performing inverse normalization.

[0017] Furthermore, the generation of chapter context tags also includes:

[0018] Perform graph density-based clustering analysis on the user feedback status graph to identify the subgroup of users with the highest concentration of feedback behavior. Within this subgroup:

[0019] Structural stability is determined by calculating the degree of edge weight fluctuation between each feedback user and the combination of feedback behaviors of its neighboring nodes. If the average edge weight variation coefficient of any feedback user is less than the preset structural stability coefficient threshold, then the feedback user is judged to have structural stability.

[0020] Feedback trajectory continuity is determined by tracking the reading context choices and feedback actions of a user across three consecutive chapters. If the trend of the four-element feedback vector changes monotonically or is within a local fluctuation range between the three chapters, then the user is judged to have feedback trajectory continuity.

[0021] Historical preference consistency is determined by comparing the cosine similarity between the current chapter's feedback behavior combination and the feedback behavior combination of the same user in the same tagged chapters in the past. If the similarity exceeds the set consistency judgment threshold, the feedback user is judged to have historical preference consistency.

[0022] The set of feedback users who simultaneously satisfy structural stability, feedback trajectory continuity, and historical preference consistency is defined as the context label reference group.

[0023] Furthermore, the generation of chapter context tags also includes:

[0024] Extract the context selection tags of feedback users from the context tag reference group, and calculate the context tag distribution density in the weather context dimension and the reading location dimension respectively. Then, calculate the Shannon entropy based on the tag probability distribution as the tag distribution entropy of that dimension.

[0025] When the label distribution entropy of any dimension is lower than the set label aggregation limit, the label with the highest frequency in that dimension is extracted as the chapter context labeling result;

[0026] When the entropy of the label distribution in any dimension is higher than the set label aggregation limit, the reading behavior trajectory data of all feedback users in the context label reference group is called. The reading behavior trajectory data is the context label and feedback action time series corresponding to the user's historical chapter reading records in the reading system.

[0027] Based on the context label of each candidate chapter, the behavioral trajectory associated with the label is identified in the reading behavior trajectory data of the context label reference group, and the proportion of context label annotations in the behavioral trajectory is calculated as the feedback trajectory aggregation degree of the context label of the candidate chapter.

[0028] If the aggregation degree of the feedback trajectory of any candidate chapter context label exceeds the aggregation consistency threshold, then the candidate chapter context label with the highest aggregation degree of feedback trajectory is selected as the chapter context label annotation result; if no label meets the aggregation consistency threshold, then no chapter context label is generated for the current chapter under this dimension.

[0029] The completed chapter context tags are attached to the chapter identifier field of the current chapter as structured fields.

[0030] Furthermore, the construction of the feedback user profile set and the current user profile set includes:

[0031] Acquire target user behavior records and feedback data in the reading system. The behavior records include reading chapter records, rating records, reading time periods and device types used. The feedback data includes page dwell time, swipe frequency and chapter context preference tags.

[0032] Based on the behavioral record data and feedback data, a user profile vector group is generated, the user profile vector group including:

[0033] The behavioral feature vector is generated based on the target user's page dwell time and swipe frequency in each chapter, and is normalized to form a reading behavior distribution vector that reflects the intensity of reading operations.

[0034] The time preference vector is generated based on the statistical results of the target user's historical reading time period. The day is divided into multiple fixed time intervals, the reading frequency in each interval is counted and normalized to form a reading time preference vector that reflects the distribution of its time activity.

[0035] Device adaptation label, which is the device type identifier that appears most frequently in the target user's historical device types, is used to indicate the reading device that the user mainly uses;

[0036] The rating preference vector is generated based on the target user's historical rating records and includes three components: average rating under different tag categories, rating variation, and high rating ratio, which are used to reflect the user's rating trend and preference concentration.

[0037] The context preference vector is generated based on the weather and location tags selected by the user in historical feedback. The frequency of occurrence of each tag is counted and weighted and normalized to obtain a context feature expression that reflects the user's typical reading preference environment.

[0038] The behavioral feature vector, time preference vector, device adaptation tag, rating preference vector, and context preference vector are used to form the corresponding feedback user profile set and the current user profile set.

[0039] Furthermore, when determining the set of ebook categories suitable for the current user, this includes:

[0040] The system calculates the reading time of each category of ebooks in the current user's reading history, and then selects the top N ebook categories by reading time to form a set of ebook categories that are suitable for the current user, where N is a preset upper limit for the number of categories.

[0041] Furthermore, behavioral similarity and historical data similarity are obtained through the following steps:

[0042] Obtain the current user profile vector group and the profile vector group of each feedback user. The profile vector group includes behavioral feature vectors, time preference labels, device adaptation labels, rating preference vectors, and a set of contextual preference labels.

[0043] Calculating behavioral similarity includes:

[0044] Cosine similarity is used to calculate the similarity between behavioral feature vectors;

[0045] If the time preference tags are completely identical, the time matching score is 1; if they are adjacent time periods, the score is calculated according to the preset reduction ratio.

[0046] The behavioral similarity is obtained by weighting and fusing the behavioral feature similarity with the time matching score according to the set weights;

[0047] Calculating historical data similarity includes:

[0048] The similarity of rating preference vectors is calculated using Euclidean distance or KL divergence.

[0049] Extract common tags from the contextual preference tag sets of the current user and the feedback user, and calculate tag similarity scores based on the degree of overlap in tag weight distribution;

[0050] The similarity scores of rating preferences and tags are weighted and fused according to a set weight to obtain the similarity of historical data;

[0051] The weighted product of behavioral similarity and historical data similarity is used as the first similarity.

[0052] Furthermore, the contextual matching degree and historical weather matching degree are obtained through the following steps:

[0053] The current user's real-time location and weather information are compared with the best reading location and best reading weather in the chapter context tags corresponding to the candidate chapters. If the current user's real-time weather information matches the best reading weather in the chapter context tags, the highest matching value is assigned; if there is only a correlation in weather type, a medium matching value is assigned; if there is no correlation, the lowest matching value is assigned. The current user's real-time location information is compared with the best reading location in the chapter context tags. If the current user is in the location type corresponding to the best reading location, the highest matching value is assigned; otherwise, a second-highest matching value is assigned based on the similarity of location types. The context matching degree between the current user and the candidate chapters is calculated based on the above matching values.

[0054] Obtain the historical weather tag sequence corresponding to the current user's historical reading record, and compare it with the historical weather information recorded by the feedback users who participated in the feedback of this chapter. Calculate the historical weather matching degree between the current user and the feedback user by weighted fusion of the co-occurrence frequency of historical weather tags and the overlap of time period distribution. The product of the context matching degree and the historical weather matching degree is used as the second similarity.

[0055] Furthermore, when generating the current user's recommended chapter list, it includes:

[0056] The first similarity and the second similarity of each candidate chapter are multiplied to obtain a comprehensive matching score; all candidate chapters are sorted in descending order based on the comprehensive matching score to construct a chapter recommendation sequence; the top few chapters in the chapter recommendation sequence are selected as the chapter recommendation list for the current user.

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0058] By jointly collecting user reading behavior data, rating data, reading time information, and environmental context information, a multi-dimensional profile construction mechanism and a hierarchical similarity calculation strategy are introduced in the recommendation stage. Specifically, this includes: generating behavioral feature vectors and rating preference vectors based on behavioral trajectories and rating distributions to quantify the similarity between different users in reading habits and rating tendencies; further, constructing a historical feedback information graph by combining the feedback user set of candidate chapters, extracting key feature tags from the perspectives of content preference, emotional preference, and time preference to achieve a high-dimensional characterization of chapter profiles; calculating the first similarity based on the matching degree between the current user and the feedback user under the multi-dimensional profile, and simultaneously introducing contextual matching degree and historical weather matching degree to calculate the second similarity; finally, a two-layer similarity scoring mechanism is used to comprehensively rank candidate chapters and generate recommendation results. This method effectively overcomes the limitations of traditional recommendation methods that only build recommendation lists based on ratings or simple behavioral data, avoiding recommendation inaccuracies caused by single samples, ambiguous behavior, or sparse ratings. It can more accurately identify the current user's true reading interests and maintain stable recommendation performance under complex environmental variables, thereby significantly improving recommendation accuracy, user acceptance, and the personalization of the reading experience.

[0059] On the other hand, the present invention proposes an intelligent e-book recommendation system based on multi-source data, comprising:

[0060] The data collection module is configured to collect reading feedback for each chapter of the e-book. After the user finishes reading a chapter, it collects the recommended reading context for the current chapter through an interactive window. The recommended reading context is the best weather and best reading location for the current chapter.

[0061] The module is configured to, when the number of valid reading feedbacks for the current chapter reaches a set threshold, statistically analyze the recommended reading context for the current chapter based on the responses of all users who participated in the reading feedback, and label the current chapter in the form of tags, which are denoted as chapter context tags; and acquire the user behavior data and historical reading data of the feedback users to construct a set of feedback user profiles for the current chapter.

[0062] The category module is configured to construct a set of user profiles to be recommended based on the current user's user behavior data and historical reading data, and determine the set of e-book categories suitable for the current user; each set of e-book categories corresponds to a set of e-book chapters containing chapter context tags;

[0063] The first calculation module is configured to calculate the behavioral similarity and historical data similarity between the current user and the feedback user profile set in the e-book chapter set, and determine the first similarity based on the behavioral similarity and historical data similarity.

[0064] The second calculation module is configured to compare the current user's real-time location and real-time weather with the best reading location and best reading weather respectively to determine the context matching degree; obtain the current user's historical weather information and compare it with the historical weather when the user participated in reading feedback to determine the historical weather similarity; and determine the second similarity degree based on the context matching degree and the historical weather matching degree.

[0065] The list generation module is configured to sort candidate chapters based on the product of the first similarity and the second similarity, and generate a recommended chapter list for the current user.

[0066] It should be noted that the intelligent e-book recommendation method based on multi-source data of the present invention has the same beneficial effects as its system, and will not be described in detail here. Attached Figure Description

[0067] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0068] Figure 1 The flowchart illustrates an intelligent e-book recommendation method based on multi-source data, as provided in this embodiment of the invention.

[0069] Figure 2 This is a functional block diagram of an intelligent e-book recommendation system based on multi-source data, provided as an embodiment of the present invention. Detailed Implementation

[0070] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0071] See Figure 1 As shown, this embodiment of the invention provides an intelligent e-book recommendation method based on multi-source data, including:

[0072] S1: Collect reading feedback for each chapter of the e-book. After the user finishes reading a chapter, collect the recommended reading context for the current chapter through an interactive window. The recommended reading context is the best weather and best reading location for the current chapter.

[0073] S2: When the number of valid reading feedbacks for the current chapter reaches a set threshold, the recommended reading context for the current chapter is statistically analyzed based on the responses of all users who participated in the reading feedback, and the current chapter is labeled in the form of tags, which are denoted as chapter context tags; user behavior data and historical reading data of the feedback users are obtained to construct a set of feedback user profiles for the current chapter;

[0074] S3: Based on the current user's behavior data and historical reading data, construct a set of user profiles to be recommended, and determine the set of e-book categories suitable for the current user; each set of e-book categories corresponds to a set of e-book chapters containing chapter context tags;

[0075] S4: In the set of chapters in the e-book, calculate the behavioral similarity and historical data similarity between the current user and the set of feedback user profiles, and determine the first similarity based on the behavioral similarity and historical data similarity;

[0076] S5: Compare the current user's real-time location and real-time weather with the best reading location and best reading weather to determine the context matching degree; obtain the current user's historical weather information and compare it with the historical weather when the user participated in the reading feedback to determine the historical weather similarity; determine the second similarity degree based on the context matching degree and the historical weather matching degree.

[0077] S6: Based on the product of the first similarity and the second similarity, sort the candidate chapters and generate a chapter recommendation list for the current user.

[0078] Understandably, by incorporating subjective contextual feedback from users after completing a chapter, the optimal reading location and weather for each chapter across different user groups are collected and statistically analyzed to construct chapter contextual tags. This is then combined with the current user's behavioral data and historical reading records to build a user profile for recommendations. A first similarity score is generated by comparing the user's behavioral similarity to the feedback user profile with historical data similarity, reflecting the degree of overlap in reading preferences among user groups. Simultaneously, the current user's real-time geographical location and weather information are compared with the recommended reading context of the chapter to calculate contextual matching. Furthermore, a second similarity score is formed by combining the current user's historical weather data with the weather during the feedback user's historical reading period, thus comprehensively constructing a multi-source fusion recommendation mechanism driven by user behavior, historical habits, and the external environment. This method not only improves the fit between recommended content and the user's actual context, enhancing the user experience of recommended content, but also significantly improves the accuracy and interpretability of personalized recommendations.

[0079] In some embodiments of the application, the generation of chapter context labels includes:

[0080] Once the number of valid reading feedbacks for the current chapter reaches a set feedback threshold, a user feedback state graph is constructed based on all users who provided reading feedback for the current chapter. Each node of the user feedback state graph is a combination of the reading context selected by a user in the current chapter and the corresponding feedback action. The combination of feedback actions is a four-element feedback vector composed of the user's page dwell time, swipe frequency, rating behavior and tag selection behavior recorded in the current chapter.

[0081] The edge weight between any two nodes in the user feedback state graph is the similarity weight between the combination of feedback behaviors of the two users. The similarity weight of the combination of feedback behaviors is obtained by calculating the Euclidean distance of the quaternary feedback vector and performing inverse normalization.

[0082] In this embodiment, to improve the accuracy and objectivity of chapter context tag generation, a user feedback status graph is introduced to comprehensively evaluate the consistency of reading preferences and the aggregation trend of feedback features among user groups in a specific chapter. The specific implementation steps are as follows:

[0083] First, once the number of valid reading feedbacks for a current ebook chapter reaches a set feedback threshold, the system centrally processes the reading feedback information from all users. Each user must submit their subjectively chosen optimal reading context after completing the chapter, including location and weather. Simultaneously, a set of objective behavioral parameters of the user during the reading process is also recorded. This parameter set includes:

[0084] Page dwell time, which is the total time a user spends on the chapter page; swipe frequency, which is the number of times the user swipes on the page during reading; rating behavior, which is the rating result of the overall reading experience of the chapter; tag selection behavior, which is whether the user actively clicks on or adds auxiliary tags that describe the mood or atmosphere of the chapter.

[0085] Based on the above four types of behavioral data and context selection results, a feedback status node for each user in this chapter is constructed. Each node is a combination of "reading context + four-element feedback vector".

[0086] Next, all the nodes are organized into a graph structure to generate the "User Feedback Status Graph" for this chapter. Each node in the graph represents the complete feedback status of a user, and the edge between any two nodes represents the similarity of the feedback behaviors between the two users. This similarity is determined in the following way:

[0087] First, extract the four-element feedback vectors of the two users (composed of page dwell time, swipe frequency, rating behavior, and tag selection behavior); calculate the Euclidean distance between the two vectors to measure the degree of difference in their feedback behavior; perform inverse normalization on the Euclidean distance (e.g., by using the reciprocal or exponential transformation after normalization to the maximum and minimum values) so that the smaller the distance, the higher the similarity, thereby generating the edge weight of the edge, representing the similarity of the behavioral feedback.

[0088] Finally, by clustering the node group with high edge weight aggregation degree in the feedback state graph, the dominant context feedback direction is determined, that is, the reading context option with the highest concentration in the group feedback, and this context is used as the recommended reading context for the chapter, thus constructing a stable chapter context label.

[0089] This method, by constructing a feedback state graph, not only comprehensively considers the combination of subjective choices and objective behaviors, but also improves the accuracy and representativeness of tag generation by judging the quantitative similarity between feedback behaviors, ensuring that chapter context tags can truly reflect the reading experience preferences of most users.

[0090] In some embodiments of the application, the generation of chapter context labels also includes:

[0091] Perform graph density-based clustering analysis on the user feedback status graph to identify the subgroup of users with the highest concentration of feedback behavior. Within this subgroup:

[0092] Structural stability is determined by calculating the degree of edge weight fluctuation between each feedback user and the combination of feedback behaviors of its neighboring nodes. If the average edge weight variation coefficient of any feedback user is less than the preset structural stability coefficient threshold, then the feedback user is judged to have structural stability.

[0093] Feedback trajectory continuity is determined by tracking the reading context choices and feedback actions of a user across three consecutive chapters. If the trend of the four-element feedback vector changes monotonically or is within a local fluctuation range between the three chapters, then the user is judged to have feedback trajectory continuity.

[0094] Historical preference consistency is determined by comparing the cosine similarity between the current chapter's feedback behavior combination and the feedback behavior combination of the same user in the same tagged chapters in the past. If the similarity exceeds the set consistency judgment threshold, the feedback user is judged to have historical preference consistency.

[0095] The set of feedback users who simultaneously satisfy structural stability, feedback trajectory continuity, and historical preference consistency is defined as the context label reference group.

[0096] In this embodiment, after the number of valid reading feedbacks for a chapter reaches a set feedback threshold, a user feedback state graph for that chapter is first constructed. Specifically, all users who provide feedback on the current chapter are treated as nodes. Each user's node information includes a four-element feedback vector consisting of their selected reading context (e.g., indoor / outdoor, morning / night) and their feedback action. The four-element feedback vector consists of four components: page dwell time (in seconds), swipe frequency (number of swipes per minute), rating behavior (user rating value), and tag selection behavior (the set of keyword codes selected by the user in the tag library, encoded as a vector). This feedback vector describes the user's actual interaction characteristics with the chapter content in a specific context.

[0097] In the user feedback state graph, the edge weights represent the similarity of behavior between two users. The calculation method is as follows: Calculate the Euclidean distance between the four-element feedback vectors of any two users. After obtaining the distance value, perform inverse normalization on this distance value between the maximum and minimum values ​​of all feedback distances to obtain the edge weight. A larger value indicates more similar feedback behaviors. The inverse normalization method is as follows: Let the distance be d, and the maximum distance be d0. max The minimum distance is d min Then the edge weight is calculated as w = 1 - (dd) min ) / (d max -d min ).

[0098] After constructing the user feedback status graph, a graph density-based clustering analysis is performed. Graph density clustering is a clustering strategy based on the strength of edge weights between nodes and the density of node connectivity. By identifying the subgraph regions with the most concentrated edge weight distribution and the tightest internal connections, it filters out user groups with strong consistency in feedback behavior. Each user subgroup identified by this clustering process represents a set of users with similar reading feedback patterns.

[0099] Next, for each feedback user subgroup, three indicators are calculated: structural stability, continuity of feedback trajectory, and consistency of historical preferences.

[0100] Structural stability: For each feedback user, the edge weights between them and all adjacent nodes are counted, and the coefficient of variation (CV, i.e., the ratio of standard deviation to mean) of their edge weights is calculated. If the coefficient of variation is less than the preset structural stability coefficient threshold (e.g., 0.2), it indicates that the user has a relatively stable behavioral structure in the entire subgroup, and is therefore considered to have structural stability.

[0101] Feedback Trajectory Continuity: For each user providing feedback, track their feedback actions in the current chapter and the two adjacent chapters before and after, constructing a four-dimensional feedback vector for each of these three chapters. Calculate the trend of change in the value of each dimension among the three sets of feedback vectors. If the trend of change in the four-dimensional vector is monotonically increasing, decreasing, or within a stable local fluctuation range (the change amplitude is less than a preset tolerance value), then it is determined that their behavioral trajectory has temporal continuity, and thus they are considered to have feedback trajectory continuity.

[0102] Historical Preference Consistency: For each user providing feedback, all chapters with the same tags as the current chapter are filtered from their past reading history, and the corresponding four-element feedback vector set is extracted. Then, the cosine similarity between the current chapter's feedback vector and the feedback vectors of historical chapters with the same tags is calculated. If the similarity value exceeds a set threshold (e.g., 0.85), the user is considered to have consistent behavior under similar content, thus possessing historical preference consistency.

[0103] Ultimately, feedback users who simultaneously meet the above three criteria will be selected from a specific subgroup of feedback users to form the contextual tag reference group for this chapter. Members of this group, whose feedback behavior can be considered reliable, stable, and consistent representatives, will have their associated reading contexts used as the source of candidate contextual tags for this chapter. Subsequent contextual tag generation will be based on integrating and labeling the shared feedback paths and contextual choices of members within this group.

[0104] In some embodiments of the application, the generation of chapter context labels also includes:

[0105] Extract the context selection tags of feedback users from the context tag reference group, and calculate the context tag distribution density in the weather context dimension and the reading location dimension respectively. Then, calculate the Shannon entropy based on the tag probability distribution as the tag distribution entropy of that dimension.

[0106] When the label distribution entropy of any dimension is lower than the set label aggregation limit, the label with the highest frequency in that dimension is extracted as the chapter context labeling result;

[0107] When the entropy of the label distribution in any dimension is higher than the set label aggregation limit, the reading behavior trajectory data of all feedback users in the context label reference group is called. The reading behavior trajectory data is the context label and feedback action time series corresponding to the user's historical chapter reading records in the reading system.

[0108] Based on the context label of each candidate chapter, the behavioral trajectory associated with the label is identified in the reading behavior trajectory data of the context label reference group, and the proportion of context label annotations in the behavioral trajectory is calculated as the feedback trajectory aggregation degree of the context label of the candidate chapter.

[0109] If the aggregation degree of the feedback trajectory of any candidate chapter context label exceeds the aggregation consistency threshold, then the candidate chapter context label with the highest aggregation degree of feedback trajectory is selected as the chapter context label annotation result; if no label meets the aggregation consistency threshold, then no chapter context label is generated for the current chapter under this dimension.

[0110] The completed chapter context tags are attached to the chapter identifier field of the current chapter as structured fields.

[0111] In this embodiment, firstly, context selection tag data from all feedback users is extracted from the constructed context tag reference group. "Context selection tags" refer to the external environmental state information of the user during the reading process, including two dimensions: "weather context" (e.g., sunny, rainy, cloudy) and "reading location" (e.g., bedroom, office, commuting). These tags are typically obtained by the reading device through location, time, and user input information, or are selected autonomously by the user during the reading process.

[0112] Then, the distribution density of context labels is statistically analyzed under two context dimensions: the weather context dimension and the reading location dimension, representing the proportion of each context label's occurrences in the total occurrences. A probability distribution is then constructed based on this. Based on this probability distribution, the Shannon entropy (information entropy) value is calculated to quantify the dispersion of the label distribution. The formula for Shannon entropy is: H = -Σ(p i ×log(p i ), where H is the information entropy, p i Let be the probability of the i-th type of label appearing. The lower the entropy value, the more concentrated the label distribution.

[0113] Based on the calculated tag distribution entropy, it is determined whether the value is lower than the preset tag aggregation threshold. If the tag distribution entropy in any dimension is lower than the threshold, the contextual tags in that dimension are considered to have high consistency, that is, user preferences are highly concentrated. Therefore, the most frequent tag in that dimension is directly extracted as the chapter contextual tag labeling result for the current chapter.

[0114] If the tag distribution entropy is higher than the set aggregation threshold, it indicates that users have diverse choices in this dimension and do not exhibit a clear central tendency. In this case, instead of directly selecting the most frequent tag, we further utilize the historical "reading behavior trajectory data" of all feedback users in the context tag reference group. This behavior trajectory data refers to the time-series record of context tags and feedback actions formed by users during the reading of historical chapters. For example, a user may have marked a science fiction chapter as "tense" in a rainy day or office environment.

[0115] Next, for each candidate context label in the current chapter, we identify the behavioral trajectories associated with that label in the aforementioned trajectory data. Specifically, we find reading record segments that contain the context label and are similar to the context of the current chapter, and calculate the proportion of these trajectories in all records, which is used as the "feedback trajectory aggregation degree." A higher value indicates a stronger structural correlation between the label and feedback behavior in historical reading trajectories.

[0116] Determine if the aggregation degree of the feedback trajectories of all candidate context labels exceeds the set "aggregation consistency threshold". If it does, select the candidate label with the highest aggregation degree as the final labeling result of the chapter context label; if the aggregation degree of no label reaches the threshold, it is considered that the current chapter lacks sufficiently consistent user feedback in this dimension, and no chapter context label for the corresponding dimension is generated.

[0117] Finally, the finalized chapter context tags are linked to the chapter identifier field corresponding to the current chapter in the form of structured fields (such as JSON key-value pairs or database table fields) to enable the tags to be called and visualized in subsequent push, categorization, recommendation and other scenarios.

[0118] In some embodiments of the application, the construction of the feedback user profile set and the current user profile set includes:

[0119] Acquire target user behavior records and feedback data in the reading system. The behavior records include reading chapter records, rating records, reading time periods and device types used. The feedback data includes page dwell time, swipe frequency and chapter context preference tags.

[0120] Based on the behavioral record data and feedback data, a user profile vector group is generated, the user profile vector group including:

[0121] The behavioral feature vector is generated based on the target user's page dwell time and swipe frequency in each chapter, and is normalized to form a reading behavior distribution vector that reflects the intensity of reading operations.

[0122] The time preference vector is generated based on the statistical results of the target user's historical reading time period. The day is divided into multiple fixed time intervals, the reading frequency in each interval is counted and normalized to form a reading time preference vector that reflects the distribution of its time activity.

[0123] Device adaptation label, which is the device type identifier that appears most frequently in the target user's historical device types, is used to indicate the reading device that the user mainly uses;

[0124] The rating preference vector is generated based on the target user's historical rating records and includes three components: average rating under different tag categories, rating variation, and high rating ratio, which are used to reflect the user's rating trend and preference concentration.

[0125] The context preference vector is generated based on the weather and location tags selected by the user in historical feedback. The frequency of occurrence of each tag is counted and weighted and normalized to obtain a context feature expression that reflects the user's typical reading preference environment.

[0126] The behavioral feature vector, time preference vector, device adaptation tag, rating preference vector, and context preference vector are used to form the corresponding feedback user profile set and the current user profile set.

[0127] In this embodiment, the first step is the collection of user behavior record data and feedback data: This involves acquiring the target user's behavior record data and feedback data from the reading system. Behavior record data refers to log information of each reading action performed by the user in the system, including but not limited to the chapter identifier read, the entry and exit times for each chapter (used to calculate page dwell time), the number or frequency of swiping actions within the chapter (such as vertical scrolling), and the rating score assigned. Feedback data includes chapter context preference tags manually selected by the user in specific chapters (such as "sunny weather" or "subway location"). These tags are typically filled out by the user through questionnaires, pop-ups, or button options. Additionally, it includes auxiliary behaviors automatically collected by the system, such as page dwell time and swipe frequency, used to reflect the depth of user interaction.

[0128] The second step is the generation of the behavioral feature vector: The behavioral feature vector is used to describe the intensity of a user's reading operations on different chapters. Specifically, for each chapter, the user's page dwell time and swipe frequency are recorded; the dwell time and swipe frequency for all chapters are normalized by dividing each value by the current user's maximum value across all chapters, thus limiting the data range to between 0 and 1; the two normalized values ​​(dwell time and swipe frequency) are concatenated by chapter to form a behavioral feature vector, which reflects the user's reading intensity on the corresponding chapter in each dimension.

[0129] The third step is the construction of the time preference vector: To depict the distribution of a user's active reading time throughout the day, the 24 hours are divided into several fixed time periods, such as 0:00–6:00, 6:00–12:00, 12:00–18:00, and 18:00–24:00, totaling four segments. For each time period, the total number of chapters read by the user within that time period is counted; the reading frequency of these four time periods is normalized (the frequency of each segment divided by the total number of times); resulting in a vector of length 4, called the time preference vector, which describes the distribution of a user's reading activity across different time periods throughout the day.

[0130] The fourth step is the generation of device compatibility tags: Device compatibility tags are used to identify the types of reading devices that users use most frequently. The types of devices used by the user across all reading records (e.g., mobile phone, tablet, computer, etc.) are counted; the most frequent device type is identified and used as the current user's device compatibility tag. This tag is a discrete identifier, mainly used for subsequent user segmentation and recommendation preference adjustments.

[0131] The fifth step is to construct the rating preference vector: This vector reflects users' rating tendencies for different types of chapter content. Specific steps include: 1) Based on the chapter content's tag classification in the system (e.g., "inspirational," "suspense," "emotional"), statistically analyze user rating records for each category; 2) For each category, calculate three statistical indicators: average rating, standard deviation of the rating (degree of variation), and the percentage of ratings greater than a certain threshold (e.g., 4 points) (high rating percentage); 3) Concatenate the three statistical indicators for all categories to form the rating preference vector, which reflects users' rating habits and concentration across different content categories.

[0132] Step 6: Constructing the Contextual Preference Vector: This vector reflects the user's preferred reading context (e.g., a preference for reading on a sunny day or on the subway). Statistical analysis is performed on the contextual tags (including weather and location) recorded in the user's historical feedback to create a tag frequency table. The frequency of each tag is weighted (e.g., recent feedback has a higher weight), and then normalized so that the sum of the weighted frequencies of all tags is 1. This results in two sub-vectors representing the preference distribution in the weather and location dimensions, respectively, which are then concatenated into a single contextual preference vector.

[0133] Step 7: Generation and Integration of User Profile Sets: The vectors and tags generated by the above five sub-modules are integrated into a unified user profile structure. The current user's various feature data sets constitute the current user profile set; the corresponding feature data sets of all historical feedback users (i.e., users who have provided feedback on chapter context tags or content) constitute the feedback user profile set; both are stored in a structured format for subsequent user matching, similarity calculation, and personalized recommendation processing.

[0134] The above method lays the foundation for context-aware chapter tag recommendation and situational adaptation by refining the multi-dimensional expression of user behavior and preference features. To further refine the similarity calculation between profiles or the construction of tag inference paths, corresponding graph structures or scoring functions can be added on this basis.

[0135] In some embodiments of the application, determining the set of ebook categories suitable for the current user includes:

[0136] The system calculates the reading time of each category of ebooks in the current user's reading history, and then selects the top N ebook categories by reading time to form a set of ebook categories that are suitable for the current user, where N is a preset upper limit for the number of categories.

[0137] It should be noted that N is the preset upper limit of the number of categories. The value of N can be set according to the actual situation, and is usually set between 3 and 5.

[0138] In some embodiments of the application, behavioral similarity and historical data similarity are obtained through the following steps:

[0139] Obtain the current user profile vector group and the profile vector group of each feedback user. The profile vector group includes behavioral feature vectors, time preference labels, device adaptation labels, rating preference vectors, and a set of contextual preference labels.

[0140] Calculating behavioral similarity includes:

[0141] Cosine similarity is used to calculate the similarity between behavioral feature vectors;

[0142] If the time preference tags are completely identical, the time matching score is 1; if they are adjacent time periods, the score is calculated according to the preset reduction ratio.

[0143] The behavioral similarity is obtained by weighting and fusing the behavioral feature similarity with the time matching score according to the set weights;

[0144] Calculating historical data similarity includes:

[0145] The similarity of rating preference vectors is calculated using Euclidean distance or KL divergence.

[0146] Extract common tags from the contextual preference tag sets of the current user and the feedback user, and calculate tag similarity scores based on the degree of overlap in tag weight distribution;

[0147] The similarity scores of rating preferences and tags are weighted and fused according to a set weight to obtain the similarity of historical data;

[0148] The weighted product of behavioral similarity and historical data similarity is used as the first similarity.

[0149] In this embodiment, the process of obtaining behavioral similarity and historical data similarity aims to measure the similarity between the current user and the user providing feedback in terms of reading behavior and historical preferences. The specific implementation method is as follows:

[0150] First, the system obtains the current user profile vector set and the feedback user profile vector set. Each user profile vector set contains five dimensions: behavioral feature vector, time preference tag, device adaptation tag, rating preference vector, and contextual preference tag set. These dimensions are used to comprehensively describe the user's reading habits and preferences.

[0151] In calculating behavioral similarity, the similarity between the current user's behavioral feature vector and that of each feedback user is first assessed. These behavioral feature vectors are constructed based on the user's page dwell time and swipe frequency across different chapters. Specifically, the user's page dwell time and swipe frequency in each chapter are first counted and normalized to form a vector reflecting the intensity distribution of the user's reading actions. The similarity between two behavioral feature vectors is calculated using cosine similarity, treating the two vectors as points in a high-dimensional space and calculating the cosine of the angle between them. The smaller the angle, the closer the cosine value is to 1, indicating a more consistent behavioral distribution.

[0152] Subsequently, the system assigns a time matching score based on the user's time preference tags. These time preference tags are formed by statistically analyzing the user's reading frequency across various fixed time periods throughout the day, reflecting the user's reading activity level at different times. If the current user's time preference tags are completely identical to the feedback user's (i.e., their active time periods are exactly the same), a time matching score of 1 is assigned. If their active time periods are adjacent (e.g., one is "18:00–21:00" and the other is "21:00–24:00"), a score is calculated based on a set reduction ratio function. For example, the score can be set to 0.8, 0.5, etc., to form an approximate matching time score value.

[0153] The cosine similarity score and the time matching score are weighted and fused according to the set weights to obtain the final behavior similarity score. For example, if the behavior feature similarity weight is set to 0.7 and the time matching weight is set to 0.3, the final behavior similarity score is the sum of the two scores multiplied by their respective weights.

[0154] Next, the similarity of historical data is calculated. The rating preference vector is constructed based on users' rating behavior under different e-book tag categories and includes three components: the average rating under that tag, the standard deviation of the rating (reflecting the degree of rating variation), and the proportion of high ratings (the proportion of ratings above a certain rating threshold). For the similarity calculation between two rating preference vectors, Euclidean distance or Kullback-Leibler divergence (KL divergence) can be selected according to specific needs. The former is suitable for numerical proximity analysis, while the latter is suitable for judging the difference in probability distributions. If Euclidean distance is used, the sum of squares of the differences between the two vectors in each dimension is calculated first, and then the square root is taken. The smaller the value, the higher the similarity. If KL divergence is used, the two vectors are normalized and then used as a probability distribution for divergence calculation. The smaller the divergence, the closer the distributions are.

[0155] In addition, the system also needs to evaluate the similarity of users' contextual preference tags across different contexts. Contextual preference tags are a set of tags selected by users in their historical reading history or inferred by the system, including typical reading weather (such as "sunny" and "rainy") and reading location (such as "indoors" and "subway"). The system extracts common tags from the contextual preference tag sets of the current user and the feedback user, and constructs a tag weight distribution vector based on the weight of each tag in the two sets (such as the proportion of a tag's frequency to the total frequency of all tags). Then, it compares the degree of overlap in the distributions between the two users. Indicators such as overlap coefficient or cross-entropy can be used; the higher the value, the more consistent the contextual preferences, thus obtaining a tag similarity score.

[0156] Finally, the rating preference similarity score and the tag similarity score are weighted and merged according to the set weights to form a complete historical data similarity score. For example, if rating preference and tag preference each account for 0.5 weights, the weighted result is the average of the two.

[0157] After obtaining behavioral similarity and historical data similarity, the system performs a weighted product of these two scores to obtain the first similarity score. This product operation effectively avoids interference from extremely low similarity in one dimension on the overall judgment, thus ensuring that only feedback users who highly match in both behavioral and historical aspects can participate in the subsequent weight allocation process of the recommendation strategy. This first similarity score serves as a quantitative basis for the comprehensive profile similarity between the current user and the feedback user, driving the subsequent e-book recommendation calculation process.

[0158] In some embodiments of the application, the context matching degree and historical weather matching degree are obtained through the following steps:

[0159] The current user's real-time location and weather information are compared with the best reading location and best reading weather in the chapter context tags corresponding to the candidate chapters. If the current user's real-time weather information matches the best reading weather in the chapter context tags, the highest matching value is assigned; if there is only a correlation in weather type, a medium matching value is assigned; if there is no correlation, the lowest matching value is assigned. The current user's real-time location information is compared with the best reading location in the chapter context tags. If the current user is in the location type corresponding to the best reading location, the highest matching value is assigned; otherwise, a second-highest matching value is assigned based on the similarity of location types. The context matching degree between the current user and the candidate chapters is calculated based on the above matching values.

[0160] Obtain the historical weather tag sequence corresponding to the current user's historical reading record, and compare it with the historical weather information recorded by the feedback users who participated in the feedback of this chapter. Calculate the historical weather matching degree between the current user and the feedback user by weighted fusion of the co-occurrence frequency of historical weather tags and the overlap of time period distribution. The product of the context matching degree and the historical weather matching degree is used as the second similarity.

[0161] In this embodiment, in order to calculate the adaptability of the chapter content to the current user's reading environment, it is necessary to extract the current user's real-time location information and real-time weather information.

[0162] Real-time location information: refers to the current geographic coordinates of the user obtained through GPS, Wi-Fi positioning or base station positioning of the mobile device; Real-time weather information: refers to the weather conditions corresponding to the user's location at the current moment, including weather type (such as sunny, rainy, snowy, cloudy), temperature range, humidity level, etc., which are obtained by the system calling the meteorological service interface.

[0163] Then, the system compares the above two pieces of information with the chapter context tags corresponding to each candidate chapter. The chapter context tags have two preset attributes: Best reading location: refers to the type of place most suitable for reading the chapter, such as "bed", "coffee shop", "subway car", etc., which are all scene classification tags; Best reading weather: refers to the weather conditions most suitable for reading the chapter, such as "sunny and hot", "light rain and cool", "cloudy and comfortable", etc., which are set by the content planner or feedback mechanism based on the user's historical preferences.

[0164] The matching rules are as follows: If the current weather is exactly the same as the best reading weather marked in the chapter, then a weather matching value of 1 is assigned; if there is a partial relationship between the weather types (for example, "light rain" and "cloudy" are semantically similar), then a medium matching value, such as 0.5, is assigned; if there is no obvious relationship between the two, then the lowest matching value, recorded as 0, is assigned.

[0165] For location matching: First, the current user's location coordinates are mapped to location types (such as "office area", "home", "outdoor leisure place" etc.). This mapping is based on the location POI (Point of Interest) database. If the location type is consistent with the best reading location set in the chapter, a location matching value of 1 is assigned. If the location types are inconsistent but belong to similar categories semantically (such as "office area" and "study room"), a second-level matching value is assigned, such as 0.6. If the types have no obvious relationship, the matching value is 0.

[0166] Ultimately, the context matching score is calculated by weighted average or product of weather matching and location matching scores, and is used to assess the degree of fit between the chapter context and the user's current reading context.

[0167] Historical weather matching accuracy acquisition method: First, extract the historical weather tag sequence associated with the current user's past reading behavior. This tag sequence comes from the weather information collected each time the user's reading behavior occurs (also obtained by calling the meteorological interface), and is converted into standardized tags (such as "high temperature and sunny day", "moderate haze", "moderate rain", etc.) to form time series data.

[0168] Secondly, the system extracts the feedback user group corresponding to each candidate chapter, which is the set of users who have historically provided positive feedback on the chapter (such as rating, favorites, recommendations, etc.). For each feedback user, the system records their historical weather information while reading the chapter and constructs a set of reading weather tags for the feedback user.

[0169] The matching degree calculation method includes: tag co-occurrence frequency: between the current user's historical weather tag sequence and the feedback user's weather tag set, the number of times the same tag appears is counted and normalized to a frequency value; time period distribution overlap: the time period of each weather tag in the current user's historical reading record is compared with the time period of the corresponding tag in the feedback user. If the time period is highly concentrated (e.g., all concentrated in the early morning or late at night), it is considered that their weather conditions are more common in physiological rhythms and are given a higher overlap score.

[0170] Finally, the co-occurrence frequency of tags and the overlap of time period distribution are weighted and fused according to the set weights to obtain the historical weather matching degree between the current user and the feedback user.

[0171] The second similarity calculation method: The two calculation results (contextual matching degree and historical weather matching degree) are treated as two independent but jointly influencing reading preferences. Their final combination is expressed as a product (or other weighted combination method), i.e.:

[0172] Second similarity = contextual matching degree × historical weather matching degree.

[0173] This second similarity score serves as one of the important indicators for measuring the suitability of user context and chapter recommendations during the recommendation process, and participates in subsequent recommendation ranking and filtering judgment processes.

[0174] In some embodiments of the application, generating the current user's chapter recommendation list includes:

[0175] The first similarity and the second similarity of each candidate chapter are multiplied to obtain a comprehensive matching score; all candidate chapters are sorted in descending order based on the comprehensive matching score to construct a chapter recommendation sequence; the top few chapters in the chapter recommendation sequence are selected as the chapter recommendation list for the current user.

[0176] See Figure 2 As shown, this embodiment of the invention provides an intelligent e-book recommendation system based on multi-source data, including:

[0177] The data collection module is configured to collect reading feedback for each chapter of the e-book. After the user finishes reading a chapter, it collects the recommended reading context for the current chapter through an interactive window. The recommended reading context is the best weather and best reading location for the current chapter.

[0178] The module is configured to, when the number of valid reading feedbacks for the current chapter reaches a set threshold, statistically analyze the recommended reading context for the current chapter based on the responses of all users who participated in the reading feedback, and label the current chapter in the form of tags, which are denoted as chapter context tags; and acquire the user behavior data and historical reading data of the feedback users to construct a set of feedback user profiles for the current chapter.

[0179] The category module is configured to construct a set of user profiles to be recommended based on the current user's user behavior data and historical reading data, and determine the set of e-book categories suitable for the current user; each set of e-book categories corresponds to a set of e-book chapters containing chapter context tags;

[0180] The first calculation module is configured to calculate the behavioral similarity and historical data similarity between the current user and the feedback user profile set in the e-book chapter set, and determine the first similarity based on the behavioral similarity and historical data similarity.

[0181] The second calculation module is configured to compare the current user's real-time location and real-time weather with the best reading location and best reading weather respectively to determine the context matching degree; obtain the current user's historical weather information and compare it with the historical weather when the user participated in reading feedback to determine the historical weather similarity; and determine the second similarity degree based on the context matching degree and the historical weather matching degree.

[0182] The list generation module is configured to sort candidate chapters based on the product of the first similarity and the second similarity, and generate a recommended chapter list for the current user.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent e-book recommendation based on multi-source data, characterized in that, include: The system collects reading feedback for each chapter of the e-book. After a user finishes reading a chapter, it collects recommended reading scenarios for the current chapter through an interactive window. These recommended scenarios include the best weather and the best reading location for the current chapter. When the number of valid reading feedbacks for the current chapter reaches a set threshold, the recommended reading context for the current chapter is statistically analyzed based on the responses of all users who participated in the reading feedback, and the current chapter is labeled in the form of tags, which are called chapter context tags; Obtain the user behavior data and historical reading data of the feedback users to construct a set of feedback user profiles for the current chapter; Based on current user behavior data and historical reading data, construct a set of user profiles to be recommended, and determine the set of e-book categories suitable for the current user; Each ebook category corresponds to a set of ebook chapters containing chapter context tags; In the e-book chapter set, the behavioral similarity and historical data similarity between the current user and the feedback user profile set are calculated respectively, and the first similarity is determined based on the behavioral similarity and historical data similarity. The current user's real-time location and real-time weather are compared with the best reading location and best reading weather to determine the context matching degree; the current user's historical weather information is obtained and compared with the historical weather when the user participated in the reading feedback to determine the historical weather similarity; and a second similarity degree is determined based on the context matching degree and the historical weather matching degree. Based on the product of the first similarity and the second similarity, the candidate chapters in the e-book chapter set containing chapter context tags corresponding to the e-book category set adapted to the current user are sorted to generate the chapter recommendation list for the current user.

2. The intelligent e-book recommendation method based on multi-source data according to claim 1, characterized in that, The generation of chapter context tags includes: Once the number of valid reading feedbacks for the current chapter reaches a set feedback threshold, a user feedback state graph is constructed based on all users who provided reading feedback for the current chapter. Each node of the user feedback state graph is a combination of the reading context selected by a user in the current chapter and the corresponding feedback action. The combination of feedback actions is a four-element feedback vector composed of the user's page dwell time, swipe frequency, rating behavior and tag selection behavior recorded in the current chapter. The edge weight between any two nodes in the user feedback state graph is the similarity weight between the combination of feedback behaviors of the two users. The similarity weight of the combination of feedback behaviors is obtained by calculating the Euclidean distance of the quaternary feedback vector and performing inverse normalization.

3. The intelligent e-book recommendation method based on multi-source data according to claim 2, characterized in that, The generation of chapter context tags also includes: Perform graph density-based clustering analysis on the user feedback status graph to identify the subgroup of users with the highest concentration of feedback behavior. Within this subgroup: Structural stability is determined by calculating the degree of edge weight fluctuation between each feedback user and the combination of feedback behaviors of its neighboring nodes. If the average edge weight variation coefficient of any feedback user is less than the preset structural stability coefficient threshold, then the feedback user is judged to have structural stability. Feedback trajectory continuity is determined by tracking the reading context choices and feedback actions of a user across three consecutive chapters. If the trend of the four-element feedback vector changes monotonically or is within a local fluctuation range between the three chapters, then the user is judged to have feedback trajectory continuity. Historical preference consistency is determined by comparing the cosine similarity between the current chapter's feedback behavior combination and the feedback behavior combination of the same user in the same tagged chapters in the past. If the similarity exceeds the set consistency judgment threshold, the feedback user is judged to have historical preference consistency. The set of feedback users who simultaneously satisfy structural stability, feedback trajectory continuity, and historical preference consistency is defined as the context label reference group.

4. The intelligent e-book recommendation method based on multi-source data according to claim 3, characterized in that, The generation of chapter context tags also includes: Extract the context selection tags of feedback users from the context tag reference group, and calculate the context tag distribution density in the weather context dimension and the reading location dimension respectively. Then, calculate the Shannon entropy based on the tag probability distribution as the tag distribution entropy of that dimension. When the label distribution entropy of any dimension is lower than the set label aggregation limit, the label with the highest frequency in that dimension is extracted as the chapter context labeling result; When the entropy of the label distribution in any dimension is higher than the set label aggregation limit, the reading behavior trajectory data of all feedback users in the context label reference group is called. The reading behavior trajectory data is the context label and feedback action time series corresponding to the user's historical chapter reading records in the reading system. Based on the context label of each candidate chapter, the behavioral trajectory associated with the label is identified in the reading behavior trajectory data of the context label reference group, and the proportion of context label annotations in the behavioral trajectory is calculated as the feedback trajectory aggregation degree of the context label of the candidate chapter. If the aggregation degree of the feedback trajectory of any candidate chapter context label exceeds the aggregation consistency threshold, then the candidate chapter context label with the highest aggregation degree of feedback trajectory is selected as the chapter context label annotation result; if no label meets the aggregation consistency threshold, then no chapter context label is generated for the current chapter under this dimension. The completed chapter context tags are attached to the chapter identifier field of the current chapter as structured fields.

5. The intelligent e-book recommendation method based on multi-source data according to claim 4, characterized in that, The construction of the feedback user profile set and the current user profile set includes: Acquire target user behavior records and feedback data in the reading system. The behavior records include reading chapter records, rating records, reading time periods and device types used. The feedback data includes page dwell time, swipe frequency and chapter context preference tags. Based on the behavioral record data and feedback data, a user profile vector group is generated, the user profile vector group including: The behavioral feature vector is generated based on the target user's page dwell time and swipe frequency in each chapter, and is normalized to form a reading behavior distribution vector that reflects the intensity of reading operations. The time preference vector is generated based on the statistical results of the target user's historical reading time period. The day is divided into multiple fixed time intervals, the reading frequency in each interval is counted and normalized to form a reading time preference vector that reflects the distribution of its time activity. Device adaptation label, which is the device type identifier that appears most frequently in the target user's historical device types, is used to indicate the reading device that the user mainly uses; The rating preference vector is generated based on the target user's historical rating records and includes three components: average rating under different tag categories, rating variation, and high rating ratio, which are used to reflect the user's rating trend and preference concentration. The context preference vector is generated based on the weather and location tags selected by the user in historical feedback. The frequency of occurrence of each tag is counted and weighted and normalized to obtain a context feature expression that reflects the user's typical reading preference environment. The behavioral feature vector, time preference vector, device adaptation tag, rating preference vector, and context preference vector are used to form the corresponding feedback user profile set and the current user profile set.

6. The intelligent e-book recommendation method based on multi-source data according to claim 5, characterized in that, When determining the set of ebook categories suitable for the current user, the following are included: The system calculates the reading time of each category of ebooks in the current user's reading history, and then selects the top N ebook categories by reading time to form a set of ebook categories that are suitable for the current user, where N is a preset upper limit for the number of categories.

7. The intelligent e-book recommendation method based on multi-source data according to claim 6, characterized in that, Behavioral similarity and historical data similarity are obtained through the following steps: Obtain the current user profile vector group and the profile vector group of each feedback user. The profile vector group includes behavioral feature vectors, time preference labels, device adaptation labels, rating preference vectors, and a set of contextual preference labels. Calculating behavioral similarity includes: Cosine similarity is used to calculate the similarity between behavioral feature vectors; If the time preference tags are completely identical, the time matching score is 1; if they are adjacent time periods, the score is calculated according to the preset reduction ratio. The behavioral similarity is obtained by weighting and fusing the behavioral feature similarity with the time matching score according to the set weights; Calculating historical data similarity includes: The similarity of rating preference vectors is calculated using Euclidean distance or KL divergence. Extract common tags from the contextual preference tag sets of the current user and the feedback user, and calculate tag similarity scores based on the degree of overlap in tag weight distribution; The similarity scores of rating preferences and labels are weighted and fused according to a set weight to obtain the historical data similarity; the weighted product of behavioral similarity and historical data similarity is used as the first similarity.

8. The intelligent e-book recommendation method based on multi-source data according to claim 7, characterized in that, Contextual matching and historical weather matching are obtained through the following steps: The current user's real-time location and weather information are compared with the best reading location and best reading weather in the chapter context tags corresponding to the candidate chapters. If the current user's real-time weather information matches the best reading weather in the chapter context tags, the highest matching value is assigned; if there is only a correlation in weather type, a medium matching value is assigned; if there is no correlation, the lowest matching value is assigned. The current user's real-time location information is compared with the best reading location in the chapter context tags. If the current user is in the location type corresponding to the best reading location, the highest matching value is assigned; otherwise, a second-highest matching value is assigned based on the similarity of location types. The context matching degree between the current user and the candidate chapters is calculated based on the above matching values. Obtain the historical weather tag sequence corresponding to the current user's historical reading record, and compare it with the historical weather information recorded by the feedback users who participated in the feedback of this chapter. Calculate the historical weather matching degree between the current user and the feedback user by weighted fusion of the co-occurrence frequency of historical weather tags and the overlap of time period distribution. The product of the context matching degree and the historical weather matching degree is used as the second similarity.

9. The intelligent e-book recommendation method based on multi-source data according to claim 8, characterized in that, When generating the current user's recommended chapter list, it includes: The first similarity and the second similarity of each candidate chapter are multiplied to obtain a comprehensive matching score; all candidate chapters are sorted in descending order based on the comprehensive matching score to construct a chapter recommendation sequence; the top few chapters in the chapter recommendation sequence are selected as the chapter recommendation list for the current user.

10. An intelligent e-book recommendation system based on multi-source data, used to implement the method described in any one of claims 1-9, characterized in that, The system includes: The data collection module is configured to collect reading feedback for each chapter of the e-book. After the user finishes reading a chapter, it collects the recommended reading context for the current chapter through an interactive window. The recommended reading context is the best weather and best reading location for the current chapter. The module is configured to, when the number of valid reading feedbacks for the current chapter reaches a set threshold, statistically analyze the recommended reading context for the current chapter based on the responses of all users who participated in the reading feedback, and label the current chapter in the form of tags, which are denoted as chapter context tags; and acquire the user behavior data and historical reading data of the feedback users to construct a set of feedback user profiles for the current chapter. The category module is configured to construct a set of user profiles to be recommended based on the current user's user behavior data and historical reading data, and determine the set of e-book categories suitable for the current user; each set of e-book categories corresponds to a set of e-book chapters containing chapter context tags; The first calculation module is configured to calculate the behavioral similarity and historical data similarity between the current user and the feedback user profile set in the e-book chapter set, and determine the first similarity based on the behavioral similarity and historical data similarity. The second calculation module is configured to compare the current user's real-time location and real-time weather with the best reading location and best reading weather respectively to determine the context matching degree; obtain the current user's historical weather information and compare it with the historical weather when the user participated in reading feedback to determine the historical weather similarity; and determine the second similarity degree based on the context matching degree and the historical weather matching degree. The list generation module is configured to sort the candidate chapters in the e-book chapter set containing chapter context tags corresponding to the e-book category set adapted to the current user based on the product of the first similarity and the second similarity, and generate a chapter recommendation list for the current user.

Images