Personalized recommendation system for reading software based on big data analytics

Abstract

This invention discloses a personalized recommendation system for reading software based on big data analysis, belonging to the field of computer software technology. Addressing the problems raised in previous technologies—that merely characterizing interests makes it difficult to depict the cognitive complexity users can process, that scene constraints are not transformed into quantitative requirements for cognitive load attributes and are not calculated in conjunction with user ability profiles, and that it is difficult to distinguish whether abnormal feedback stems from ability transfer or scene constraint distortion—this invention proposes the following solution, including data collection: the client SDK collects real-time environmental context data of the user's deep behavioral sequence during the reading process. This invention avoids recommending content that is short in form but cognitively overloaded, can distinguish whether abnormal feedback originates from user ability transfer or scene constraint definition deviation, and performs incremental calibration on the reading ability preference vector and scene constraint vector respectively, reducing the candidate size of the ranking model, reducing the amount of data transmission between the server and the client, and improving the overall system operating efficiency.

Description

Technical Field

[0001] This invention relates to the field of computer software technology, and in particular to a personalized recommendation system for reading software based on big data analysis. Background Technology

[0002] With the popularization of mobile internet and digital reading, various reading software have accumulated massive amounts of user behavior data. Personalized recommendation technology has become a key means to improve user experience and enhance user stickiness. Traditional reading recommendation systems are mostly based on collaborative filtering or content matching algorithms, mainly relying on explicit feedback such as users' historical clicks, favorites, and ratings, as well as the content's metadata (such as author, category, and tags) for recommendations.

[0003] However, as an information acquisition activity with high cognitive load and strong immersion requirements, reading has the following limitations with existing technologies: 1. It only describes interests but fails to capture the cognitive complexity that users can process; 2. Scene perception is mostly limited to genre or length filtering, and does not transform scene constraints into quantitative requirements for cognitive load attributes and calculate them in conjunction with user capability profiles; 3. In-session optimization and profile updates are disconnected, making it difficult to distinguish whether abnormal feedback comes from capability migration or scene constraint distortion. Summary of the Invention

[0004] This invention provides a personalized recommendation system for reading software based on big data analysis. It solves the problems of existing personalized recommendation systems that only describe interests but fail to describe the cognitive complexity that users can handle, fail to transform scene constraints into quantitative requirements for cognitive load attributes and calculate them in conjunction with user ability profiles, and have difficulty distinguishing whether abnormal feedback comes from ability transfer or scene constraint distortion.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Personalized recommendation systems for reading software based on big data analytics include: The data acquisition module is used to collect real-time user reading depth behavior data sequences and environmental context data; The user profile modeling module is used to construct and update user profiles based on the reading depth behavior sequence. The user profile includes a reading ability preference vector, which is generated by analyzing the correlation model between the text complexity features of the user's historical reading content and the corresponding behavioral feedback, and represents the user's tolerance range for the cognitive complexity of the content. The scene parsing module is used to identify the current reading scene mode based on environmental context data, and associate scene constraint vectors with the identified scene mode. The scene constraint vectors include at least the expected reading time, the allowed frequency of interaction interruption, and the level of environmental interference. A hybrid recommendation engine is connected to the user profile modeling module and the scene parsing module, respectively. The hybrid recommendation engine includes: The capability-scenario joint recaller is configured to: prune the tolerance interval in the reading capability preference vector under the cognitive load constraint obtained by mapping the scenario constraint vector to obtain a scenario-based capability window; and weight and synthesize the vector representing the scenario-based capability window with the interest topic vector to generate a co-adaptation vector; wherein the pruning includes at least setting upper limit thresholds for the information density index and logical complexity index of the content and narrowing or shifting the upper boundary of the tolerance interval, and using the co-adaptation vector as the core retrieval condition to recall an initial candidate set from the content library whose content feature vector matches the co-adaptation vector; A real-time session optimizer is configured to: within a single reading session, dynamically adjust the weight ratio of capability factors and scenario factors in the co-adaptation vector based on the user's real-time interactive feedback on the initial candidate set using an online learning model, and re-rank the initial candidate set based on the adjusted co-adaptation vector to generate a final recommendation list; wherein the online learning model is a contextual multi-armed gambling machine or an online logistic regression model, and its input features explicitly include the weight ratio state; The profile-scene collaborative updater is configured to: when the real-time session optimizer detects that the user's feedback behavior continuously deviates from the expectation based on the historical profile and the current scene within a preset time window, it triggers an incremental calibration update of the reading ability preference vector and the scene constraint vector, wherein the deviation is obtained by the divergence or distance measure between the feedback difference vector formed by the real-time feedback and the expected feedback vector.

[0006] Preferably, the text complexity features include at least: lexical difficulty, syntactic tree depth, reference chain length, topic jump rate, and information entropy; the logical complexity index is obtained by calculating the graph complexity of the content's text structure graph.

[0007] Preferably, the reading depth behavior data sequence includes at least: chapter-by-chapter reading speed vector, page dwell distribution, reread rate, interaction heatmap, and note semantic anchor data; the environmental context data includes at least: time tag, location tag, device type, network status, device battery level, and motion status of the accelerometer.

[0008] Preferably, the scene analysis module uses sensor fusion and clustering models to identify reading scene patterns, and the scene patterns include at least a commuting short reading mode and a deep focus mode; the sensor fusion includes Kalman filtering fusion of GPS and WiFi positioning results.

[0009] Preferably, the tolerance interval learning includes: constructing a multi-dimensional behavior-complexity correlation matrix; predicting the user's expected completion rate and interaction depth for a given combination of complexity features through a regression model; defining the complexity feature space with predicted values ​​higher than a threshold as a dynamic complexity tolerance interval, wherein the threshold is a threshold that adaptively changes with the scene constraint vector.

[0010] Preferably, the real-time interactive feedback includes at least negative feedback signals and positive feedback signals; negative feedback signals include recommendations being closed or swiped away within a preset time after exposure; positive feedback signals include reading time exceeding the historical average and accompanied by high interaction behavior; and different signals are assigned weights to generate a feedback difference vector.

[0011] Preferably, the triggering logic of the profile-scene collaborative updater includes: if, in a certain scene mode, the user gives negative feedback multiple times to content at the boundary of the scene-based capability window, it is determined that the scene constraint vector is inaccurate or the user's capability preference has shifted, and incremental calibration is performed on the scene constraint vector or the reading capability preference vector respectively.

[0012] Preferably, the content feature vector of the content library includes a cognitive load attribute vector and an interest topic vector; the capability-scenario joint recaller uses approximate nearest neighbor retrieval to achieve candidate recall, and limits the number of recalled candidates to a preset upper limit to reduce computational overhead.

[0013] The beneficial effects of this invention are as follows: 1. Achieve dynamic cognitive matching from static interest matching: Under scenario constraints, trim the capability tolerance range and generate scenario-based capability windows to avoid recommending content that is short in form but has an excessive cognitive load, thus significantly improving reading completion rate and user satisfaction; 2. Forming an evolutionary closed loop between profiles and scenarios: Through a real-time feedback deviation detection mechanism, the system can distinguish whether abnormal feedback stems from user ability migration or scenario constraint definition deviation, and perform incremental calibration on the reading ability preference vector and scenario constraint vector respectively. 3. Reduce computational and network overhead caused by invalid recommendations: During the recall phase, the retrieval space is narrowed by co-adaptive vectors, reducing the candidate size of the ranking model, reducing the amount of data transmission between the server and the client, and improving the overall operating efficiency of the system.

[0014] In summary, this invention can avoid recommending content that is short in form but overloaded with cognitive load, can distinguish between abnormal feedback stemming from user ability migration or deviation in the definition of scenario constraints, and incrementally calibrates the reading ability preference vector and scenario constraint vector respectively, thereby reducing the candidate size of the ranking model, reducing the amount of data transmission between the server and the client, and improving the overall operating efficiency of the system. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall architecture of the personalized recommendation system for reading software based on big data analysis proposed in this invention. Detailed Implementation

[0016] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Example 1, referring to Figure 1 A personalized recommendation system for reading software based on big data analytics includes a data acquisition module, a user profile modeling module, a scenario analysis module, and a hybrid recommendation engine.

[0018] The system architecture and data flow involve the client SDK collecting real-time user reading depth behavior sequences and environmental context data, which are then uploaded to the server via an encrypted channel. Reading depth behavior sequences include chapter-by-chapter reading speed (average reading time per chapter), reread rate (number of paragraphs revisited / total number of paragraphs), interaction heatmaps (distribution of clicks, swipes, and long presses), and semantic anchors for notes (text snippets highlighted or commented by the user). Environmental context includes time tags (accurate to the minute), location tags (GPS coordinates or WiFi hotspot), device type (phone model, screen size), network status (4G / 5G / WiFi), battery level, and motion status from the accelerometer (stationary, walking, riding in a vehicle). To reduce transmission overhead, the client performs sliding window aggregation on the behavioral sequences: a sliding window with a length of 30 seconds and a step size of 5 seconds is used to extract statistical features (mean, variance, maximum value) from the time-series data such as reading speed, dwell time, and number of interactions within the window, and only the aggregated statistical features and necessary anonymized device identifiers are applied. The original sequence data is cached locally and then periodically cleared. Complexity feature extraction and tolerance interval learning: For each piece of content read by a user, the server extracts the following text complexity features: Vocabulary difficulty: Flesch-Kincaid level based on word frequency statistics; Syntax tree depth: The average depth of a sentence parse tree; Reference chain length: the average distance between a pronoun and its referent within a paragraph; Topic jump rate: the cosine distance between topic vectors of adjacent paragraphs; Information entropy: Shannon entropy of the content word distribution; Logical complexity: By constructing a chapter structure graph (nodes are paragraphs, edges are references or causal relationships), calculate the number of nodes, the number of edges, and the average shortest path length of the graph; The above features are combined with the user's deep behavioral features on the content (completion rate, note density, reread rate) to form a behavior-complexity correlation matrix. An XGBoost regression model is used for training. The complexity features are input, and the expected completion rate and interaction depth (normalized to [0, 1]) are output. After the model is trained, for any combination of complexity features, its completion rate and interaction depth are predicted. The feature space with predicted values ​​higher than the dynamic threshold is defined as the user's dynamic complexity tolerance interval. The dynamic threshold is adaptively adjusted according to the current scenario: for example, the threshold is 0.6 in commuting mode and 0.8 in focus mode; it can also be dynamically set based on the percentile of the user's historical completion rate (such as the 70th percentile). Scene constraint vectors and scene-based capability windows; the scene parsing module uses sensor fusion (GPS and WiFi positioning fused by Kalman filtering) and clustering models (such as DBSCAN) to identify scene patterns, including at least: Commute Short Reading Mode: Expected reading time <5 minutes, allows for high frequency of interactive interruptions and high level of environmental interference; Deep Focus Mode: Expected reading time > 30 minutes, low frequency of interruptions, and low level of environmental interference; A scenario constraint vector is predefined for each scenario mode. For example, the constraint vector for the commuting mode is (expected duration = 3 minutes, interruption frequency = 0.8, interference level = 0.9). Then, the scenario constraints are mapped to constraints on the cognitive load attributes of the content. The mapping method uses a piecewise linear function. The upper limit of information density is D = 0.1 × expected duration + 0.2 (when the expected duration ≤ 10 minutes) or D = 0.05 × expected duration + 0.5 (when the expected duration > 10 minutes). The upper limit of logical complexity is L = 0.08 × expected duration + 0.1 (when expected duration ≤ 10 minutes) or L = 0.03 × expected duration + 0.6 (when expected duration > 10 minutes). Within the user tolerance range (e.g., complexity range [4.0, 8.5]), the above constraints are used to prune the content to obtain the scenario-based capability window (e.g., after pruning, it becomes [4.0, 6.2]), which is the range of content complexity that the user can actually accept in the current scenario. In-session online learning and collaborative updates: Within a single reading session, the real-time session optimizer generates a weighted feedback difference vector P based on real-time feedback such as exposure, clicks, dwell time, swipes, and notes. Specifically, positive feedback (reading time exceeding the historical average and accompanied by high interaction) is encoded as (1, 0), and negative feedback (swipes away within 2 seconds of exposure) is encoded as (0, 1), and weights are assigned (e.g., positive +0.5, negative -0.3), which are then accumulated to form a vector. The weight ratio of capability factors and scenario factors in the collaborative adaptation vector is dynamically updated using online logistic regression or contextual multi-armed gambling machine (LinUCB). The model input features explicitly include the current weight ratio state and candidate content features, and the output is the re-ranking score. The condition for triggering collaborative updates between the profile and the scene is: Calculate the KL divergence between the feedback difference vector P and the expected feedback vector Q (based on the predicted feedback distribution of historical profiles and the current scene). The threshold is set to 0.1. If the KL divergence exceeds the threshold for three consecutive time windows (each time window includes the most recent 10 exposures), the updater is triggered to perform incremental calibration (using stochastic gradient descent fine-tuning) on ​​the reading ability preference vector or the scene constraint vector respectively, in order to distinguish between ability transfer and scene definition distortion.

[0019] Example 2, comparative experiment: In an offline playback experiment containing 100,000 reading sessions, Example 1 was compared with the baseline method of "interest-only recall + scene and genre filtering", and the following table was obtained: In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0020] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A personalized recommendation system for reading software based on big data analysis, characterized in that: include: The data acquisition module is used to collect real-time user reading depth behavior data sequences and environmental context data; The user profile modeling module is used to construct and update user profiles based on the reading depth behavior sequence. The user profile includes a reading ability preference vector. The tolerance range in the reading ability preference vector is a dynamic complexity tolerance range obtained by constructing a correlation matrix between the text complexity features of the user's historical reading content and the corresponding behavioral feedback, and by using a regression model. The scene parsing module is used to identify the current reading scene mode based on environmental context data, and associate scene constraint vectors with the identified scene mode. The scene constraint vectors include at least the expected reading time, the allowed frequency of interaction interruption, and the level of environmental interference. A hybrid recommendation engine is connected to the user profile modeling module and the scene parsing module, respectively. The hybrid recommendation engine includes: The capability-scenario joint recaller is configured to: prune the tolerance interval in the reading capability preference vector under the cognitive load constraint obtained by mapping the scenario constraint vector to obtain a scenario-based capability window; and weight and synthesize the vector representing the scenario-based capability window with the interest topic vector to generate a co-adaptation vector; wherein the pruning includes at least setting upper limit thresholds for the information density index and logical complexity index of the content, and narrowing or shifting the upper boundary of the tolerance interval; and using the co-adaptation vector as the core retrieval condition to recall an initial candidate set from the content library whose content feature vector matches the co-adaptation vector. A real-time session optimizer is configured to: within a single reading session, dynamically adjust the weight ratio of capability factors and scenario factors in the co-adaptation vector based on the user's real-time interactive feedback on the initial candidate set using an online learning model, and re-rank the initial candidate set based on the adjusted co-adaptation vector to generate a final recommendation list; wherein the online learning model is a contextual multi-armed gambling machine or an online logistic regression model, and its input features explicitly include the weight ratio state; The profile-scene collaborative updater is configured to: when the real-time session optimizer detects that the user's feedback behavior continuously deviates from the expectation based on the historical profile and the current scene within a preset time window, trigger an incremental calibration update of the reading ability preference vector and the scene constraint vector, wherein the deviation is obtained by the divergence or distance measure between the feedback difference vector formed by the real-time feedback and the expected feedback vector, and the divergence is calculated using KL divergence.

2. The personalized recommendation system for reading software based on big data analysis according to claim 1, characterized in that, The text complexity features include at least: lexical difficulty, syntactic tree depth, reference chain length, topic jump rate, and information entropy; the logical complexity index is obtained by calculating the graph complexity of the content's text structure graph.

3. The personalized recommendation system for reading software based on big data analysis according to claim 1, characterized in that, The reading depth behavior data sequence mentioned above includes at least: chapter-by-chapter reading speed vector, page dwell distribution, reread rate, interaction heatmap, and note semantic anchor data; the environmental context data includes at least: time tag, location tag, device type, network status, device battery level, and motion status of the accelerometer.

4. The personalized recommendation system for reading software based on big data analysis according to claim 1, characterized in that, The scene analysis module uses sensor fusion and clustering models to identify reading scene patterns, which include at least commuting short reading mode and deep focus mode; the sensor fusion includes Kalman filtering fusion of GPS and WiFi positioning results.

5. The personalized recommendation system for reading software based on big data analysis according to claim 1, characterized in that, The tolerance interval learning includes: constructing a multi-dimensional behavior-complexity correlation matrix; predicting the expected completion rate and interaction depth of users for a given combination of complexity features through a regression model; defining the complexity feature space with predicted values ​​higher than a threshold as a dynamic complexity tolerance interval, wherein the threshold is a threshold that adaptively changes with the scene constraint vector.

6. The personalized recommendation system for reading software based on big data analysis according to claim 1, characterized in that, The real-time interactive feedback includes at least negative feedback signals and positive feedback signals; negative feedback signals include recommendations being closed or swiped away within a preset time after exposure; positive feedback signals include reading time exceeding the historical average and accompanied by high interaction behavior; and different signals are assigned weights to generate a feedback difference vector.

7. The personalized recommendation system for reading software based on big data analysis according to claim 1, characterized in that, The triggering logic of the profile-scene collaborative updater includes: if, in a certain scene mode, a user gives negative feedback multiple times to content at the boundary of the scene-based capability window, it is determined that the scene constraint vector is not defined accurately or the user's capability preference has shifted, and incremental calibration is performed on the scene constraint vector or the reading capability preference vector respectively.

8. The personalized recommendation system for reading software based on big data analysis according to claim 1, characterized in that, The content feature vector of the content library includes a cognitive load attribute vector and an interest topic vector; the capability-scenario joint recaller uses approximate nearest neighbor retrieval to achieve candidate recall and limits the number of recalled candidates to a preset upper limit to reduce computational overhead.

Images