APP user reading data monitoring and analyzing method and system based on big data

Abstract

The invention discloses an APP user reading data monitoring and analyzing method and system based on big data, and the method comprises the steps: collecting and preprocessing the reading behavior data of an APP user, and generating a standardized data set; constructing a user-content-behavior relationship network based on the standardized data set; performing iteration on the initial label assignment of the nodes and a label propagation algorithm to obtain a label distribution result of all the nodes; forming a high-dimensional sparse feature set based on a label distribution result; performing feature interaction modeling by the factorization machine, and outputting a behavior prediction result; and according to a behavior prediction result, monitoring and analyzing interests and preferences, and dynamically updating label distribution. According to the method and the system, intelligent monitoring and analysis of the reading data of the APP user are realized, and the accuracy and the real-time performance of personalized content recommendation are improved.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a method and system for monitoring and analyzing APP user reading data based on big data. Background Technology

[0002] With the rapid development of the internet, mobile applications (APPs) have become an important part of people's daily lives. APP user behavior data, especially reading behavior data, can provide a wealth of information about user interests and preferences. Existing recommendation systems typically generate personalized recommendations by analyzing users' historical behavior and using technologies such as collaborative filtering and content recommendation. However, traditional recommendation systems mainly rely on simple user behavior data, such as clicks and browsing time, and usually depend on static models for prediction, making it difficult to flexibly respond to rapid changes in user interests and dynamic behavioral patterns.

[0003] Currently, many recommender systems rely on traditional collaborative filtering algorithms or content-based recommendation methods. While effective to some extent, these methods also have significant drawbacks. Collaborative filtering algorithms face the problem of data sparsity, particularly in the cold start scenario with new users or new content, where their recommendation performance is poor. Content-based recommendation methods often rely too heavily on content features, neglecting the complex interactions between users. This results in lower accuracy for the recommender system and makes it difficult to capture deeper changes in user interests.

[0004] While existing label propagation algorithms can effectively propagate label information within network structures, they still have some problems in practical applications. Traditional label propagation algorithms rely on pre-set initial labels, and the label update process lacks sufficient dynamism and flexibility, causing the label distribution to fail to reflect changes in user interests in a timely manner, thus affecting the accuracy and real-time performance of recommendations.

[0005] Therefore, how to provide a method and system for monitoring and analyzing APP user reading data based on big data is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method and system for monitoring and analyzing APP user reading data based on big data. This invention fully utilizes tag propagation algorithms and factorization machine technology, and details an intelligent analysis method that constructs a user-content-behavior relationship network and uses feature interaction modeling. By dynamically updating user interests, behavioral patterns, and content preferences, combined with tag distribution optimization, this invention achieves continuous monitoring and accurate analysis of APP user reading data. It possesses efficient personalized recommendation capabilities, real-time user interest capture, and accurate behavior prediction capabilities, effectively improving the accuracy and personalization level of recommendation results, and solving the problems of data sparsity and tag update lag in existing technologies.

[0007] A method for monitoring and analyzing APP user reading data based on big data according to an embodiment of the present invention includes: Collect reading behavior data from APP users, preprocess the reading behavior data, and generate a standardized dataset; Based on a standardized dataset, a user-content-behavior relationship network is constructed, with user nodes and content nodes as vertices and user reading behavior as edges, forming a heterogeneous behavior network structure. For some nodes in the user-content-behavior relationship network, initial labels are assigned based on existing interest tags or content topic information. A label propagation algorithm is then applied to iteratively propagate the initial label information to unlabeled nodes through the edge relationships between nodes, resulting in the label distribution of all nodes. Based on the label distribution results, the label features of each user node and content node are extracted, and the label features are merged with the relevant features in the standardized dataset to form a high-dimensional sparse feature set. A joint analysis of high-dimensional sparse feature sets is performed using factorization machine. The factorization machine is used to jointly model the second-order and higher-order feature interaction relationships between user, content, behavior and tag features, and output the prediction results of user behavior towards content. Based on the behavioral prediction results, user interests, reading behavior patterns, and content preferences are monitored and analyzed, and the tag distribution in the user-content-behavior relationship network is dynamically updated to achieve continuous monitoring and intelligent analysis of APP user reading data.

[0008] Optionally, the reading behavior data specifically includes users' browsing, clicking, collecting, sharing, and commenting actions on APP content, as well as corresponding time information and content identifiers.

[0009] Optionally, the preprocessing of the reading behavior data specifically includes deduplication, missing value imputation, outlier removal, and standardization of the collected reading behavior data.

[0010] Optionally, based on a standardized dataset, a user-content-behavior relationship network is constructed, with user nodes and content nodes as vertices and user reading behavior as edges, forming a heterogeneous behavior network structure, including: In the standardized dataset, all user identifiers, content identifiers, and information on the types and timing of reading behaviors between users and content are identified and extracted to generate a list of basic information containing users, content, and behaviors. Based on the list of basic information, establish a set of user nodes and a set of content nodes respectively, and assign each user identifier to a user node and each content identifier to a content node; Traverse the list of basic information, and based on each user's reading behavior, establish an edge between the corresponding user node and content node, and mark the type and time of the reading behavior on the edge; Collect all reading behavior records between each user node and content node, and assign weights to the edges between user nodes and content nodes based on different types of reading behaviors and their frequency. The weights are determined comprehensively based on the behavior type and frequency. The weights of the edges are adjusted based on the time of the behavior, with the weights of historical behaviors that are far removed from the current time being reduced and the weights of recent behaviors being increased, reflecting the user's latest interests and behavioral activity. By utilizing the established set of user nodes, set of content nodes, and weighted user-content edges, a user-content-behavior relationship network is constructed.

[0011] Optionally, for some nodes in the user-content-behavior relationship network, initial labels are assigned based on existing interest tags or content topic information. A label propagation algorithm is then applied to iteratively propagate the initial label information to unlabeled nodes through edge relationships between nodes, resulting in the label distribution of all nodes, including: In the user-content-behavior relationship network, some nodes with known interest tags or content topic information are identified and designated as labeled nodes. Each labeled node is assigned an initial label distribution result and a label confidence of 1, while the remaining unlabeled nodes are assigned an empty label distribution result and a confidence of 0. For each edge between a user node and a content node, a temporal decay weight is assigned to each edge based on the time when the user's reading behavior occurs. The closer the behavior is to the current time, the higher the weight is assigned, and the farther the behavior is to the current time, the lower the weight is assigned, forming an edge weight set containing temporal information. During the label propagation iteration process, each unlabeled node collects the current label and confidence information of all neighboring nodes, and combines the temporal decay weight of the neighboring edges to perform weighted score statistics on all labels; The label with the highest weighted score is used as the new label for the current unlabeled node. The confidence of the node's current label is updated based on the confidence of the neighboring node labels involved in the calculation and the temporal decay weight of the edge. The confidence is taken as the weighted average. A diversity test is performed on the new labels and confidence scores assigned to each node. If the node label distribution is too concentrated on a single label, the second-best label with the highest confidence score is introduced to adjust the label distribution of the node. After each iteration, the label distribution of all nodes is statistically analyzed. If a node with an abnormally concentrated label distribution is found, the label distribution of the node is automatically adjusted based on the node's historical label distribution and the diversity of neighboring node labels, and self-correction is performed. Repeat the propagation of label distribution results, confidence updates, diversity adjustments and self-corrections until the label distribution of all nodes remains stable or the preset number of iterations is reached. Finally, output the label distribution results and corresponding confidence scores of all nodes after convergence.

[0012] Optionally, based on the label distribution results, the label features of each user node and content node are extracted, and the label features are merged with relevant features in the standardized dataset to form a high-dimensional sparse feature set, including: From the tag distribution results, tag features are extracted for each user node and each content node. The tag features include the distribution of the interest category to which the node belongs and the distribution of the content topic. In the standardized dataset, for each pair of user and content behavior records, user behavior features, content features and contextual features of the behavior are extracted. The behavior features include user activity, historical behavior type and frequency, the content features include content category and content popularity, and the contextual features include the time of the behavior and device type. The obtained tag features are concatenated with user behavior features, content features, and behavior context features to form a joint feature vector for each user-content behavior relationship; For all user and content behavior records, generate joint feature vectors sequentially, and combine all feature vectors to form a high-dimensional sparse feature set; The high-dimensional sparse feature set is indexed and organized, and the organized high-dimensional sparse feature set is output.

[0013] Optionally, the high-dimensional sparse feature set is jointly analyzed using a factorization machine. The factorization machine is used to jointly model the second-order and higher-order feature interaction relationships between user, content, behavior, and tag features, outputting user behavior prediction results for the content, including: The obtained high-dimensional sparse feature set is used as input. The high-dimensional sparse feature set includes user features, content features, behavior features and tag features. For each pair of features in the high-dimensional sparse feature set, a learnable weight parameter is assigned. The weight parameter is adjusted together with the factorization-related parameters during the parameter update process to reflect the importance of different features to the interaction. Different factor vector dimensions are set for user features, content features, behavior features, and tag features in the high-dimensional sparse feature set, and a factor vector of corresponding length is used for each category feature; When performing interactive computation on a high-dimensional sparse feature set, the factor vector of the corresponding dimension is selected according to the feature category. Then, second-order and higher-order interactive operations are performed between the factor vector and the factor vector of other categories of features to form the interactive expression result between features. The interactive result is stored as an interactive feature vector. Based on the obtained interaction feature vector, it is fused with user features, content features, behavioral features and tag features in the high-dimensional sparse feature set to form a fused feature set, which contains the interaction information between all features; For each pair of feature interactions in the fused feature set, the impact on the final behavior prediction result is calculated. If the interaction contribution of a certain feature pair is lower than the set threshold, the corresponding factor vector parameter is set to zero, and the invalid feature pair is removed from the feature interaction calculation. Repeatedly perform feature interaction calculation, factor vector update, and low-impact feature pair removal operations to continuously optimize feature interaction expression and parameter distribution, so that the effective interaction between features is fully expressed and redundant and invalid feature pairs are effectively removed. Based on the optimized feature interaction representation and parameter distribution, the system outputs the user's behavior prediction results for the content.

[0014] Optionally, based on the behavior prediction results, the monitoring and analysis of user interests, reading behavior patterns, and content preferences, and the dynamic updating of tag distribution in the user-content-behavior relationship network, to achieve continuous monitoring and intelligent analysis of APP user reading data, includes: Based on the predicted results of user behavior towards content, we analyze users' reading behavior patterns, changes in interests, and content preferences to identify users' behavioral characteristics and interest trends in different time periods or scenarios. By comparing the analysis results with historical behavioral data, we can continuously track and update changes in user interests and content preferences, assess the dynamic changes in user behavior, and discover potential points of interest or changes in behavioral patterns. Based on real-time monitoring of user behavior and content preferences, and utilizing updated tag information, the tag distribution in the user-content-behavior relationship network is dynamically adjusted. When a user's interests or behavior change, the corresponding tag distribution will be adjusted in real time to maintain the accuracy of tag allocation. By combining user interests and behavioral changes, the node tags in the user-content-behavior relationship network are continuously updated through a tag propagation algorithm. For new user behaviors and interests, the corresponding tag information is automatically inferred and added to the tag propagation process. Regularly evaluate the effectiveness of tag updates and dissemination results, and monitor the relationship between tag distribution and changes in user behavior; Based on the optimized tag distribution, we continuously monitor and analyze user interests, behavioral patterns, and content preferences, and provide real-time data support for content recommendation.

[0015] According to an embodiment of the present invention, an APP user reading data monitoring and analysis system based on big data includes the following modules: The data acquisition and preprocessing module is used to collect reading behavior data of APP users, preprocess the reading behavior data, and generate standardized datasets. The User-Content-Behavior Relationship Network Construction Module is used to build a user-content-behavior relationship network based on a standardized dataset, connecting users and content nodes through reading behavior to form a heterogeneous behavior network; The tag assignment and propagation module is used to assign initial tags to some nodes based on interest tags or content topic information, and apply a tag propagation algorithm to propagate the tag information to unlabeled nodes to obtain the tag distribution results of all nodes; The label feature extraction and merging module is used to extract label features from the label distribution results and merge them with other features in the standardized dataset to form a high-dimensional sparse feature set. The factorization machine analysis module is used to perform factorization machine analysis on high-dimensional sparse feature sets, model the second-order and higher-order interaction relationships of user, content, behavior and tag features, and output the prediction results of user behavior towards content. The user interest and behavior monitoring module is used to continuously monitor users' interests, reading behavior patterns and content preferences based on behavior prediction results, and dynamically update the tag distribution. The tag update and feedback module is used to dynamically update tags in the user-content-behavior relationship network based on monitoring results, supporting personalized recommendations and predictive optimization.

[0016] The beneficial effects of this invention are: This invention effectively solves some key problems in existing recommender systems by combining a label propagation algorithm and a factorization machine. By constructing a user-content-behavior relationship network, it more accurately captures complex patterns in user behavior data, enabling in-depth analysis of user interests, reading behavior patterns, and content preferences. Unlike traditional recommender methods, this invention dynamically updates the label distribution in the user-content-behavior relationship network, reflecting changes in user interests in real time. The recommender system can flexibly adapt to changes in user preferences, rather than being limited to static historical user behavior data.

[0017] This invention utilizes factorization machines to jointly analyze high-dimensional sparse feature sets, successfully modeling second-order and higher-order interaction relationships among user, content, behavior, and tag features, thereby improving prediction accuracy. The use of factorization machines enables the model to handle data sparsity issues, avoiding the cold-start and recommendation accuracy problems common in traditional collaborative filtering methods. By dynamically adjusting tag distribution and continuously optimizing feature interactions, this invention allows the recommendation system to continuously learn and optimize, ensuring more accurate personalized recommendation results.

[0018] This invention improves the accuracy and real-time performance of recommendations, enhances the ability of personalized recommendations, and is particularly adaptable to rapidly changing user interests and preferences. It enables more accurate prediction and analysis of APP users' reading behavior, ultimately achieving more efficient and accurate content recommendations and providing users with a better user experience. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for monitoring and analyzing APP user reading data based on big data, as proposed in this invention. Figure 2 This is a schematic diagram of the structure of an APP user reading data monitoring and analysis system based on big data proposed in this invention. Detailed Implementation

[0020] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0021] refer to Figure 1 A method for monitoring and analyzing APP user reading data based on big data includes: Collect reading behavior data from APP users, preprocess the reading behavior data, and generate a standardized dataset; Based on a standardized dataset, a user-content-behavior relationship network is constructed, with user nodes and content nodes as vertices and user reading behavior as edges, forming a heterogeneous behavior network structure. For some nodes in the user-content-behavior relationship network, initial labels are assigned based on existing interest tags or content topic information. A label propagation algorithm is then applied to iteratively propagate the initial label information to unlabeled nodes through the edge relationships between nodes, resulting in the label distribution of all nodes. Based on the label distribution results, the label features of each user node and content node are extracted, and the label features are merged with the relevant features in the standardized dataset to form a high-dimensional sparse feature set. A joint analysis of high-dimensional sparse feature sets is performed using factorization machine. The factorization machine is used to jointly model the second-order and higher-order feature interaction relationships between user, content, behavior and tag features, and output the prediction results of user behavior towards content. Based on the behavioral prediction results, user interests, reading behavior patterns, and content preferences are monitored and analyzed, and the tag distribution in the user-content-behavior relationship network is dynamically updated to achieve continuous monitoring and intelligent analysis of APP user reading data.

[0022] In this embodiment, the reading behavior data specifically includes the user's browsing, clicking, collecting, sharing, and commenting operations on APP content, as well as the corresponding time information and content identifiers.

[0023] In this embodiment, the preprocessing of reading behavior data specifically includes deduplication, missing value filling, outlier removal, and standardization of the collected reading behavior data.

[0024] In this embodiment, the construction of a user-content-behavior relationship network based on a standardized dataset, with user nodes and content nodes as vertices and user reading behavior as edges, forms a heterogeneous behavior network structure, including: In the standardized dataset, all user identifiers, content identifiers, and information on the types and timing of reading behaviors between users and content are identified and extracted to generate a list of basic information containing users, content, and behaviors. Based on the list of basic information, establish a set of user nodes and a set of content nodes respectively, and assign each user identifier to a user node and each content identifier to a content node; Traverse the list of basic information, and based on each user's reading behavior, establish an edge between the corresponding user node and content node, and mark the type and time of the reading behavior on the edge; Collect all reading behavior records between each user node and content node, and assign weights to the edges between user nodes and content nodes based on different types of reading behaviors and their frequency. The weights are determined comprehensively based on the behavior type and frequency. The weights of the edges are adjusted based on the time of the behavior, with the weights of historical behaviors that are far removed from the current time being reduced and the weights of recent behaviors being increased, reflecting the user's latest interests and behavioral activity. By utilizing the established set of user nodes, set of content nodes, and weighted user-content edges, a user-content-behavior relationship network is constructed.

[0025] In this embodiment, the initial label assignment for some nodes in the user-content-behavior relationship network based on existing interest tags or content topic information, and the application of a label propagation algorithm to iteratively propagate the initial label information to unlabeled nodes through the edge relationships between nodes, to obtain the label distribution result of all nodes, includes: In the user-content-behavior relationship network, some nodes with known interest tags or content topic information are identified and designated as labeled nodes. Each labeled node is assigned an initial label distribution result and a label confidence of 1, while the remaining unlabeled nodes are assigned an empty label distribution result and a confidence of 0. For each edge between a user node and a content node, a temporal decay weight is assigned to each edge based on the time when the user's reading behavior occurs. The closer the behavior is to the current time, the higher the weight is assigned, and the farther the behavior is to the current time, the lower the weight is assigned, forming an edge weight set containing temporal information. During the label propagation iteration process, each unlabeled node collects the current label and confidence information of all neighboring nodes, and combines the temporal decay weight of the neighboring edges to perform weighted score statistics on all labels; The label with the highest weighted score is used as the new label for the current unlabeled node. The confidence of the node's current label is updated based on the confidence of the neighboring node labels involved in the calculation and the temporal decay weight of the edge. The confidence is taken as the weighted average. A diversity test is performed on the new labels and confidence scores assigned to each node. If the node label distribution is too concentrated on a single label, the second-best label with the highest confidence score is introduced to adjust the label distribution of the node. After each iteration, the label distribution of all nodes is statistically analyzed. If a node with an abnormally concentrated label distribution is found, the label distribution of the node is automatically adjusted based on the node's historical label distribution and the diversity of neighboring node labels, and self-correction is performed. Repeat the propagation of label distribution results, confidence updates, diversity adjustments and self-corrections until the label distribution of all nodes remains stable or the preset number of iterations is reached. Finally, output the label distribution results and corresponding confidence scores of all nodes after convergence.

[0026] In this embodiment, the step of extracting tag features for each user node and content node based on the tag distribution results, and merging the tag features with relevant features in the standardized dataset to form a high-dimensional sparse feature set, includes: From the tag distribution results, tag features are extracted for each user node and each content node. The tag features include the distribution of the interest category to which the node belongs and the distribution of the content topic. In the standardized dataset, for each pair of user and content behavior records, user behavior features, content features and contextual features of the behavior are extracted. The behavior features include user activity, historical behavior type and frequency, the content features include content category and content popularity, and the contextual features include the time of the behavior and device type. The obtained tag features are concatenated with user behavior features, content features, and behavior context features to form a joint feature vector for each user-content behavior relationship; For all user and content behavior records, generate joint feature vectors sequentially, and combine all feature vectors to form a high-dimensional sparse feature set; The high-dimensional sparse feature set is indexed and organized, and the organized high-dimensional sparse feature set is output.

[0027] In this embodiment, the joint analysis of the high-dimensional sparse feature set using factorization machine (FDM) is performed. The FDM is used to jointly model the second-order and higher-order feature interaction relationships between user, content, behavior, and tag features, outputting user behavior prediction results for the content, including: The obtained high-dimensional sparse feature set is used as input. The high-dimensional sparse feature set includes user features, content features, behavior features and tag features. For each pair of features in the high-dimensional sparse feature set, a learnable weight parameter is assigned. The weight parameter is adjusted together with the factorization-related parameters during the parameter update process to reflect the importance of different features to the interaction. Different factor vector dimensions are set for user features, content features, behavior features, and tag features in the high-dimensional sparse feature set, and a factor vector of corresponding length is used for each category feature; When performing interactive computation on a high-dimensional sparse feature set, the factor vector of the corresponding dimension is selected according to the feature category. Then, second-order and higher-order interactive operations are performed between the factor vector and the factor vector of other categories of features to form the interactive expression result between features. The interactive result is stored as an interactive feature vector. Based on the obtained interaction feature vector, it is fused with user features, content features, behavioral features and tag features in the high-dimensional sparse feature set to form a fused feature set, which contains the interaction information between all features; For each pair of feature interactions in the fused feature set, the impact on the final behavior prediction result is calculated. If the interaction contribution of a certain feature pair is lower than the set threshold, the corresponding factor vector parameter is set to zero, and the invalid feature pair is removed from the feature interaction calculation. Repeatedly perform feature interaction calculation, factor vector update, and low-impact feature pair removal operations to continuously optimize feature interaction expression and parameter distribution, so that the effective interaction between features is fully expressed and redundant and invalid feature pairs are effectively removed. Based on the optimized feature interaction representation and parameter distribution, the system outputs the user's behavior prediction results for the content.

[0028] In this embodiment, the monitoring and analysis of user interests, reading behavior patterns, and content preferences based on behavior prediction results, and the dynamic updating of tag distribution in the user-content-behavior relationship network, to achieve continuous monitoring and intelligent analysis of APP user reading data, includes: Based on the predicted results of user behavior towards content, we analyze users' reading behavior patterns, changes in interests, and content preferences to identify users' behavioral characteristics and interest trends in different time periods or scenarios. By comparing the analysis results with historical behavioral data, we can continuously track and update changes in user interests and content preferences, assess the dynamic changes in user behavior, and discover potential points of interest or changes in behavioral patterns. Based on real-time monitoring of user behavior and content preferences, and utilizing updated tag information, the tag distribution in the user-content-behavior relationship network is dynamically adjusted. When a user's interests or behavior change, the corresponding tag distribution will be adjusted in real time to maintain the accuracy of tag allocation. By combining user interests and behavioral changes, the node tags in the user-content-behavior relationship network are continuously updated through a tag propagation algorithm. For new user behaviors and interests, the corresponding tag information is automatically inferred and added to the tag propagation process. Regularly evaluate the effectiveness of tag updates and dissemination results, and monitor the relationship between tag distribution and changes in user behavior; Based on the optimized tag distribution, we continuously monitor and analyze user interests, behavioral patterns, and content preferences, and provide real-time data support for content recommendation.

[0029] refer to Figure 2 A big data-based APP user reading data monitoring and analysis system includes the following modules: The data acquisition and preprocessing module is used to collect reading behavior data of APP users, preprocess the reading behavior data, and generate standardized datasets. The User-Content-Behavior Relationship Network Construction Module is used to build a user-content-behavior relationship network based on a standardized dataset, connecting users and content nodes through reading behavior to form a heterogeneous behavior network; The tag assignment and propagation module is used to assign initial tags to some nodes based on interest tags or content topic information, and apply a tag propagation algorithm to propagate the tag information to unlabeled nodes to obtain the tag distribution results of all nodes; The label feature extraction and merging module is used to extract label features from the label distribution results and merge them with other features in the standardized dataset to form a high-dimensional sparse feature set. The factorization machine analysis module is used to perform factorization machine analysis on high-dimensional sparse feature sets, model the second-order and higher-order interaction relationships of user, content, behavior and tag features, and output the prediction results of user behavior towards content. The user interest and behavior monitoring module is used to continuously monitor users' interests, reading behavior patterns and content preferences based on behavior prediction results, and dynamically update the tag distribution. The tag update and feedback module is used to dynamically update tags in the user-content-behavior relationship network based on monitoring results, supporting personalized recommendations and predictive optimization. Example 1:

[0030] To verify the feasibility of this invention in practice, it was applied to a mobile reading app. The platform needs to accurately push massive amounts of different types of content, such as novels, news, columns, and comics, to millions of users daily. Previously, most platforms used collaborative filtering-based recommendation methods, which, while able to uncover interest connections between users to some extent, are highly sensitive to data sparsity and have poor adaptability to the cold start of new users or new content. Traditional methods struggle to track dynamic changes in user interests in a timely manner, causing recommended content to lag behind users' actual needs, thus reducing platform activity and user satisfaction.

[0031] To address the aforementioned issues, this invention provides a big data-based method for monitoring and analyzing APP user reading data. In practical applications, the platform continuously collects daily user behavior data, including reading, clicking, collecting, sharing, and commenting, through the client. After preprocessing, a standardized dataset is generated, constructing a user-content-behavior relationship network. Using tag information from core users and popular content, a tag propagation algorithm is employed to dynamically spread interests and topics, promptly capturing changes in user interests across the entire network. The tag features of all user and content nodes are fused with standardized behavioral features to generate a high-dimensional sparse feature set. A factorization machine is used to jointly analyze this feature set, deeply modeling the interaction relationships between users, content, behavior, and tag features, predicting each user's reading probability and interest tendencies for various types of content. The system automatically adjusts its content push strategy daily based on the prediction results and dynamically updates user interest tags, achieving real-time tracking of user interests and accurate content recommendation.

[0032] In the actual operation test in April 2025, the platform randomly selected 300,000 users in a certain region and divided them into a control group (using the traditional collaborative filtering method) and an experimental group (using the method of this invention). During the 30-day comparative experiment, the system recorded key indicators for both groups of users, including average daily recommendation click-through rate, average reading time, interest tag response speed, new user retention, and cold start recommendation effect of new content.

[0033] Table 1 Comparison of key indicators between the method of this invention and the traditional method

[0034] As clearly shown in Table 1, the method of this invention demonstrates comprehensive superiority over traditional collaborative filtering techniques in mobile reading recommendation scenarios. Regarding the core recommendation click-through rate (CTR), the experimental group achieved 14.3%, a significant improvement over the control group's 10.2%, indicating that the personalized recommendations implemented through this invention can more accurately match user interests and increase users' actual willingness to click on recommended content. In terms of increasing user activity, the experimental group's average daily reading time was 24.7 minutes, nearly 4 minutes longer than the control group, indicating a significant enhancement in users' content consumption enthusiasm and platform stickiness under the recommendation mechanism of this invention.

[0035] Regarding the dynamic responsiveness of user interest tags, the experimental group's average update cycle for user interest tags was only 2.3 days, while the control group's was 5.2 days. This demonstrates that the platform can capture and reflect changes in user interests more quickly, adjusting recommendation strategies in real time and avoiding content delivery delays. New user retention has also been significantly improved, with the experimental group achieving a 7-day retention rate of 39.5%, nearly 11 percentage points higher than the control group. This fully demonstrates the outstanding effectiveness of this invention in solving the challenges of cold start and new user interest identification.

[0036] The method of this invention also demonstrates excellent performance in promoting new content and improving user experience. The average cold start click-through rate for new content increased from 7.2 to 11.1, significantly enhancing the exposure and appeal of new content in the experimental group. This contributes to the healthy cycle of the platform's content ecosystem. The proportion of duplicate recommended content decreased from 22.7% to 13.2%, and the user complaint rate decreased from 1.01% to 0.89%. The method of this invention reduces content homogenization and user dissatisfaction, further improving overall user satisfaction and activity.

[0037] From recommendation accuracy, user stickiness, interest capture, cold start performance to platform satisfaction, the method of this invention outperforms traditional collaborative filtering recommendation technology, improving the platform's intelligent recommendation capabilities and user service level.

[0038] 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 method for monitoring and analyzing APP user reading data based on big data, characterized in that, include: Collect reading behavior data from APP users, preprocess the reading behavior data, and generate a standardized dataset; Based on a standardized dataset, a user-content-behavior relationship network is constructed, with user nodes and content nodes as vertices and user reading behavior as edges, forming a heterogeneous behavior network structure. For some nodes in the user-content-behavior relationship network, initial labels are assigned based on existing interest tags or content topic information. A label propagation algorithm is then applied to iteratively propagate the initial label information to unlabeled nodes through the edge relationships between nodes, resulting in the label distribution of all nodes. Based on the label distribution results, the label features of each user node and content node are extracted, and the label features are merged with the relevant features in the standardized dataset to form a high-dimensional sparse feature set. A joint analysis of high-dimensional sparse feature sets is performed using factorization machine. The factorization machine is used to jointly model the second-order and higher-order feature interaction relationships between user, content, behavior and tag features, and output the prediction results of user behavior towards content. Based on the behavioral prediction results, user interests, reading behavior patterns, and content preferences are monitored and analyzed, and the tag distribution in the user-content-behavior relationship network is dynamically updated to achieve continuous monitoring and intelligent analysis of APP user reading data.

2. The method for monitoring and analyzing APP user reading data based on big data according to claim 1, characterized in that, The reading behavior data specifically includes users' browsing, clicking, collecting, sharing, and commenting actions on APP content, as well as the corresponding time information and content identifiers.

3. The method for monitoring and analyzing APP user reading data based on big data according to claim 1, characterized in that, The preprocessing of reading behavior data specifically includes deduplication, missing value imputation, outlier removal, and standardization of the collected reading behavior data.

4. The method for monitoring and analyzing APP user reading data based on big data according to claim 1, characterized in that, Based on a standardized dataset, a user-content-behavior relationship network is constructed, with user nodes and content nodes as vertices and user reading behavior as edges, forming a heterogeneous behavior network structure, including: In the standardized dataset, all user identifiers, content identifiers, and information on the types and timing of reading behaviors between users and content are identified and extracted to generate a list of basic information containing users, content, and behaviors. Based on the list of basic information, establish a set of user nodes and a set of content nodes respectively, and assign each user identifier to a user node and each content identifier to a content node; Traverse the list of basic information, and based on each user's reading behavior, establish an edge between the corresponding user node and content node, and mark the type and time of the reading behavior on the edge; Collect all reading behavior records between each user node and content node, and assign weights to the edges between user nodes and content nodes based on different types of reading behaviors and their frequency. The weights are determined comprehensively based on the behavior type and frequency. The weights of the edges are adjusted based on the time of the behavior, with the weights of historical behaviors that are far removed from the current time being reduced and the weights of recent behaviors being increased, reflecting the user's latest interests and behavioral activity. By utilizing the established set of user nodes, set of content nodes, and weighted user-content edges, a user-content-behavior relationship network is constructed.

5. The method for monitoring and analyzing APP user reading data based on big data according to claim 1, characterized in that, The method involves assigning initial labels to some nodes in the user-content-behavior relationship network based on existing interest tags or content topic information. Then, a label propagation algorithm is applied to iteratively propagate the initial label information to unlabeled nodes through the edge relationships between nodes, resulting in the label distribution of all nodes, including: In the user-content-behavior relationship network, some nodes with known interest tags or content topic information are identified and designated as labeled nodes. Each labeled node is assigned an initial label distribution result and a label confidence of 1, while the remaining unlabeled nodes are assigned an empty label distribution result and a confidence of 0. For each edge between a user node and a content node, a temporal decay weight is assigned to each edge based on the time when the user's reading behavior occurs. The closer the behavior is to the current time, the higher the weight is assigned, and the farther the behavior is to the current time, the lower the weight is assigned, forming an edge weight set containing temporal information. During the label propagation iteration process, each unlabeled node collects the current label and confidence information of all neighboring nodes, and combines the temporal decay weight of the neighboring edges to perform weighted score statistics on all labels; The label with the highest weighted score is used as the new label for the current unlabeled node. The confidence of the node's current label is updated based on the confidence of the neighboring node labels involved in the calculation and the temporal decay weight of the edge. The confidence is taken as the weighted average. A diversity test is performed on the new labels and confidence scores assigned to each node. If the node label distribution is too concentrated on a single label, the second-best label with the highest confidence score is introduced to adjust the label distribution of the node. After each iteration, the label distribution of all nodes is statistically analyzed. If a node with an abnormally concentrated label distribution is found, the label distribution of the node is automatically adjusted based on the node's historical label distribution and the diversity of neighboring node labels, and self-correction is performed. Repeat the propagation of label distribution results, confidence updates, diversity adjustments and self-corrections until the label distribution of all nodes remains stable or the preset number of iterations is reached. Finally, output the label distribution results and corresponding confidence scores of all nodes after convergence.

6. The method for monitoring and analyzing APP user reading data based on big data according to claim 1, characterized in that, Based on the label distribution results, label features are extracted for each user node and content node, and these label features are merged with relevant features in the standardized dataset to form a high-dimensional sparse feature set, including: From the tag distribution results, tag features are extracted for each user node and each content node. The tag features include the distribution of the interest category to which the node belongs and the distribution of the content topic. In the standardized dataset, for each pair of user and content behavior records, user behavior features, content features and contextual features of the behavior are extracted. The behavior features include user activity, historical behavior type and frequency, the content features include content category and content popularity, and the contextual features include the time of the behavior and device type. The obtained tag features are concatenated with user behavior features, content features, and behavior context features to form a joint feature vector for each user-content behavior relationship. For all user and content behavior records, generate joint feature vectors sequentially, and combine all feature vectors to form a high-dimensional sparse feature set; The high-dimensional sparse feature set is indexed and organized, and the organized high-dimensional sparse feature set is output.

7. The method for monitoring and analyzing APP user reading data based on big data according to claim 1, characterized in that, The high-dimensional sparse feature set is jointly analyzed using factorization machine. Factorization machine is used to jointly model the second-order and higher-order feature interaction relationships among user, content, behavior, and tag features, outputting user behavior prediction results for the content, including: The obtained high-dimensional sparse feature set is used as input. The high-dimensional sparse feature set includes user features, content features, behavior features and tag features. For each pair of features in the high-dimensional sparse feature set, a learnable weight parameter is assigned. The weight parameter is adjusted together with the factorization-related parameters during the parameter update process to reflect the importance of different features to the interaction. Different factor vector dimensions are set for user features, content features, behavior features, and tag features in the high-dimensional sparse feature set, and a factor vector of corresponding length is used for each category feature; When performing interactive computation on a high-dimensional sparse feature set, the factor vector of the corresponding dimension is selected according to the feature category. Then, second-order and higher-order interactive operations are performed between the factor vector and the factor vector of other categories of features to form the interactive expression result between features. The interactive result is stored as an interactive feature vector. Based on the obtained interaction feature vector, it is fused with user features, content features, behavioral features and tag features in the high-dimensional sparse feature set to form a fused feature set, which contains the interaction information between all features; For each pair of feature interactions in the fused feature set, the impact on the final behavior prediction result is calculated. If the interaction contribution of a certain feature pair is lower than the set threshold, the corresponding factor vector parameter is set to zero, and the invalid feature pair is removed from the feature interaction calculation. Repeatedly perform feature interaction calculation, factor vector update, and low-impact feature pair removal operations to continuously optimize feature interaction expression and parameter distribution, so that the effective interaction between features is fully expressed and redundant and invalid feature pairs are effectively removed. Based on the optimized feature interaction representation and parameter distribution, the system outputs the user's behavior prediction results for the content.

8. The method for monitoring and analyzing APP user reading data based on big data according to claim 1, characterized in that, Based on the behavior prediction results, the system monitors and analyzes user interests, reading behavior patterns, and content preferences, and dynamically updates the tag distribution in the user-content-behavior relationship network to achieve continuous monitoring and intelligent analysis of APP user reading data, including: Based on the predicted results of user behavior towards content, we analyze users' reading behavior patterns, changes in interests, and content preferences to identify users' behavioral characteristics and interest trends in different time periods or scenarios. By comparing the analysis results with historical behavioral data, we can continuously track and update changes in user interests and content preferences, assess the dynamic changes in user behavior, and discover potential points of interest or changes in behavioral patterns. Based on real-time monitoring of user behavior and content preferences, and utilizing updated tag information, the tag distribution in the user-content-behavior relationship network is dynamically adjusted. When a user's interests or behavior change, the corresponding tag distribution will be adjusted in real time to maintain the accuracy of tag allocation. By combining user interests and behavioral changes, the node tags in the user-content-behavior relationship network are continuously updated through a tag propagation algorithm. For new user behaviors and interests, the corresponding tag information is automatically inferred and added to the tag propagation process. Regularly evaluate the effectiveness of tag updates and dissemination results, and monitor the relationship between tag distribution and changes in user behavior; Based on the optimized tag distribution, we continuously monitor and analyze user interests, behavioral patterns, and content preferences, and provide real-time data support for content recommendation.

9. A big data-based APP user reading data monitoring and analysis system, comprising executing the big data-based APP user reading data monitoring and analysis method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to collect reading behavior data of APP users, preprocess the reading behavior data, and generate standardized datasets. The User-Content-Behavior Relationship Network Construction Module is used to build a user-content-behavior relationship network based on a standardized dataset, connecting users and content nodes through reading behavior to form a heterogeneous behavior network; The tag assignment and propagation module is used to assign initial tags to some nodes based on interest tags or content topic information, and apply a tag propagation algorithm to propagate the tag information to unlabeled nodes to obtain the tag distribution results of all nodes; The label feature extraction and merging module is used to extract label features from the label distribution results and merge them with other features in the standardized dataset to form a high-dimensional sparse feature set. The factorization machine analysis module is used to perform factorization machine analysis on high-dimensional sparse feature sets, model the second-order and higher-order interaction relationships of user, content, behavior and tag features, and output the prediction results of user behavior towards content. The user interest and behavior monitoring module is used to continuously monitor users' interests, reading behavior patterns and content preferences based on behavior prediction results, and dynamically update the tag distribution. The tag update and feedback module is used to dynamically update tags in the user-content-behavior relationship network based on monitoring results, supporting personalized recommendations and predictive optimization.

Images