A book recommendation method based on reader behavior feedback update

Abstract

This application provides a book recommendation method based on reader behavior feedback updates, belonging to the field of intelligent information processing technology. The method includes: acquiring reader behavior data in a book service platform; constructing a behavior time-series dataset; forming a behavior sequence feature dataset; generating a current reading state dataset; forming a reader interest feature dataset; generating a recommendation matching dataset; outputting recommendation results; and generating a feedback behavior dataset. This application solves the problem that existing recommendation systems mainly rely on static interest tags or single-time behavior data, failing to dynamically perceive changes in reader interests, leading to lagging recommendation results and insufficient accuracy. By establishing a multi-level state determination mechanism, state transition calculation, and recursive updating of book topic features, the method improves the recommendation system's ability to perceive dynamic changes in reader interests, enhances recommendation accuracy and personalization, and ensures real-time, accurate, and personalized recommendations.

Description

Technical Field

[0001] This invention relates to the field of intelligent information processing technology, specifically to a book recommendation method based on reader behavior feedback. Background Technology

[0002] With the widespread adoption of digital reading platforms and online book services, book recommendation technology is extensively used in digital libraries, e-book platforms, and online bookstores to help readers quickly find content of interest from a vast library of books. Existing recommendation methods are typically based on readers' historical borrowing and browsing records, or manually set book category tags, and their recommendation rules are mostly statically configured or periodically updated. While these methods can reflect readers' long-term reading preferences to some extent, their ability to respond to recent changes in readers' reading behavior is limited, making it difficult to capture the dynamic evolution of interests in a timely manner, resulting in recommendation results that lag behind readers' actual needs.

[0003] Furthermore, existing technologies utilize reader behavior in a rather rudimentary way, typically focusing only on whether an action occurs, lacking analysis of the sequence and continuity of these actions. This makes it difficult to distinguish between browsing and sustained reading, resulting in recommendations that fail to accurately reflect the reader's current interest. Simultaneously, most recommendation systems lack dynamic correction mechanisms based on subsequent behavior; recommendation strategy updates rely on manual intervention or offline updates, failing to achieve continuous optimization of recommendation results. Therefore, existing technologies lack a method that can combine reader behavior feedback to determine reading status and dynamically adjust recommended content based on changes in reading status, leading to a discrepancy between recommendation results and the reader's actual interests, thus impacting recommendation effectiveness. Summary of the Invention

[0004] This application provides a book recommendation method based on reader behavior feedback updates, aiming to solve the problems of existing recommendation systems that mainly rely on static interest tags or single-time behavior data for matching, failing to continuously perceive changes in reader reading status and interest migration, resulting in lagging recommendation results, insufficient accuracy, and low personalization. This invention generates reading status and status confidence information by performing multi-level state determination on the reader behavior sequence, and recursively updates it by combining it with book topic feature vectors to form interest-oriented feature data; simultaneously, it introduces a state adjustment mechanism to calculate the recommendation matching value, realizing dynamic adjustment and adaptive output of recommendation results, thereby improving the recommendation system's responsiveness to the evolution of reader interests and recommendation accuracy.

[0005] In view of the above problems, this application provides a book recommendation method based on reader behavior feedback updates, including: acquiring reader behavior data in a book service platform; performing structured processing on the behavior data to generate a set of behavior record units containing behavior type identifiers, behavior timestamps, book identifiers, and behavior durations, and storing this set as a reader original behavior dataset; sorting the reader original behavior dataset by behavior timestamps to construct a behavior time series dataset; performing segmented statistical analysis on the behavior time series dataset within a preset time window, calculating the behavior frequency and behavior duration values ​​within each time period, generating a behavior intensity index sequence, and forming a behavior sequence feature dataset; based on the behavior sequence feature dataset, calling a preset state determination function to perform interval mapping on the behavior intensity index sequence to generate reading state identifiers, constructing a reading state transition matrix based on the reading state identifiers of adjacent time windows, and generating a current reading state dataset; and, based on the current reading state dataset and the reading state transition matrix, and in conjunction with the behavior time series dataset, calling... A book feature association mapping table is used to extract book topic vectors corresponding to the current reading state. Based on the reading state transition matrix, the book topic vectors are recursively accumulated to generate interest-oriented feature vectors, forming a reader interest feature dataset. This dataset and the current reading state identifier are input into a recommendation rule triggering function to calculate a recommendation matching value and generate a recommendation matching dataset. When the recommendation matching value meets a preset trigger threshold, a candidate book set is generated from the book resource library based on the recommendation matching dataset, sorted according to the recommendation matching value, and the recommendation result is output. After the recommendation result is output, subsequent behavioral data of the reader on the candidate book set is collected to generate a feedback behavior dataset, which is then merged into the original reader behavior dataset. A consistency comparison is performed between the feedback behavior dataset and the current reading state dataset to calculate the state deviation value. When the state deviation value exceeds a preset correction threshold, the parameter sets of the state determination function and the recommendation rule triggering function are updated.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: By performing multi-level state determination on reader behavior sequences and introducing a state confidence mechanism, refined identification of reading states is achieved, avoiding misjudgments caused by relying solely on single behaviors or simple threshold judgments, thus improving the accuracy and stability of state identification. By analyzing the state evolution relationship between adjacent time windows, the trend of reader interest changes can be characterized, enabling the recommendation system to respond to the interest migration process. By mapping state information to the book topic vector space, a state-driven interest-oriented feature vector is generated, and historical interest information is integrated using a recursive update method to achieve dynamic evolution of interest features, enhancing the recommendation system's adaptability to real-time behavioral feedback. By setting recommendation rule trigger functions and state adjustment factors, the recommendation matching value can be adaptively adjusted according to the current reading state, strengthening interest relevance while taking into account content diversity, thereby improving the accuracy, personalization, and user experience of recommendation results.

[0007] In summary, the book recommendation method based on reader behavior feedback provided in this application achieves continuous characterization and adaptive recommendation updates of the dynamic evolution of reader interests by performing multi-level state determination on the behavior sequence, constructing a reading state transition matrix, and generating state-driven interest-oriented feature vectors. This significantly improves the recommendation system's ability to perceive real-time changes in reader interests, the accuracy of recommendation results, and the level of personalization.

[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating a book recommendation method based on reader behavior feedback, provided as an embodiment of this application. Detailed Implementation

[0010] This application provides a book recommendation method based on reader behavior feedback, which solves the problem in the prior art that recommendation systems rely only on historical behavior data or static interest tags and cannot dynamically capture changes in readers' reading status and interest migration, resulting in lagging recommendation results, insufficient accuracy, and low personalization. This method achieves continuous perception and adaptive updating of the dynamic evolution of readers' interests, thereby improving the real-time response capability, recommendation accuracy, and personalization level of the recommendation system.

[0011] Example 1, as Figure 1 As shown in the figure, this application provides a book recommendation method based on reader behavior feedback, the method including: S1: Obtain the reader's behavior data in the book service platform, perform structured processing on the behavior data, generate a set of behavior record units containing behavior type identifier, behavior timestamp, book identifier and behavior duration, and store it as the reader's original behavior dataset.

[0012] Specifically, the system acquires reader behavior data from the book service platform, performs structured processing on this data, generates a set of behavior record units containing behavior type identifiers, behavior timestamps, book identifiers, and behavior durations, and stores this set as the reader's original behavior dataset. This behavior data includes, but is not limited to, click behavior, reading behavior, borrowing behavior, collection behavior, rating behavior, and sharing behavior. Click behavior refers to a reader's clicks on book covers, chapters, or recommendation lists; reading behavior refers to the start and end times of a reader opening a book and the cumulative reading time; borrowing behavior includes the borrowing of physical books or e-books and the borrowing duration; collection behavior refers to a reader adding books to their personal bookshelf or favorites; rating behavior includes readers rating, commenting on, or tagging books; and sharing behavior refers to readers sharing books or reading notes to social media platforms or friends. The sources of this behavior data include the book service platform's log database, mobile or web application operation records, third-party reading plugin data interfaces, and the backend statistics system. During data collection, the raw behavioral data undergoes structured processing. Each behavioral record is uniformly formatted into a behavioral record unit, including a behavior type identifier, a behavior timestamp accurate to the second, a unique book identifier, the behavior duration, and a unique user identifier for the reader. This facilitates subsequent construction of behavior time series, determination of reading status, and generation of interest feature vectors. After processing, all behavioral record units are archived and stored by reader, forming a raw reader behavior dataset. This achieves standardized storage and management of multi-source, multi-type behavioral data, ensuring data integrity, traceability, and analyzability. It provides a reliable data foundation for subsequent analysis of reader behavior patterns, identification of reading status, and dynamic updates to recommendation strategies, thereby improving the recommendation system's responsiveness to changes in reader interests and the effectiveness of personalized recommendations.

[0013] S2: Sort the original reader behavior dataset by behavior timestamp to construct a behavior time series dataset.

[0014] Specifically, the process begins by extracting behavioral record units from the original reader behavior dataset. These units include fields such as behavior type identifier, behavior timestamp, book identifier, behavior duration, and user identifier. All behavior records are then categorized by user identifier, and the behavior data of the same reader is processed centrally. Next, each reader's behavior records are sorted in ascending order based on their behavior timestamps, ensuring that each behavior is arranged chronologically to form a continuous time series that accurately reflects the reader's actual operational trajectory and reading pace on the platform. During the sorting process, missing or abnormal timestamp data is simultaneously validated and processed, including standardizing records with inconsistent timestamp formats and deduplicating or merging duplicate or conflicting behavior records to guarantee the completeness and accuracy of the sorting results. Furthermore, the time series can be labeled with behavior types, such as the start and end times of reading, borrowing, or clicking behaviors. This allows for rapid identification of the contribution of different types of behavior to reader interest in subsequent time window segmentation and behavior intensity calculation. The completed behavior time series dataset can be further segmented and statistically analyzed according to preset time windows, providing foundational data for generating behavior intensity indices, determining reading status, and calculating interest-oriented feature vectors. The data primarily originates from raw reader behavior datasets, including logs from the book service platform, mobile application operation records, and data collected by third-party plugins. Standardization and time-sequencing ensure that data from different behavior types and time periods can be analyzed and calculated within a unified timeframe. By organizing reader behavior chronologically, structured and continuous time-series data is created, guaranteeing that subsequent analysis steps accurately reflect trends in reader behavior and the evolution of their interests. This provides a reliable data foundation for dynamically updating book recommendation strategies and enhances the recommendation system's sensitivity and responsiveness to changes in reader interests.

[0015] S3: Perform segmented statistics on the behavior time series dataset within a preset time window, calculate the behavior frequency value and behavior duration value within each time period, generate a behavior intensity index sequence, and form a behavior sequence feature dataset.

[0016] Furthermore, step S3 includes: S31: Divide the behavior time series dataset into multiple time subsequence datasets according to a preset time window length; S32: For each time subsequence dataset, construct a behavior category hierarchical matrix according to behavior type identifiers. The behavior category hierarchical matrix is ​​used to characterize the distribution structure of different types of behavior within the current time window; S33: Based on the behavior category hierarchical matrix, calculate the behavior coupling coefficient between each behavior category. The behavior coupling coefficient is used to characterize the degree of correlation between different behaviors within the same time window, and generate behavior correlation feature data; S34: Based on the behavior correlation feature data, statistically analyze the behavior frequency vector and behavior duration vector within the current time window, and incorporate the behavior coupling coefficient into the calculation process of the behavior frequency vector to generate a coupling frequency vector; S35: Normalize the coupling frequency vector and behavior duration vector, and based on the time decay recursive function, recursively accumulate the calculation results of different time windows to generate the behavior intensity index of the current time window; S36: Arrange the behavior intensity indices generated by each time window in chronological order to form a behavior intensity index sequence, and construct a behavior sequence feature dataset by combining it with the behavior correlation feature data.

[0017] Specifically, the behavioral time series dataset is segmented and statistically analyzed within a preset time window. The frequency and duration of behaviors within each time period are calculated to generate a behavior intensity index sequence, forming a behavioral sequence feature dataset. The behavioral time series data obtained from the original reader behavior dataset is divided according to a preset time window length, generating multiple time sub-series datasets. Each time sub-series corresponds to a time window, the length of which can be set according to platform requirements, such as by hour, day, or fixed minutes, to capture changes in reader behavior patterns and reading rhythms across different time periods. During the segmentation process, key fields such as behavior type identifier, behavior timestamp, book identifier, behavior duration, and user identifier are retained to ensure the integrity and traceability of the time series. For each time sub-series dataset, a hierarchical behavior category matrix is ​​constructed based on the behavior type identifier. Behavior types include, but are not limited to, clicking, browsing, borrowing, reading, collecting, rating, sharing, and dwell time, with each type of behavior corresponding to a dimension in the matrix. The hierarchical behavior category matrix clearly represents the quantity, proportion, and distribution structure of various behaviors within the current time window, providing basic data for subsequent correlation analysis. Meanwhile, to consider the differences in the contribution of different behaviors to interest expression, preliminary weights can be assigned to different types of behaviors, so that the matrix reflects both the quantity and quality of behaviors. Based on the hierarchical matrix of behavior categories, the behavior coupling coefficient between each behavior category is calculated. The coupling coefficient is used to quantify the correlation strength between different behaviors within the same time window, such as the correlation between click and read behaviors, collection and evaluation behaviors, and sharing and dwell time behaviors. The calculation of the coupling coefficient can be achieved by statistically analyzing the frequency of simultaneous occurrence of behaviors, the time interval between behaviors, and co-occurrence patterns, thereby generating behavior association feature data. Behavior association feature data not only reflects the independent distribution of behaviors but also reveals the interaction and potential interest-driven relationships between different behaviors, providing more granular information support for personalized recommendations. The behaviors within the current time window are comprehensively quantified. The behavior frequency vector and behavior duration vector of each behavior type are statistically analyzed, and the behavior coupling coefficient is integrated into the behavior frequency calculation process to generate a coupling frequency vector. The coupling frequency vector can reflect the synergistic relationship between different behaviors, so that when calculating the behavior intensity, not only the quantity and duration of individual behaviors are considered, but also the mutual correlation and potential interest signals between behaviors are considered. The coupling frequency vector and behavior duration vector are normalized to eliminate the influence of differences in units and time windows on the index calculation. Then, a time-decay recursive function is introduced to recursively accumulate the calculation results for different time periods to generate the behavior intensity index for the current time window. The time-decay recursive function assigns higher weights to recent behaviors, making the system more sensitive to dynamic changes in reader interests, while ensuring that historical behaviors still have a reasonable impact on the index. The behavior intensity index comprehensively reflects the number of behaviors, duration, and coupling characteristics between behaviors, and is an important indicator describing the strength and activity level of reader interests.The behavioral intensity indices generated in each time window are arranged chronologically to form a complete behavioral intensity index sequence. This sequence is then combined with behavioral association feature data to construct a behavioral sequence feature dataset. This dataset not only contains the behavioral intensity of each time window but also records behavioral type coupling information, time decay weighting information, and historical accumulated data. This provides standardized, structured, and traceable input data for subsequent reading status determination, interest-oriented feature vector generation, and recommendation rule triggering. Data sources include original reader behavior datasets, platform log databases, mobile and web application records, and data collected by third-party reading plugins or extension modules. This approach accurately quantifies the intensity and association features of different types of behavior across different time periods, achieving a comprehensive characterization of reader behavior patterns and interest evolution trends. This provides a solid data foundation for personalized recommendations and dynamic feedback adjustments, improving the recommendation system's response speed and accuracy to changes in interest. By segmenting time, hierarchically classifying behavior, coupling computation, normalization, and time decay recursion, a structured, standardized behavior sequence feature dataset containing behavioral correlation information is formed. This enables the comprehensive quantification of reader behavior across multiple types and time periods, providing accurate data support for subsequent book recommendations based on behavioral feedback, thereby significantly improving the personalization and dynamic adjustment capabilities of the recommendation system.

[0018] S4: Based on the behavior sequence feature dataset, call the preset state determination function to perform interval mapping on the behavior intensity index sequence, generate reading state identifiers, construct a reading state transition matrix based on the reading state identifiers of adjacent time windows, and generate the current reading state dataset.

[0019] Furthermore, step S4 includes: S41: Input the behavior intensity index of different time periods in the behavior sequence feature dataset into a preset multi-level state determination function. The multi-level state determination function includes an interval mapping module, a rate of change analysis module, a behavior fusion module, and a confidence generation module, which are used to output the preliminary reading state identifier and the state confidence value; S42: Construct a state transition probability matrix based on the preliminary reading state identifiers of adjacent time windows. P(k) Each matrix element P(i,j) Representing reading state i To state j The probability of transition is calculated and stored as a time window transition matrix sequence, and the state transition history is recorded; S43: Combine the state confidence value with the state transition probability matrix. P(k)S44: Calculate a weighted state score to measure the stability and reliability of the reading state in the current time window; S45: When the weighted state score is lower than a preset threshold, correct the initial reading state, generate a corrected time window reading state identifier, and record the state correction log; S46: Arrange the corrected time window reading state identifiers in chronological order to form a current reading state identifier sequence, and output a state confidence sequence for interest-oriented feature vector generation and recommendation rule triggering; S47: Construct a final reading state transition matrix based on the corrected time window reading state identifiers, and generate a current reading state dataset, which includes the corrected reading state sequence, state transition matrix, and state confidence information, providing complete input data for subsequent recommendations and dynamic adjustments.

[0020] Furthermore, step S41 includes: S41-1: The interval mapping module divides the behavior intensity index within the current time window into multi-level intervals, and maps the behavior intensity index to a predefined multi-level reading state set S = according to preset interval mapping rules. {Browsing state, shallow reading state, deep reading state, interest decay state}, generating candidate reading state identifiers; S41-2: The change rate analysis module calculates the change rate of behavior intensity within the current time window, extracts the behavior intensity fluctuation trend, and generates a state stability index to characterize the persistence and fluctuation characteristics of the reading state; S41-3: The behavior fusion module weights and combines the behavior intensity indices corresponding to different behavior types within the same time window to form a comprehensive behavior intensity index; the weighting coefficient is preset according to the influence weight of the behavior type or dynamically adjusted based on historical behavior performance; S41-4: The confidence generation module combines the comprehensive behavior intensity index, the state stability index, and the reading state identifier of the previous time window to perform a consistency evaluation of the current candidate reading state and generate a state confidence value; S41-5: The multi-level state determination function outputs the preliminary reading state identifier and the corresponding state confidence value of the current time window based on the comprehensive behavior intensity index and the interval mapping result.

[0021] Furthermore, step S42 includes: S42-1: Number and encode the multi-level reading state set to establish a reading state index set; S42-2: Generate single-step state transition records based on the preliminary reading state identifiers of adjacent time windows, and construct a single-step state transition indicator matrix; S42-3: Combine the state confidence values ​​to perform confidence weighting on the single-step state transition indicator matrix to generate a weighted state transition matrix, which is used to reduce the impact of low-confidence states on the transition probability calculation; S42-4: Recursively accumulate the weighted state transition matrix of the current time window with the historical cumulative transition matrix, and introduce a time decay factor to attenuate and update the historical transition intensity to generate a cumulative state transition intensity matrix. A(k) The update method satisfies the formula: A(k)=λ· A(k - 1) + ( k ) ;in, A(k) For the first k The cumulative state transition intensity matrix of the time window, where A(k-1) is the cumulative state transition intensity matrix of the previous time window. ( k S42-5: The weighted state transition matrix for the current time window, where λ is the time decay factor, ranging from 0 to 1, used to control the influence of historical state transitions on the current cumulative result; S42-5: The cumulative state transition intensity matrix A(k) After normalization, the state transition probability matrix of the current time window is obtained. P(k) ; where matrix elements P(i,j) This indicates the reading status within the current time window. i Towards reading state j The conditional probability of the transition, which is determined by the cumulative state transition strength matrix. A(k) The ratio of the corresponding element in the middle to the sum of the elements in that row is determined for any reading state. i The sum of all corresponding transition probabilities satisfies the preset normalization condition; S42-6: The state transition probability matrix generated in each time window P(k) The states are stored in chronological order as a sequence of state transition matrices, and a state transition history log is recorded synchronously. The state transition history log includes the reading state identifiers of adjacent time windows and the corresponding state confidence values, which are used for subsequent state stability analysis and dynamic correction of recommended strategies.

[0022] Specifically, based on the behavior sequence feature dataset, a preset state determination function is invoked to perform interval mapping on the behavior intensity index sequence, generating reading state identifiers. A reading state transition matrix is ​​then constructed based on the reading state identifiers of adjacent time windows, generating the current reading state dataset. The behavior sequence feature dataset originates from the behavior intensity index sequence and behavior-related feature data generated in step S3. The data includes structured fields such as time window number, behavior intensity index value, behavior coupling feature vector, behavior persistence index, and time decay weight value. This data is collected and structured from the book service platform log system, mobile reading client, web page operation records, and third-party reading plugin modules, exhibiting temporal continuity and behavioral correlation. The aim is to convert the continuous numerical form of the behavior intensity index into discrete reading state identifiers and to construct a state transition probability matrix through statistical analysis of the temporal evolution relationship between states, thereby forming a reading state dataset capable of depicting the changing trends of reader interest. The multi-level state determination function includes an interval mapping module, a rate of change analysis module, a behavior fusion module, and a confidence generation module. Each module is executed sequentially according to a preset logical order, forming a complete state determination process. The input data consists of the behavior intensity index and related feature data for the current time window, and the output data consists of the preliminary reading state identifier and the corresponding state confidence value. To ensure consistency and interpretability in the judgment process, a multi-level reading state set is predefined: S = {Browsing state, shallow reading state, deep reading state, interest decay state} where: Browsing state: low behavior intensity, mainly consisting of clicks and short dwell times; Shallow reading state: continuous reading behavior exists, but the dwell time is relatively short; Deep reading state: long reading duration and frequent interaction; Interest decay state: behavior intensity decreases significantly or fluctuates drastically. The interval mapping module divides the behavior intensity index within the current time window into multi-level intervals. The interval 0 – first threshold corresponds to the browsing state; the interval from the first threshold to the second threshold corresponds to the shallow reading state; the interval from the second threshold to the third threshold corresponds to the deep reading state; when the behavior intensity index is higher than the third threshold but shows a significant downward trend, it is judged as the interest decay state. The interval mapping module outputs candidate reading state identifiers, providing a basis for subsequent consistency verification. For example, the following thresholds can be set: 0–0.25 → browsing state, 0.25–0.55 → light reading state, 0.55–0.85 → deep reading state, and ≥0.85 with a continuous downward trend → interest decay state. When the behavior intensity index in a certain time window is 0.62, it is initially determined to be in a "deep reading state" according to the interval mapping rules, and a candidate reading state identifier is generated. The interval threshold can be determined based on historical sample data or preset according to the platform's statistical rules. The rate of change analysis module is used to calculate the rate of change of the behavior intensity index in the current time window relative to the behavior intensity index in the previous time window. The rate of change is used to reflect the strengthening or weakening trend of reader interest.By comparing the differences and directions of change in the behavior intensity index between adjacent time windows, the fluctuation trend of behavior intensity is extracted to generate a state stability index. If the behavior intensity index continues to rise and the change range is stable, the state stability is high; if the behavior intensity index fluctuates drastically or drops sharply, the state stability is low. The state stability index is used for subsequent confidence calculations to avoid misjudgment of the state due to short-term abnormal behavior. For example, if the behavior intensity rises from 0.60 to 0.78, the rate of change is positive, and the state stability is enhanced; if the behavior intensity drops from 0.75 to 0.40, the rate of change is negative, and interest decay may be entering a period of decline. The "state stability index" is generated based on the change range to measure whether the current reading state is continuously stable. If the behavior intensity fluctuates drastically within a certain time window, even if the interval is mapped to "deep reading," its stability may be low. The behavior fusion module performs weighted combination processing on the behavior intensity indices corresponding to different types of behavior within the current time window. Behavior types include click behavior, browsing behavior, reading pause behavior, page turning behavior, collection behavior, evaluation behavior, and sharing behavior, etc. Each behavior type corresponds to a different level of interest expression, therefore, corresponding weights need to be set during the fusion process. These weights can be determined using a preset method or dynamically adjusted based on the contribution of historical behaviors to the recommendation effect. For example, the weight of reading dwell behavior can be higher than that of click behavior, and the weight of collection behavior can be higher than that of browsing behavior. A comprehensive behavior intensity index is generated through weighted fusion to correct the interval mapping results, making the state judgment more consistent with the actual reading behavior characteristics. For example: reading dwell weight = 0.4, collection weight = 0.2, rating weight = 0.15, sharing weight = 0.15, and click weight = 0.1. If, in a certain time window, reading dwell intensity = 0.8, collection intensity = 0.6, and sharing intensity = 0.5, then a comprehensive behavior intensity index is formed after weighted fusion to correct the interval mapping results. The confidence generation module integrates the following factors: comprehensive behavior intensity index, state stability index, reading state in the previous time window, and behavior coupling feature data. A consistency evaluation mechanism is used to generate the state confidence value for the current time window. If the current state is consistent with the previous state and has high stability, the confidence value is high; if the state changes frequently or the behavior intensity fluctuates drastically, the confidence value decreases accordingly. For example, if the current interval is determined to be "deep reading," and the previous window was also "deep reading," and the rate of change is stable, the confidence value is high. If the current interval is determined to be "shallow reading," but the previous window was "deep reading" and the decrease was significant, the confidence value is moderately reduced. The confidence value can be a value in the range of 0–1 for subsequent weighted processing. Finally, the multi-level state determination function outputs: the preliminary reading state identifier for the current time window and the corresponding state confidence value. For example: Time window T3: Behavior intensity index = 0.63, state identifier = deep reading state, state confidence = 0.82.To depict the dynamic evolution of reading states over time, a state transition probability matrix is ​​constructed based on the reading state identifiers of adjacent time windows. First, the multi-level reading state set is numbered and encoded to establish a reading state index set. For example: browsing state is encoded as 1; shallow reading state as 2; deep reading state as 3; and interest decay state as 4. This encoding method facilitates matrix calculation and probability statistics. Based on the preliminary reading state identifiers of adjacent time windows, single-step state transition records are generated. For example, if a user's state in three consecutive time windows is: browsing state → shallow reading state → deep reading state, then two single-step transition records are generated: a transition from browsing state to shallow reading state; and a transition from shallow reading state to deep reading state. A single-step state transition indicator matrix is ​​constructed based on these records. To reduce the interference of low-confidence states on the overall transition statistics, the single-step state transition indicator matrix is ​​weighted by combining the corresponding state confidence values ​​to generate a weighted state transition matrix. State transition records with higher confidence contribute more to the matrix, while records with lower confidence contribute less. To incorporate historical information and maintain sensitivity to recent state changes, a recursive update mechanism is used to construct a cumulative state transition strength matrix. A(k) Its update method satisfies: A(k)=λ· A(k - 1)+ (k) ;in, A(k) For the first k The cumulative state transition intensity matrix within the time window. A(k - 1) This is the cumulative state transition intensity matrix for the previous time window. ( k () is the weighted state transition matrix for the current time window. λ The time decay factor, λ The value ranges from 0 to 1, and is used to control the degree of influence of historical state transitions on the current cumulative result; when λ A larger value indicates a higher degree of retention of historical state transition information; when... λ A smaller value indicates that the system pays more attention to state changes within the current time window. To ensure this parameter is achievable and repeatable, λ The acquisition method can be determined through one or more of the following methods. In the initial stage of system deployment, λ can be set empirically based on the user behavior characteristics and reading scenario patterns of the book service platform. For example, when the platform's readers' reading behavior is relatively stable and their interest change cycle is long, λ can be set to a fixed value within the range of 0.7 to 0.9; when the platform's readers' interests fluctuate frequently and their behavior changes rapidly, λ can be set within the range of 0.4 to 0.6. Those skilled in the art can directly select appropriate values ​​according to the actual business scenario. The cumulative state transition intensity matrix... A(k)After normalization, the state transition probability matrix of the current time window is obtained. P(k) ; where matrix elements P(i,j) This indicates the reading status within the current time window. i Towards reading state j The conditional probability of the transition, which is determined by the cumulative state transition strength matrix. A(k) The ratio of the corresponding element in the middle to the sum of the elements in that row is determined for any reading state. i The sum of all corresponding transition probabilities satisfies the preset normalization condition; the state transition probability matrix generated for each time window... P(k) The data is stored chronologically as a sequence of state transition matrices, and a state transition history log is recorded synchronously. This history log includes reading state identifiers for adjacent time windows and their corresponding state confidence values, used for subsequent state stability analysis and dynamic correction of recommendation strategies. After the above processing, a current reading state dataset is generated. This dataset includes: a corrected reading state sequence; a state confidence sequence; a state transition probability matrix sequence; a cumulative state transition strength matrix; and a state transition history log. This dataset is used for subsequent interest-oriented feature vector generation, recommendation matching value calculation, and recommendation rule triggering. Compared to existing solutions that judge reading states based solely on a single action, this solution, through a multi-level state determination function and a state transition probability matrix construction mechanism, achieves continuous characterization and trend analysis of the dynamic changes in readers' reading interests. This improves the accuracy and stability of reading state identification, enhances the recommendation system's responsiveness and adaptive adjustment capabilities to changes in interests, and thus significantly improves the personalization and matching accuracy of book recommendations.

[0023] S5: Based on the current reading state dataset and reading state transition matrix, and in conjunction with the behavior time series dataset, call the book feature association mapping table to extract the book topic vector corresponding to the current reading state. Based on the reading state transition matrix, perform recursive cumulative calculation on the book topic vector to generate interest-oriented feature vectors and form a reader interest feature dataset.

[0024] Furthermore, step S5 includes: S51: Extract the current time window reading state identifier from the current reading state dataset, and calculate the future state trend distribution of the current reading state based on the reading state transition matrix to characterize the reader's state evolution trend; S52: Call the book feature association mapping table, which establishes a mapping relationship between reading state and book topic, mapping the current reading state identifier to the corresponding book topic vector; wherein, each reading state corresponds to multiple topic dimensions, and the book topic vector is used to represent the content interest direction of the current reading state; S53: Based on the state transition probability matrix P(k)Based on the transition probability distribution corresponding to the current state, the book topic vector is subjected to state transition weighted expansion processing to generate a state-driven topic expansion vector. T′k S54: Used to represent the direction of potential interest migration; S54: Drives the topic expansion vector based on the state of the current time window. T′k Historical interest-oriented feature vector H(k - 1) Perform recursive cumulative updates to generate the interest-oriented feature vector for the current time window. H(k) The formula is: H(k) = α×H(k - 1) + β×T′k ;in, H(k) The feature vector pointing to the interest of the current time window. H(k - 1) The feature vector pointing to the interest of the previous time window. T′k For state-driven topic extension vectors, α This represents the historical interest decay coefficient. β This represents the current theme reinforcement coefficient. α、β All are dynamically adjustable parameters, and are adaptively adjusted according to the corresponding transition probabilities in the current reading state transition matrix, so that the deep reading state corresponds to... β The value is greater than the browsing status corresponding to β Values; where the driving topic expansion vector T′k The degree of contribution is determined by the state transition probability distribution, and historical interest points to the feature vector. H(k - 1) The influence gradually diminishes over time; S55: The interest-oriented feature vectors generated in each time window are... H(k) Construct a dataset of reader interest features in chronological order.

[0025] Specifically, based on the current reading state dataset and reading state transition matrix, and combined with the behavior time series dataset, the book feature association mapping table is invoked to extract the book topic vector corresponding to the current reading state. Based on the reading state transition matrix, the book topic vector is recursively accumulated to generate interest-oriented feature vectors, thus forming a reader interest feature dataset. The current reading state dataset originates from the processing result of step S4 and includes at least the following structured fields: time window number, corrected reading state identifier, state confidence value, and state transition probability matrix. P(k) Cumulative state transition strength matrix A(k)And state transition history. The behavior time series dataset comes from the sorting processing result of step S2, and includes behavior timestamps, behavior type identifiers, book identifiers, behavior durations, and corresponding time window indices. The above data are all collected by the book service platform server log system, mobile reading client, and web reading system, and generated after structured processing, time segmentation, and state determination, possessing temporal continuity and state correlation. The discretized reading states are further transformed into continuous interest feature vectors that can be used for recommendation matching calculations to characterize the reader's content interest direction and its evolution trend under different time windows. First, the reading state identifier of the current time window is extracted from the current reading state dataset. For example, in the first... k Within a given time window, the reader is identified as being in a "deep reading state." Subsequently, based on the state transition probability matrix corresponding to the current time window... P(k) Calculate the future state tendency distribution of the current state. The state transition probability matrix... P(k) matrix elements P(i,j) Indicates reading status i Towards reading state j The conditional probability of the transition. Therefore, if the current state is state... i Then the matrix of the first iThe probability distribution corresponding to each row represents the probability of transitioning to each state in the next time window. By extracting this probability distribution, a future state trend distribution vector can be obtained to characterize the evolution trend of the reader's reading state. For example, when the current state is "shallow reading state", its corresponding state transition probability distribution is: 0.5 for transitioning to shallow reading state; 0.3 for transitioning to deep reading state; 0.1 for transitioning to browsing state; and 0.1 for transitioning to interest decay state. A pre-built book feature association mapping table is invoked to map the current reading state to the corresponding book topic vector. The book feature association mapping table is stored in a book resource database, and its data sources include: book metadata (book classification, tags, keywords, author domain, etc.); book content semantic analysis results (topic feature vectors calculated based on text word frequency statistics and feature mapping); and historical user behavior statistics (topic preference distribution results under different reading states based on behavior frequency statistics). The mapping table establishes a correspondence between "reading state - topic dimension". For example: the browsing state corresponds to the theme dimensions biased towards "popular recommendations," "short content," and "light reading"; the shallow reading state corresponds to the theme dimensions biased towards "introductory," "general knowledge," and "bestsellers"; the deep reading state corresponds to the theme dimensions biased towards "professional," "academic," and "long series"; and the interest decay state corresponds to the theme dimensions biased towards "diverse exploration" or "related recommendations." Each reading state corresponds to multiple theme dimensions, forming a multi-dimensional theme vector. For example, a theme vector can be defined as an N-dimensional vector, where each dimension corresponds to a theme category, and the value represents the interest weight. When the current time window state is "deep reading state," the corresponding theme vector is extracted through a mapping table. This is based on the state transition probability matrix. P(k) Based on the transition probability distribution corresponding to the current state, the book topic vector is subjected to state transition weighted expansion processing to generate a state-driven topic expansion vector. T′k This is used to represent the potential direction of interest migration; it considers not only the topic vector corresponding to the current state, but also the topic vectors corresponding to possible future transition states, and combines them according to the transition probability. For example, if the current state is "shallow reading state", its future transition probability is: 0.3 to deep reading state; 0.5 to shallow reading state; 0.1 to browsing state; 0.1 to interest decay state. Then, the state-driven topic expansion vector... T′k The interest vector is formed by weighting and fusing the current topic vector with the aforementioned future state topic vectors according to their corresponding probabilities. This process ensures that the interest vector not only reflects the current interest direction but also embodies potential interest migration trends. The current time window's state drives the topic expansion vector. T′k Historical interest-oriented feature vector H(k - 1) Perform recursive cumulative updates to generate the interest-oriented feature vector for the current time window. H(k)The formula is: H(k) = α×H(k - 1) + β×T′k ;in, H(k) The feature vector pointing to the interest of the current time window. H(k - 1) The feature vector pointing to the interest of the previous time window. T′k For state-driven topic extension vectors, α This represents the historical interest decay coefficient. β This represents the current theme reinforcement coefficient. α、β All are dynamically adjustable parameters, and are adaptively adjusted according to the corresponding transition probabilities in the current reading state transition matrix, so that the deep reading state corresponds to... β The value is greater than the browsing status corresponding to β Value; wherein, the contribution of the current topic vector is determined by the reading state transition probability, and the influence of historical interest features decays over time; in this embodiment, the historical interest decay coefficient α and the current theme reinforcement coefficient β All parameters are set as dynamically adjustable parameters that change with the time window, and their values ​​are adaptively determined based on the reading state transition matrix of the current time window. Within each time window, the state transition probability distribution corresponding to the current reading state is extracted from the reading state transition matrix. Based on the self-circulation probability of this state and the transition probability structure to other states, a stability index for the current reading state is calculated. This stability index characterizes the reader's interest persistence and behavioral intensity in the current state. When the self-circulation probability corresponding to the current reading state is high, the state is considered to have high stability, indicating that the reader's interest expression is clear and persistent. In this case, the value of the current topic reinforcement coefficient β is increased, so that the state-driven topic expansion vector generated in the current time window occupies a larger weight in the interest-oriented feature vector update process; at the same time, the historical interest decay coefficient α is correspondingly reduced, so that the retention ratio of historical interest features decreases, thereby enhancing the influence of the current reading behavior on the interest feature structure. When the probability of self-looping corresponding to the current reading state is low and the probability of transitioning to other states is high, the stability of this state is considered weak, indicating that the expression of interest is volatile or sporadic. In this case, the value of the current topic enhancement coefficient β is reduced to weaken the influence of the current topic vector on the overall interest feature structure; simultaneously, the historical interest decay coefficient α is increased to maintain a high proportion of historical interest features during the update process, thus avoiding excessive disturbance to interest features caused by short-term browsing behavior. At the state type level, different reading states have preset enhancement priorities, ensuring that the enhancement priorities corresponding to deep reading states are optimized. β The overall value range is higher than that corresponding to the browsing state. βThe value range is determined to ensure that the contribution of deep reading behavior to the interest update process is greater than that of browsing behavior. The priority setting can be configured based on historical behavior statistics or empirical rules, and dynamically fine-tuned in conjunction with state transition probabilities during actual operation. Through a dynamic weight adjustment mechanism, the contribution of the current topic vector is determined by the reading state transition probability, and the influence of historical interest features gradually decays over time, achieving an adaptive balance between long-term stable interests and short-term fluctuating interests, thus improving the accuracy and stability of characterizing changes in reader interests. First, the interest vector for each time window is obtained. H(k) This vector is driven by the state of the current time window, forming a topic expansion vector. T′k Interest vectors from the previous time window H(k - 1) It is obtained by updating according to the recursive rule, where each H(k) It includes multiple book theme dimensions to characterize the distribution of readers' interest intensity in each theme within this time window. H(k) The generation depends on the behavioral time series dataset, the current reading state dataset, and the state transition probability matrix. P(k) The former provides reader behavior records and their timestamps, including browsing, clicking, favorites, borrowing, and commenting. The current reading state dataset provides the reading state and corresponding confidence level of each time window output by a multi-level state determination function. The reading state transition matrix reflects the transition probability between reading states in each time window, providing a basis for the cumulative update and historical decay of interest vectors. The interest vectors of all time windows are arranged in chronological order to form a sequence { H(1) , H(2) , …, H(k)},in kThe total number of time windows can be defined and annotated with timestamp information to ensure temporal continuity and the traceability of interest evolution. To enhance the operability and stability of the dataset, vectors in the sequence can be standardized or normalized to unify the values ​​of each topic dimension. Simultaneously, state confidence can be used to weight interest intensity, reducing the impact of short-term behavioral fluctuations on overall interest characteristics. The completed dataset can be stored in the form of relational database tables, time series data files, or in-memory vector caches. Each record contains a time window identifier, the value of each topic interest, the corresponding reading state, and state confidence information, thereby achieving structured management and dynamic tracking of interest features. For example, if a day is divided into 24 time windows, and each interest vector contains 10 topic dimensions, the final dataset contains 24 ordered vector records, each reflecting the reader's interest intensity and stability indicators for 10 topics within the corresponding time window. Through the above processing, dynamic behavioral data can be effectively transformed into structured time-series interest feature data, providing a complete and operable input foundation for personalized recommendation strategies, recommendation rule triggering, and long-term interest evolution analysis. At the same time, it ensures the recursive accumulation and historical decay processing of interest features over time, making the recommendation results more accurate, stable, and consistent with the actual interest change trends of readers.

[0026] S6: Input the reader interest feature dataset and the current reading status identifier into the recommendation rule trigger function, calculate the recommendation matching value, and generate the recommendation matching dataset.

[0027] Furthermore, step S6 includes: S61: Construct a set of candidate book theme feature vectors, wherein the candidate book theme feature vectors are extracted from the theme feature data of each book in the book resource database, and are compared with the interest orientation feature vector. H(k) Perform dimension alignment processing; S62: Based on the current reading state In the state transition probability matrix P(k) Calculate the state adjustment factor from the transition probability distribution of the corresponding row. The calculation formula is: ;in, To the current reading state j Given the transition probabilities of each state, the maximum transition probability in that row is taken as the state stability. The state-sensitive adjustment coefficient enhances the recommendation strength when the state transition probability concentration is high, and reduces the strength of the enhancement and enhances the diversity of recommendations when the state transition probability distribution is relatively dispersed; S63: Define the recommendation rule trigger function, which is used to calculate the book matching degree based on the current reading state identifier and interest orientation feature vector. Its calculation process includes a similarity calculation step and a state adjustment step. The similarity calculation formula is: ;in, For the first The recommended match value for this candidate book. The feature vector pointing to the interest of the current time window. For the first Thematic feature vector of this candidate book; S64: Recommended matching value based on candidate books A recommendation matching dataset is generated, which includes book identifiers, matching values, and state adjustment factor information; S65: The recommendation matching dataset is sorted, and candidate books are arranged from high to low according to the matching value. When the recommendation matching value meets the preset trigger threshold condition, the corresponding book is output as the recommendation result.

[0028] Specifically, the constructed reader interest feature dataset and the reading status identifier of the current time window are input into the recommendation rule trigger function to achieve personalized recommendations of candidate books. First, a set of topic feature vectors for candidate books is extracted based on the book information in the book resource library. Each candidate book's topic feature vector includes multiple topic dimensions, covering content categories such as literature, science and technology, history, art, and children's books. This vector can be obtained through book category tags, content analysis, keyword extraction, or book metadata mapping. Then, the topic feature vectors of the candidate books are compared with the interest-oriented feature vectors of the current time window. Dimension alignment is performed to ensure that the topic dimensions of both are consistent, thus facilitating subsequent similarity calculations. (Based on current reading status) In the state transition probability matrix P(k) The state adjustment factor is calculated from the transition probability distribution of the corresponding row, using the following formula: : ;in, To the current reading state j Given the transition probabilities of each state, the maximum transition probability in that row is taken as the state stability. The state-sensitive adjustment coefficient enhances the recommendation strength when the state transition probability concentration is high, and reduces the strength of the enhancement and enhances the diversity of recommendations when the state transition probability distribution is relatively dispersed; S63: Define the recommendation rule trigger function, which is used to calculate the book matching degree based on the current reading state identifier and interest orientation feature vector. Its calculation process includes a similarity calculation step and a state adjustment step. The similarity calculation formula is: ;in, For the first The recommended match value for this candidate book. The feature vector pointing to the interest of the current time window. For the first Thematic feature vector of this candidate book This is a state adjustment factor, adaptively determined based on the current reading state identifier and the transition probability distribution of the corresponding state in the reading state transition matrix; when the concentration of transition probabilities is high, A larger value is used to enhance the strength of the recommendation, prioritizing books with higher matching degrees; when the transition probability distribution is relatively dispersed, A lower value is used to reduce the reinforcement intensity, thereby enhancing the diversity of recommendations and ensuring that readers can access different categories of books even when their interest fluctuates. This recommendation matching dataset and ranking information can be stored and recorded for subsequent recommendation strategy optimization, historical behavior analysis, and interest evolution analysis. The data sources involved in the entire process include: user behavior datasets on the book service platform, and interest-oriented feature vectors generated through this process. Current time window reading status indicator Thematic feature data of each book in the book resource database, and the state transition probability matrix. P(k) The processing includes data extraction, feature alignment, state adjustment factor calculation, similarity calculation, matching value generation, sorting, and threshold determination. Through the above embodiments, it is possible to accurately characterize changes in reader interests across different time windows, making the recommendation results more consistent with the dynamic changes in reader interests, while also considering the personalization and diversity of recommendations, ensuring the stability and effectiveness of the system's recommendations.

[0029] S7: When the recommended matching value meets the preset trigger threshold condition, a candidate book set is generated from the book resource library based on the recommended matching dataset, and the recommended results are output according to the recommended matching value.

[0030] Furthermore, step S7 includes: S71. Detect the recommendation matching value of each candidate book in the recommendation matching dataset. S71. Determine whether the preset trigger threshold condition is met. When the recommended matching value is greater than or equal to the preset trigger threshold, add the corresponding book identifier to the candidate filtering queue; S72. Based on the book identifier added to the candidate filtering queue, retrieve the corresponding book resource information from the book resource library. The book resource information includes book theme feature data, category tag data, and resource availability information to generate a candidate book set; S73. Sort the books in the candidate book set according to the recommended matching value from high to low. When there are cases where the recommended matching values ​​are the same, perform a secondary sorting according to the priority adjustment of the current reading status identifier; S74. Control the number of books in the sorted candidate book set. Extract the top N books according to the preset output quantity limit to form a recommendation result set. The output quantity limit is dynamically adjusted according to the current reading status identifier or system resource load status; S75. Output the recommendation result set to the recommendation display module and record the recommendation output log for subsequent status correction.

[0031] Specifically, when the recommended matching value of the candidate book When a preset trigger threshold is met, candidate book filtering, sorting, and output operations are performed based on the recommendation matching dataset to generate the final personalized recommendation result. First, the matching value of each candidate book in the recommendation matching dataset is detected. Compared with the preset trigger threshold, when When the value is greater than or equal to the threshold, the corresponding book identifier is added to the candidate filtering queue to ensure that only books highly matching the current reader's interests and status enter the candidate set. After candidate filtering is completed, the corresponding book resource information is retrieved from the book resource library. This book resource information includes the book's subject feature data, category tag data, and resource availability information, such as the book's inventory status, digital version accessibility, and borrowing permissions. By integrating this information, a candidate book set is generated, providing a complete data foundation for subsequent sorting and output. Subsequently, the books in the candidate book set are sorted according to the recommended matching value. The books are sorted from highest to lowest match value to ensure that the books with the highest match value are displayed first. When multiple books have the same match value, they are further sorted according to the reading status priority corresponding to the current time window. For example, books in the deep reading state may be prioritized over books in the shallow reading state, thus realizing dynamic adjustment of the recommendation results based on the interest status. After sorting, the system controls the number of candidate books in the set, selecting the top N books according to the preset output limit to form the recommendation result set. The output limit can be dynamically adjusted according to the current reading status, reader activity, system resource load, etc., to ensure the rationality of the recommendation results and the system response performance. Finally, the generated recommendation result set is output to the recommendation display module to realize the visual recommendation display for readers. At the same time, the system records the recommendation output log, including the recommendation match value, candidate book identifier, sorting result, and status adjustment factor information, to provide data support for subsequent dynamic optimization of the recommendation algorithm, status correction, and interest feature updates. The data types involved in the entire process include: reader interest feature datasets, current time window reading status identifiers, recommendation matching datasets, book resource library information, and status adjustment priority tables; data sources include reader behavior records, book metadata, category tags, and system-set threshold parameters; data processing includes operations such as filtering, feature comparison, sorting, quantity control, and log recording. It can achieve accurate filtering of highly matched books, ensuring that the recommendation results not only conform to reader interests and preferences but also take into account dynamic status adjustment and system resource constraints, while providing complete data support and traceability for subsequent recommendation strategy optimization.

[0032] S8: After the recommendation results are output, collect the reader's subsequent behavior data on the candidate book set, generate a feedback behavior dataset, and merge the feedback behavior dataset into the reader's original behavior dataset.

[0033] Specifically, after the recommendation result set is output to the recommendation display module, the subsequent behaviors of readers on the candidate book set within the specified time window are continuously collected in real time or periodically to generate a feedback behavior dataset. This feedback behavior data includes, but is not limited to, reader clicks on recommended books, browsing duration, adding to favorites or book lists, borrowing, downloading or reading digital content, and interactive behaviors such as ratings or comments. Each behavior record includes a behavior type identifier, behavior timestamp, corresponding book identifier, behavior duration, and behavior intensity value to accurately quantify the reader's interest response. The data sources mainly include front-end interaction logs from the book service platform, mobile client operation records, library or online reading system borrowing databases, and third-party content access records. More refined user behavior characteristics can also be obtained by combining device-side sensor data (such as reader dwell time, scrolling behavior, page-turning speed, etc.). During data processing, the collected feedback behavior data is first structured, parsing the original operation events into a unified data format, removing duplicate or abnormal data, and supplementing with behavioral context information, such as the time window period, user activity tags, and recommendation strategy version number, forming standardized feedback behavior record units. Subsequently, each record is sorted chronologically and merged with the existing original reader behavior dataset to update the completeness and continuity of the reader behavior data, thereby generating a comprehensive behavior dataset containing historical behavior and the latest feedback. The merging process includes field matching, unique identifier verification, and behavior time-series adjustment to ensure consistency between historical data and new feedback data in the time series, while retaining the source information and behavior type label of each record for accurate retrieval during subsequent state determination, interest vector updates, and recommendation rule triggering. This embodiment enables closed-loop collection of reader interest responses, ensuring that feedback generated by each recommendation behavior is reflected in the reader behavior dataset in a timely and complete manner. This provides data support for subsequent reading state determination, multi-level state transition matrix updates, interest-oriented feature vector generation, and adaptive optimization of recommendation strategies, achieving dynamic adjustment and refined optimization of personalized recommendations.

[0034] S9: Based on the consistency comparison between the feedback behavior dataset and the current reading state dataset, calculate the state deviation value. When the state deviation value exceeds the preset correction threshold, update the parameter set of the state determination function and the recommendation rule triggering function.

[0035] Specifically, after obtaining the reader feedback behavior dataset on the recommendation results, the dataset is compared with the reading state dataset for the current time window to assess the deviation between the reader's actual behavior and the system's predicted reading state. This consistency comparison first extracts information such as behavior type identifier, behavior duration, behavior intensity, and behavior occurrence time from each behavior record in the feedback behavior dataset. Simultaneously, it structures the time window reading state identifier, state confidence score, and historical state transition matrix in the current reading state dataset to ensure comparability between the two types of data in both the time and behavior dimensions. The comparison process includes matching the expected reading state with the actual behavior within the same time window. For example, if the system determines the reader is in a "deep reading state" within the current time window, and the feedback behavior shows that the reader only performed a short browsing session or quick jump, it is considered a state deviation. By setting a behavior weight matrix, the contribution of different types of behavior deviation to the state deviation value is weighted and calculated to generate a comprehensive state deviation value. The state deviation value reflects the degree of consistency between the current predicted state and the actual behavior; a larger value indicates a more significant deviation. Regarding data sources, deviation calculation relies on the aforementioned feedback behavior dataset, including various behavior types such as clicks, browsing, favorites, borrowing, ratings, comments, page turning, and dwell time, combined with time window identifiers and state confidence information from the current reading state dataset. Furthermore, user tag information, recommendation strategy version numbers, and historical interest feature data can be used as auxiliary inputs to improve the accuracy of deviation calculation. In the data processing, feedback behavior data is first aggregated according to time windows and matched with the reading state identifiers of the corresponding time windows. The deviation contribution value for each behavior is calculated based on behavior type and intensity. Subsequently, the deviation contribution values ​​of each behavior are weighted and accumulated according to a preset weighting rule to generate the state deviation value for the current time window. When the calculated state deviation value exceeds a preset correction threshold, an adaptive adjustment mechanism is triggered to dynamically update the parameter sets of the state determination function and the recommendation rule triggering function. The parameters of the state determination function include behavior intensity interval division rules, change rate analysis weights, behavior fusion coefficients, and confidence generation strategies; the parameters of the recommendation rule triggering function include state adjustment factors, similarity calculation weights, and recommendation thresholds. Parameter updates employ an iterative correction method based on feedback behavior. For example, weights corresponding to behaviors with significant deviations are increased or interval boundaries are adjusted to enhance the system's sensitivity to changes in readers' actual interests. Simultaneously, this ensures the recommendation strategy can adaptively adjust when deviations occur, thereby improving the accuracy and personalization of the recommendation results. This embodiment enables dynamic optimization of the recommendation loop, allowing feedback data after each recommendation action to effectively guide the next round of state determination and recommendation matching, ensuring continuous and refined adjustments to reader interest characteristics and recommendation strategies.All functional modules described in this application are implemented based on explicit statistical rules, probability calculations, and vector operations, without relying on neural network structures or training algorithms. All parameters are determined based on statistical distributions or preset interval rules.

[0036] In summary, the book recommendation method based on reader behavior feedback provided in this application has the following technical effects: This application constructs a multi-level reading state determination and state transition mechanism based on reader behavior sequence features. By combining the reading state transition matrix with a book feature association mapping table, it achieves recursive cumulative calculation of interest-oriented feature vectors. This solves the problem in existing technologies that rely solely on static interest tags or single-time behavior data, failing to dynamically capture changes in reader interests, resulting in insufficient personalization and low matching accuracy in recommendation results. This application performs time-series analysis on reader behavior data and feedback behavior data, mapping the intensity of behavior at different times to multi-level reading states, generating a current reading state dataset. Based on state transition probabilities, it weights and accumulates book topic vectors to form interest-oriented feature vectors reflecting potential interest migration directions, thereby achieving dynamic calculation of recommendation matching values ​​and screening and ranking of candidate book sets. Simultaneously, this application collects and analyzes user feedback behavior on recommendation results, calculates state deviation values, and dynamically adjusts the parameters of state determination and recommendation rules based on deviation, achieving closed-loop optimization of the recommendation strategy. Compared to existing technologies, this application significantly improves the personalization accuracy and real-time response capability of book recommendations. It also achieves continuous updating of interest features, dynamic adjustment of recommendation matching values, and enhanced consistency between recommendation results and users' actual interests, thereby improving the system's adaptability to changes in reader behavior and the diversity of interests. The above description of the disclosed embodiments enables those skilled in the art to implement or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A book recommendation method based on reader behavior feedback, characterized in that, include: S1. Obtain the reader's behavior data in the book service platform, perform structured processing on the behavior data, generate a set of behavior record units containing behavior type identifier, behavior timestamp, book identifier and behavior duration, and store it as the reader's original behavior dataset; S2. Sort the original reader behavior dataset by behavior timestamp to construct a behavior time series dataset; S3. Within a preset time window, the behavior time series dataset is segmented and statistically analyzed to calculate the behavior frequency and duration values ​​in each time period, generate a behavior intensity index sequence, and form a behavior sequence feature dataset. S4. Based on the behavior sequence feature dataset, call the preset state determination function to perform interval mapping on the behavior intensity index sequence, generate reading state identifiers, construct a reading state transition matrix based on the reading state identifiers of adjacent time windows, and generate the current reading state dataset. S5. Based on the current reading state dataset and reading state transition matrix, and in conjunction with the behavior time series dataset, call the book feature association mapping table to extract the book theme vector corresponding to the current reading state. Based on the reading state transition matrix, perform recursive cumulative calculation on the book theme vector to generate interest-oriented feature vectors and form a reader interest feature dataset. S6. Input the reader interest feature dataset and the current reading status identifier into the recommendation rule trigger function, calculate the recommendation matching value, and generate the recommendation matching dataset; S7. When the recommended matching value meets the preset trigger threshold condition, a candidate book set is generated from the book resource library based on the recommended matching dataset, and the books are sorted according to the recommended matching value to output the recommendation result. S8. After the recommendation results are output, collect the reader's subsequent behavior data on the candidate book set, generate a feedback behavior dataset, and merge the feedback behavior dataset into the reader's original behavior dataset. S9. Based on the consistency comparison between the feedback behavior dataset and the current reading state dataset, calculate the state deviation value. When the state deviation value exceeds the preset correction threshold, update the parameter set of the state determination function and the recommendation rule triggering function.

2. The book recommendation method based on reader behavior feedback update according to claim 1, characterized in that, Within a preset time window, the behavioral time series dataset is segmented and statistically analyzed to calculate the frequency and duration of behaviors in each time period, generating a behavioral intensity index sequence and forming a behavioral sequence feature dataset, including: S31. Divide the behavioral time series dataset into time segments according to a preset time window length to generate multiple time subseries datasets; S32. For each time subsequence dataset, construct a behavior category hierarchical matrix according to the behavior type identifier. The behavior category hierarchical matrix is ​​used to characterize the distribution structure of different types of behaviors in the current time window. S33. Based on the behavior category hierarchical matrix, calculate the behavior coupling coefficient between each behavior category. The behavior coupling coefficient is used to characterize the degree of correlation between different behaviors within the same time window and generate behavior correlation feature data. S34. Based on the behavior-related feature data, calculate the behavior frequency vector and behavior duration vector within the current time window, and incorporate the behavior coupling coefficient into the calculation process of the behavior frequency vector to generate a coupling frequency vector. S35. Normalize the coupling frequency vector and behavior duration vector, and based on the time decay recursive function, recursively accumulate the calculation results of different time windows to generate the behavior intensity index of the current time window. S36. Arrange the behavior intensity indices generated in each time window in chronological order to form a behavior intensity index sequence, and construct a behavior sequence feature dataset by combining the behavior association feature data.

3. The book recommendation method based on reader behavior feedback update according to claim 1, characterized in that, Based on the behavior sequence feature dataset, a preset state determination function is invoked to perform interval mapping on the behavior intensity index sequence, generating reading state identifiers. A reading state transition matrix is ​​constructed based on the reading state identifiers of adjacent time windows, generating the current reading state dataset, including: S41. Input the behavior intensity index of different time periods in the behavior sequence feature dataset into a preset multi-level state determination function. The multi-level state determination function includes an interval mapping module, a rate of change analysis module, a behavior fusion module, and a confidence generation module, which are used to output the preliminary reading state identifier and the state confidence value. S42. Construct a state transition probability matrix based on the reading state identifiers of adjacent time windows. P(k) Each matrix element P(i,j) Representing reading state i To state j The probability of transition is calculated and stored as a sequence of transition matrices within a time window, and the state transition history is recorded. S43. Combining the state confidence value with the state transition probability matrix P(k) Calculate a weighted state score to measure the stability and reliability of the reading state in the current time window; S44. When the weighted state score is lower than the preset threshold, the initial reading state is corrected, a corrected time window reading state identifier is generated, and the state correction log is recorded. S45. Arrange the corrected time window reading status identifiers in chronological order to form the current reading status identifier sequence, and output the status confidence sequence for interest-oriented feature vector generation and recommendation rule triggering. S46. Construct the final reading state transition matrix based on the corrected time window reading state identifier, and generate the current reading state dataset, which includes the corrected reading state sequence, state transition matrix and state confidence information, to provide complete input data for subsequent recommendation and dynamic adjustment.

4. The book recommendation method based on reader behavior feedback update according to claim 3, characterized in that, The behavioral intensity index of different time periods in the behavioral sequence feature dataset is input into a preset multi-level state determination function. This multi-level state determination function includes an interval mapping module, a rate of change analysis module, a behavior fusion module, and a confidence generation module, used to output preliminary reading state identifiers and state confidence values, including: S41-1. The interval mapping module divides the behavior intensity index within the current time window into multi-level intervals. According to the preset interval mapping rules, the behavior intensity index is mapped to a predefined multi-level reading state set S = {browsing state, shallow reading state, deep reading state, interest decay state}, and candidate reading state identifiers are generated. S41-2. The rate of change analysis module calculates the rate of change of behavior intensity within the current time window, extracts the trend of behavior intensity fluctuation, and generates a state stability index to characterize the persistence and fluctuation characteristics of the reading state. S41-3. The behavior fusion module performs a weighted combination of the behavior intensity indices corresponding to different behavior types within the same time window to form a comprehensive behavior intensity index. S41-4. The confidence generation module combines the comprehensive behavior intensity index, the state stability index and the reading state identifier of the previous time window to perform a consistency evaluation on the current candidate reading state and generate a state confidence value. S41-5. The multi-level state determination function outputs the preliminary reading state identifier and the corresponding state confidence value of the current time window based on the comprehensive behavior intensity index and the interval mapping result.

5. The book recommendation method based on reader behavior feedback update according to claim 4, characterized in that, Based on the reading status identifiers of adjacent time windows, a state transition probability matrix is ​​constructed. P(k) Each matrix element P(i,j) Representing reading state i To state j The probability of transitions is calculated and stored as a sequence of transition matrices within a time window, and the state transition history is recorded, including: S42-1. Number and encode the multi-level reading state set to establish a reading state index set; S42-2. Based on the reading status identifiers of adjacent time windows, generate single-step state transition records and construct a single-step state transition indication matrix; S42-3. Based on the state confidence value, the single-step state transition indication matrix is ​​subjected to confidence weighting to generate a weighted state transition matrix, which is used to reduce the impact of low-confidence states on the calculation of transition probabilities. S42-4. Recursively accumulate the weighted state transition matrix of the current time window with the historical cumulative transition matrix, and introduce a time decay factor to attenuate and update the historical transition intensity, generating the cumulative state transition intensity matrix. A(k) The update method satisfies the formula: A(k) = λ · A(k-1) ＋ ( k ) ； in, A(k) For the first k The cumulative state transition intensity matrix within the time window. A(k-1) This is the cumulative state transition intensity matrix for the previous time window. ( k ) is the weighted state transition matrix for the current time window, and λ is the time decay factor, which ranges from 0 to 1 and is used to control the degree of influence of historical state transitions on the current cumulative result; S42-5, Regarding the cumulative state transition strength matrix A(k) After normalization, the state transition probability matrix of the current time window is obtained. P(k) ; Among them, matrix elements P(i,j) This indicates the reading status within the current time window. i Towards reading state j The conditional probability of the transition, which is determined by the cumulative state transition strength matrix. A(k) The ratio of the corresponding element in the middle to the sum of the elements in that row is determined for any reading state. i The sum of all its corresponding transition probabilities satisfies the preset normalization condition; S42-6. Calculate the state transition probability matrix generated for each time window. P(k) The states are stored in chronological order as a sequence of state transition matrices, and a state transition history log is recorded synchronously. The state transition history log includes the reading state identifiers of adjacent time windows and the corresponding state confidence values, which are used for subsequent state stability analysis and dynamic correction of recommended strategies.

6. The book recommendation method based on reader behavior feedback update according to claim 1, characterized in that, Based on the current reading state dataset and reading state transition matrix, and in conjunction with the behavioral time series dataset, the book feature association mapping table is invoked to extract the book topic vector corresponding to the current reading state. Based on the reading state transition matrix, the book topic vector is recursively accumulated to generate interest-oriented feature vectors, thus forming a reader interest feature dataset, including: S51. Extract the current time window reading state identifier from the current reading state dataset, and calculate the future state trend distribution of the current reading state based on the reading state transition matrix to characterize the reader's state evolution trend. S52. Call the book feature association mapping table, which establishes a mapping relationship between reading status and book topic, and maps the current reading status identifier to the corresponding book topic vector. Each reading state corresponds to multiple topic dimensions, and the book topic vector is used to represent the content interest direction in the current reading state; S53, Based on the state transition probability matrix P(k) Based on the transition probability distribution corresponding to the current state, the book topic vector is subjected to state transition weighted expansion processing to generate a state-driven topic expansion vector. T′k This is used to indicate the direction of potential interest migration; S54. Drive the topic expansion vector with the state of the current time window. T′k Historical interest-oriented feature vector H(k-1) Perform recursive cumulative updates to generate the interest-oriented feature vector for the current time window. H(k) The formula is: H(k) = α×H(k-1) ＋ β×T′k ； in, H(k) The feature vector pointing to the interest of the current time window. H(k-1) The feature vector pointing to the interest of the previous time window. T′k For state-driven topic extension vectors, α This represents the historical interest decay coefficient. β This represents the current theme reinforcement coefficient. α、β All are dynamically adjustable parameters, and are adaptively adjusted according to the corresponding transition probabilities in the current reading state transition matrix, so that the deep reading state corresponds to... β The value is greater than the browsing status corresponding to β value; Among them, the state-driven topic expansion vector T′k The degree of contribution is determined by the state transition probability distribution, and historical interest points to the feature vector. H(k-1) The effect gradually diminishes over time. S55. Analyze the interest-oriented feature vectors generated in each time window. H(k) Construct a dataset of reader interest features in chronological order.

7. The book recommendation method based on reader behavior feedback update according to claim 6, characterized in that, The reader interest feature dataset and the current reading state identifier are input into the recommendation rule trigger function to calculate the recommendation matching value and generate a recommendation matching dataset, including: S61. Construct a set of candidate book theme feature vectors, wherein the candidate book theme feature vectors are extracted from the theme feature data of each book in the book resource database, and are compared with the interest orientation feature vector. H(k) Perform dimension alignment. S62. Based on the current reading status In the state transition probability matrix P(k) Calculate the state adjustment factor from the transition probability distribution of the corresponding row. The calculation formula is: ； in, To the current reading state j Given the transition probabilities of each state, the maximum transition probability in that row is taken as the state stability. This is a state-sensitive adjustment coefficient. When the concentration of state transition probabilities is high, it enhances the strength of recommendation reinforcement; when the distribution of state transition probabilities is relatively dispersed, it reduces the strength of reinforcement and enhances the diversity of recommendations. S63. Define a recommendation rule trigger function. This function calculates the book matching degree based on the current reading state identifier and interest-oriented feature vector. The calculation process includes a similarity calculation step and a state adjustment step. The similarity calculation formula is as follows: ； in, For the first The recommended match value for this candidate book. The feature vector pointing to the interest of the current time window. For the first Thematic feature vector of this candidate book; S64. Based on the recommended matching value of candidate books Generate a recommendation matching dataset, which includes book identifiers, matching values, and state adjustment factor information; S65. Sort the recommended matching dataset and arrange the candidate books from high to low according to the matching value. When the recommended matching value meets the preset trigger threshold condition, output the corresponding book as the recommendation result.

8. The book recommendation method based on reader behavior feedback update according to claim 1, characterized in that, When the recommended matching value meets the preset trigger threshold condition, a candidate book set is generated from the book resource library based on the recommended matching dataset, and the books are sorted according to the recommended matching value. The recommendation result is then output, including: S71. Detect the recommendation matching value of each candidate book in the recommendation matching dataset. Determine whether the preset trigger threshold condition is met. When the recommended matching value is greater than or equal to the preset trigger threshold, add the corresponding book identifier to the candidate filtering queue. S72. Based on the book identifiers added to the candidate filtering queue, retrieve the corresponding book resource information from the book resource library. The book resource information includes book theme feature data, category label data, and resource availability information to generate a candidate book set. S73. Sort the books in the candidate book set from high to low according to the recommended matching value. When there are books with the same recommended matching value, perform a second sorting according to the priority of the current reading status identifier. S74. Control the number of candidate books after sorting, and extract the top N books according to the preset output quantity limit to form a recommendation result set. The output quantity limit is dynamically adjusted according to the current reading status identifier or the system resource load status. S75. Output the set of recommended results to the recommendation display module and record the recommendation output log for subsequent status correction.

Images