A method and system for analyzing user behavior data in a BOSS system

Abstract

The application relates to the technical field of data processing technology and discloses a method and system for analyzing user behavior data in a BOSS system, which comprises the following steps: collecting multi-source behavior data of a user in the BOSS system; generating a dynamic feature set containing at least role weight features and behavior attenuation features according to the user role, the business stage and real-time behavior; according to an adaptive weight distribution algorithm based on a reinforcement learning framework, the weighted fusion of the dynamic features is optimized in real time according to the user retention rate and the path conversion rate; based on the fused features, user grouping, path optimization and loss prediction are performed through collaborative analysis of a graph neural network and a multi-task learning model; and the analysis result is pushed to the BOSS system in real time through a stream processing engine to trigger interface permission adjustment or resource pushing, thereby enhancing the refinement degree of business decision-making and significantly improving the user retention, conversion efficiency and adaptive ability of system operation through a real-time closed loop.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing technology, in particular to a user behavior data analysis method and system in a BOSS system. BACKGROUND

[0002] In a business operation support system (BOSS system), user behavior data analysis is a core link for optimizing system performance, improving user experience, and ensuring business security. As a key infrastructure for enterprise operation, the BOSS system is responsible for handling user registration, service subscription, billing management, fault reporting, and other full-process operations. Its data flow is complex and large in scale, containing multi-dimensional information such as user login frequency, operation path, service request mode, and abnormal behavior. However, traditional user behavior analysis methods face many challenges in dealing with the complexity of the BOSS system.

[0003] Firstly, traditional analysis methods mainly rely on log statistics and fixed rule engines, such as extracting user operation logs through ETL tools and generating static reports. Although this method can provide a basic overview of the data, it has significant limitations: first, the data processing process has high latency, which cannot achieve real-time analysis, resulting in a lag in responding to sudden abnormalities such as account theft or bulk fraud; second, the analysis dimension is single, making it difficult to capture the dynamic patterns of user behavior (such as cross-service operation correlation), resulting in the lack of key business insights; third, it relies on manually defined rules, which are difficult to adapt to rapid changes in business scenarios, such as new service launches or user behavior pattern shifts.

[0004] Secondly, with the development of big data technology, some systems have begun to introduce machine learning and data mining techniques. For example, through clustering analysis to identify user group characteristics, or using time series prediction models to predict user churn risk. However, existing technologies still have bottlenecks in the application of BOSS systems: on the one hand, data sources are scattered in multiple subsystems (such as billing modules, customer service systems, and background management), making data integration difficult, resulting in biased analysis results; on the other hand, model training relies on historical data, which cannot quickly and accurately respond to user behavior sequences in a timely manner. SUMMARY

[0005] In order to provide a user behavior data analysis method in a BOSS system that can analyze multi-dimensional user behavior and respond quickly and accurately, the present application provides a user behavior data analysis method and system in a BOSS system.

[0006] In a first aspect, the present application provides a user behavior data analysis method in a BOSS system, which adopts the following technical solution:

[0007] A user behavior data analysis method in a BOSS system, comprising:

[0008] Collecting multi-source behavior data of the user in the BOSS system; the multi-source behavior data includes operation logs, business database records, third-party system data, and user role attributes;

[0009] According to the user role, the business stage, and the real-time behavior, a dynamic feature set containing at least role weight features and behavior attenuation features is generated;

[0010] According to an adaptive weight allocation algorithm based on a reinforcement learning framework, the weighted fusion of dynamic features is optimized in real time according to the user retention rate and the path conversion rate;

[0011] Based on the fused features, user clustering, path optimization, and churn prediction are performed through collaborative analysis of a graph neural network and a multi-task learning model;

[0012] The analysis results are pushed to the BOSS system in real time through a stream processing engine to trigger interface permission adjustment or resource pushing.

[0013] By adopting the above technical solutions, by integrating operation logs, business data, and third-party information, the system can comprehensively capture multi-dimensional features of user behavior, especially the dynamic changes of role attributes and business stages, so that the feature description is both accurate and timely, avoiding the one-sidedness caused by data fragmentation in traditional analysis. Secondly, the adaptive weight allocation mechanism based on reinforcement learning can dynamically adjust the feature fusion strategy according to the user retention and conversion target. For example, in the high-risk scenario of user payment failure, the system will automatically strengthen the weight of recent behavior or adjust the influence of role privileges on decision-making, so that the model responds more closely to actual needs in different business stages, reducing the limitations of artificial preset parameters.

[0014] By modeling user behavior sequences through a graph neural network, the system can mine the temporal correlation and business logic relationship between behaviors, such as identifying the abnormal path of "multiple payment cancellations-returning to the shopping cart", so as to more accurately locate user churn risk or optimize the recommendation path. The multi-task learning framework further collaboratively trains the clustering, path optimization, and prediction tasks, avoiding the one-sidedness of a single model, such as identifying high-value users while simultaneously generating personalized interface adjustment strategies.

[0015] Finally, real-time analysis and business system linkage are realized through a stream processing engine, so that strategy adjustment and user behavior are almost synchronized. For example, when the model detects that the user has a churn risk, it can trigger the pushing of coupons or interface simplification operations in real time. This closed-loop mechanism not only improves the user's immediate experience, but also optimizes the model through continuous data feedback, forming a "analysis-decision-verification" continuous iteration capability, reducing the cost of business intervention.

[0016] Overall, the user behavior analysis is upgraded from traditional offline statistics to a dynamic perception and intelligent response proactive strategy system, which not only enhances the refinement of business decision-making, but also significantly improves user retention, conversion efficiency and system operation adaptive ability through real-time closed loop.

[0017] Optionally, the role weight feature The generation process comprises:

[0018] ;

[0019] ;

[0020] ;

[0021] wherein, is an adaptive role weight determined by the adaptive weight allocation algorithm, is a business stage dynamic weight, is a user active dynamic weight, is a total number of stages, is a stage factor of the i-th stage, is a user value, is a user churn rate, is an average user value, is the number of times the user's finger slides in a specified product interface within a specified time length, is the total time length of staying in a specified product interface, is the time taken by the user to complete the target behavior, is a user portrait label quantization factor.

[0022] Optionally, the generation process of the behavior decay feature comprises:

[0023] ;

[0024] ;

[0025] wherein, is a segmented decay function, is a behavior type weight, is a time demarcation point, is a recent decay coefficient, is a long-term decay coefficient, and are time decay coefficients determined by the adaptive weight allocation algorithm, is a business type sensitive coefficient, .

[0026] ​By adopting the technical scheme, the limitations of the traditional static role label can be overcome, the weight distribution can be flexibly adjusted according to real-time business targets (for example, VIP user privileges are focused on during a promotion period, and high-risk role monitoring is strengthened during a risk control period), the model can focus on key role attributes that truly affect user behavior, and the synergistic effect of role weight and behavior decay characteristics can effectively capture the influence of role changes on behavior.

[0027] Optionally, the adaptive weight distribution algorithm comprises:

[0028] a state space S is defined with a user role and a business stage as state inputs;

[0029] an action space A is defined as an adjustment range of the weight parameter;

[0030] a reward function is set :

[0031] ;

[0032] wherein, a scenario factor related to a business scenario, a user retention rate, indicating a probability that a user does not flow away within a TC hour after the weight is adjusted, a path conversion rate, indicating a proportion of users completing a target behavior in an optimized path, a penalty coefficient weight, a penalty coefficient.

[0033] Optionally, the graph neural network adopts a graph attention network, aggregates neighbor node information through a self-attention mechanism, and extracts high-order features of user behavior, wherein the neighbor node is composed of a time edge and a logical edge.

[0034] Optionally, the multi-task learning model shares a bottom feature representation of the graph attention network, and simultaneously outputs user clustering, path optimization and flow prediction.

[0035] Optionally, comprising:

[0036] each user behavior is a node;

[0037] time edge establishment: if two user behaviors occur continuously within a preset time interval, a time edge is established;

[0038] logical edge establishment based on logical association of a business process;

[0039] neighbor nodes include time neighbors and logical neighbors; the time neighbors represent all user behaviors connected to the node i through the time edge, and the logical neighbors represent all user behaviors connected through the logical edge.

[0040] Optionally, the stream processing engine is implemented by Apache Flink, and the condition triggering the strategy adjustment includes that the user enters a high churn risk group or a business path conversion rate is lower than a preset threshold.

[0041] In a second aspect, the application provides a user behavior data analysis system in a BOSS system, which adopts the following technical solution:

[0042] A user behavior data analysis system in a BOSS system, comprising:

[0043] A data acquisition module for acquiring multi-source behavior data of a user in a BOSS system; the multi-source behavior data includes operation logs, business database records, third-party system data and user role attributes;

[0044] A dynamic weight generation module for generating a dynamic feature set containing at least role weight features and behavior decay features according to user roles, business stages and real-time behaviors;

[0045] A fusion weighting module for using a reinforcement learning framework according to an adaptive weight distribution algorithm to real-time optimize the weighted fusion of dynamic features according to user retention rates and path conversion rates;

[0046] A collaborative analysis module for collaborative analysis based on the fused features through a graph neural network and a multi-task learning model to perform user grouping, path optimization and churn prediction;

[0047] A push module for real-time pushing the analysis results to the BOSS system through a stream processing engine to trigger interface permission adjustment or resource pushing.

[0048] In summary, the application includes at least one of the following beneficial technical effects:

[0049] By integrating operation logs, business data and third-party information, the system can comprehensively capture multi-dimensional features of user behavior, especially the dynamic changes of role attributes and business stages, so that the feature description is not only accurate but also time-effective, avoiding the one-sidedness caused by data fragmentation in traditional analysis. Secondly, the adaptive weight distribution mechanism based on reinforcement learning can dynamically adjust the feature fusion strategy according to user retention and conversion targets. By modeling user behavior sequences through a graph neural network, the system can mine the temporal correlation and business logic relationship between behaviors, thereby more accurately positioning user churn risks or optimizing recommended paths. The multi-task learning framework further collaboratively trains the grouping, path optimization and prediction tasks, avoiding the one-sidedness of a single model; finally, the real-time analysis and business system linkage are realized through a stream processing engine, not only improving the user's immediate experience, but also optimizing the model through continuous data feedback, reducing the cost of business intervention. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a flowchart of a method for analyzing user behavior data in a BOSS system. DETAILED DESCRIPTION

[0051] Embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings.

[0052] In the description of the present specification, the description of the terms "certain embodiments", "one embodiment", "some embodiments", "illustrative embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the described embodiments or examples are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily mean the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0053] The embodiments of the present application disclose a method for analyzing user behavior data in a BOSS system, referring to Figure 1 A method for analyzing user behavior data in a BOSS system, comprising:

[0054] Collecting multi-source behavior data of users in the BOSS system; the multi-source behavior data includes operation logs, business database records, third-party system data and user role attributes;

[0055] According to the user role (such as VIP level, permission type), the business stage (such as registration process, promotion period, system maintenance, payment confirmation, complaint handling, etc.) and real-time behavior (such as specific operation and operation type, timestamp, page path), a dynamic feature set containing at least role weight feature and behavior decay feature is generated;

[0056] According to the adaptive weight allocation algorithm based on the reinforcement learning framework, the weighted fusion of the dynamic features is optimized in real time according to the user retention rate and the path conversion rate;

[0057] Based on the fused features, through the cooperation of the graph neural network and the multi-task learning model, user clustering, path optimization and loss prediction are performed;

[0058] The analysis results are pushed to the BOSS system in real time through the stream processing engine, triggering interface permission adjustment or resource pushing.

[0059] By adopting the above technical solutions and integrating operation logs, business data, and third-party information, the system can comprehensively capture multi-dimensional characteristics of user behavior, especially the dynamic changes in role attributes and business stages. This makes the feature description both accurate and timely, avoiding the one-sidedness caused by data fragmentation in traditional analysis. Secondly, the adaptive weight allocation mechanism based on reinforcement learning can dynamically adjust the feature fusion strategy according to user retention and conversion goals. For example, in high-risk scenarios where users fail to pay, the system will automatically strengthen the weight of recent behavior or adjust the impact of role privileges on decision-making, making the model's response at different business stages more in line with actual needs and reducing the limitations of manually preset parameters.

[0060] By modeling user behavior sequences using graph neural networks, the system can uncover temporal correlations and business logic relationships between behaviors. For example, it can identify abnormal paths such as "multiple payment cancellations followed by returning to the shopping cart," thereby more accurately pinpointing user churn risks or optimizing recommendation paths. The multi-task learning framework further integrates clustering, path optimization, and prediction tasks for collaborative training, avoiding the limitations of a single model. For instance, while identifying high-value users, it can simultaneously generate personalized interface adjustment strategies.

[0061] Finally, the stream processing engine enables real-time analysis and integration with business systems, allowing strategy adjustments to be completed almost simultaneously with user behavior. For example, when the model detects a risk of user churn, it can immediately trigger coupon pushes or simplify interface operations. This closed-loop mechanism not only improves the user's immediate experience but also optimizes the model through continuous data feedback, forming a continuous iterative capability of "analysis-decision-verification" and reducing business intervention costs.

[0062] Optionally, the role weight feature The generation process includes:

[0063] ;

[0064] ;

[0065] ;

[0066] in, The adaptive role weights are determined by the adaptive weight allocation algorithm. Dynamic weights for business stages. Dynamic weighting based on user activity The total number of stages, For the first The stage factor for each stage (configurable to 1.2 during the promotion period and 1.0 during the non-promotion period). For user value, User churn rate, user value Churn rate can be measured by ARPU (average revenue per user) and calculated from historical data, as the average user value, as the number of times the user's finger slides in the specified product interface within the specified time (configurable to 30 seconds), as the total time spent in the specified product interface, as the time it takes for the user to complete the target behavior (such as clicking on the product category bar - adding to the shopping cart - confirming the order), as the user portrait label quantification factor, the user portrait label can be obtained by combining the available third-party data, and the specific value of the corresponding quantification factor is determined according to the preset matching rule.

[0067] Optionally, the generation process of the behavior decay feature includes:

[0068] ;

[0069] ;

[0070] wherein, is a piecewise decay function, used to distinguish between recent behavior and long-term behavior, is a behavior type weight, which can include key behaviors and secondary behaviors, key behaviors such as successful payment is 1.5, and secondary behaviors such as page browsing is 0.8, is a time demarcation point, which can be selected as 24 hours, is a recent decay coefficient, which can be 0.02, is a long-term decay coefficient, which can be 0.1, and are time decay coefficients determined by an adaptive weight allocation algorithm, respectively; is a business type sensitive coefficient, different business scenarios have different sensitivity to behavior timeliness, such as financial business, which can be 0.05, because it has high sensitivity and needs to respond quickly to abnormal transactions, and e-commerce promotion business, which can be 0.1, which can focus on long-term user retention.

[0071] Optionally, the adaptive weight allocation algorithm includes:

[0072] The role vector, business stage ID, stage duration, and behavior sequence encoding are used as state inputs to define the state space S;

[0073] The adjustment range of the weight parameter is defined as the action space A, which adjusts the contribution proportion of the business stage dynamic weight and the user activity dynamic weight to determine the adaptive role weight the value range of , by discretizing the continuous action space with a step size of 0.01, 70 possible action values can be formed; the decay coefficient the value range is controlled in , also discretized into 90 action values with a step size of 0.01, each action is composed of and value combination.

[0074] The reward function is centered on business goals, balancing retention rate and path conversion rate through dynamic weighting;

[0075] Set the reward function :

[0076] ;

[0077] wherein, is a scene factor related to the business scenario, is the user retention rate, indicating the probability that the user does not flow away within 24 hours after adjusting the weight, is the path conversion rate, indicating the proportion of users who complete the target behavior (such as successful payment) in the optimized path, is the penalty coefficient weight, is the penalty coefficient, to prevent the weight parameter from fluctuating sharply, , is the adaptive role weight value of the jth action, Similarly.

[0078] Specifically:

[0079] User A is currently a VIP user (role vector [0, 1, 0]) and is in the payment stage (stage ID = 2, duration = 5 minutes), the last 5 behaviors include "product browsing - adding to shopping cart - clicking to pay - canceling payment - returning to homepage", according to the current state space, the output action is 0.55, is 0.03, after the system applies the new weight, user A does not flow away within 1 hour (retention rate reward + 0.8), but does not complete payment (conversion rate reward - 0.2), at the same time, the weight adjustment amplitude is small (stability reward + 0.1), the final total reward Rt = 0.7; through the PPO algorithm to calculate the advantage function, update the policy network parameters, so that in similar states, it is more inclined to reduce (prolong the behavior decay period) or adjust (balance role permission and influence coefficient) to improve subsequent rewards. Through real-time updating of the state space, the model automatically increases , strengthen the impact of recent cancellation behavior, and trigger intervention strategies in advance. The stability penalty term controls the weight parameter fluctuations within a specified range, preventing model collapse due to excessive pursuit of short-term conversion rates

[0080] Optionally, the graph neural network adopts a graph attention network (GAT) that aggregates neighbor node information through a self-attention mechanism to extract high-order features of user behavior and output an aggregated feature vector for each node to represent the context relevance of the user's current behavior.

[0081] The neighbor nodes are composed of time edges and logical edges, with each user behavior being a node.

[0082] Time edge establishment: If two user behaviors occur consecutively within a preset time interval (e.g., within 2 hours), a time edge is established.

[0083] Logical edge establishment based on logical associations in business processes, such as "browsing goods - adding to shopping cart" being a natural business step.

[0084] Neighbor nodes include time neighbors and logical neighbors; the time neighbors represent all user behaviors connected to node i through time edges, and the logical neighbors represent all user behaviors connected through logical edges.

[0085] Optionally, the multi-task learning model shares the underlying feature representation of the graph attention network while simultaneously outputting user clustering (dividing users into predefined groups such as "high-value VIP users" and "churn risk users" through a Softmax classification layer), path optimization (predicting the optimal path of users in the business process such as "registration - real-name authentication - payment" by generating path weights through reinforcement learning), and churn prediction (binary classification task predicting the probability of user churn in the next 7 days).

[0086] Optionally, the stream processing engine uses Apache Flink to collect user behavior logs, business stage states, and role change events in real time, and the conditions for triggering strategy adjustments include the user entering a high churn risk group or the business path conversion rate being lower than a preset threshold.

[0087] If the user is classified as a "high-risk churn user," a 50-yuan no-threshold coupon is issued through API calls to the BOSS system. If the user fails multiple times in the payment stage (the path optimization result predicts a conversion rate below the threshold), the BOSS system is triggered to adjust the interface and simplify the payment steps and UI display.

[0088] Model inference is optimized using TensorRT or ONNX and performed on a GPU server. Link stream processing tasks achieve millisecond-level latency through a state backend (such as RocksDB). Model updates use shadow deployment, with new and old models running in parallel. The model is switched after A / B testing to verify the effect.

[0089] Full-process scenario example:

[0090] VIP users clicked "Cancel Payment" three times consecutively during the payment process.

[0091] It is 0.6 (enhanced VIP privileges). With a value of 0.03 (focusing on short-term behavior during the promotion period), construct a dynamic graph structure that includes cancellation behavior;

[0092] User segmentation results: changed from "high-value users" to "high-risk churn users";

[0093] Path optimization suggestion: Shorten the payment process steps;

[0094] Churn prediction probability: increased from 15% to 60%.

[0095] Real-time push and execution:

[0096] Flink triggered a push notification of a 50 yuan coupon to the user's account;

[0097] The BOSS system has adjusted the payment interface to a simplified "one-click payment" version.

[0098] Real-time tracking records subsequent user behaviors (such as whether coupons are claimed or payments are successful), and the data is fed back to the reinforcement learning framework to optimize weight parameters.

[0099] Through the above design, the system achieves full-link automation from dynamic feature fusion to business strategy execution, significantly improving user retention and conversion efficiency, while ensuring real-time performance and stability.

[0100] This application also discloses an analysis system for user behavior data in a BOSS system, including:

[0101] The data acquisition module is used to collect multi-source behavioral data of users in the BOSS system; the multi-source behavioral data includes operation logs, business database records, third-party system data and user role attributes;

[0102] The dynamic weight generation module is used to generate a dynamic feature set that includes at least role weight features and behavior decay features based on user roles, business stages, and real-time behaviors.

[0103] The fusion weighting module is configured to adopt a reinforcement learning framework according to an adaptive weight distribution algorithm, and to optimize the weighted fusion of dynamic features in real time according to user retention rate and path conversion rate.

[0104] The collaborative analysis module is configured to perform user grouping, path optimization and loss prediction by means of a graph neural network and a multi-task learning model based on the fused features.

[0105] The push module is configured to push the analysis results to a BOSS system in real time through a stream processing engine, and trigger interface permission adjustment or resource pushing.

[0106] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A method for analyzing user behavior data in a BOSS system, characterized by, The method comprises the following steps: Collecting multi-source behavior data of users in a BOSS system; The multi-source behavior data comprises operation logs, business database records, third-party system data, and user role attributes; Generating a dynamic feature set comprising at least role weight features and behavior decay features according to user roles, business stages, and real-time behaviors; The role weight feature The generation process includes: ； ； ； wherein, is an adaptive role weight determined by an adaptive weight allocation algorithm, is a business stage dynamic weight, is a user active dynamic weight, is a stage total number, is a stage factor of the th stage, is a user value, is a user churn rate, is an average user value, is a number of times that a user's finger slides in a specified product interface within a specified time length, is a total time length of staying in a specified product interface, is a time used by a user to complete a target behavior, is a user portrait label quantification factor; The generation process of the behavior decay features comprises: ； ； wherein, is a piecewise decay function, is a behavior type weight, is a time demarcation point, is a recent decay coefficient, is a distant decay coefficient, and are time decay coefficients determined by an adaptive weight assignment algorithm, respectively, is a service type sensitivity coefficient, ; According to an adaptive weight allocation algorithm based on a reinforcement learning framework, the weighted fusion of dynamic features is optimized in real time according to user retention rates and path conversion rates, wherein the adaptive weight allocation algorithm comprises: Taking user roles and business stages as state inputs, a state space S is defined; The adjustment range of the weight parameter is defined as an action space A; Setting a reward function : ； wherein, is a scenario factor related to a business scenario, is a user retention rate, indicating the probability that a user does not flow away within the TC hours after the weight adjustment, is a path conversion rate, indicating the proportion of users completing the target behavior in the optimized path, is a penalty coefficient weight, is a penalty coefficient; Based on the fused features, a graph neural network and a multi-task learning model are used for collaborative analysis to perform user clustering, path optimization, and churn prediction; The analysis results are pushed to the BOSS system in real time through a stream processing engine to trigger interface permission adjustment or resource pushing.

2. The method of claim 1, wherein the BOSS system analyzes the user behavior data, and The graph neural network uses a graph attention network to aggregate neighbor node information through a self-attention mechanism to extract high-order features of user behaviors, and the neighbor nodes are composed of time edges and logical edges.

3. The method of claim 2, wherein the BOSS system analyzes the user behavior data by, The multi-task learning model shares the bottom-level feature representation of the graph attention network and simultaneously outputs user clustering, path optimization, and churn prediction.

4. The method of claim 3, wherein the BOSS system analyzes the user behavior data by, The method comprises the following steps: Each user behavior is a node; Time edges are established: if two user behaviors occur continuously within a preset time interval, a time edge is established; Logical edges are established based on logical associations of business processes; Neighbor nodes include time neighbors and logical neighbors; the time neighbors represent all user behaviors connected to node i through time edges, and the logical neighbors represent all user behaviors connected through logical edges.

5. The method of claim 1, wherein the BOSS system analyzes the user behavior data, and The stream processing engine is implemented using Apache Flink, and the conditions for triggering strategy adjustment include the user entering a high-risk churn group or the business path conversion rate being lower than a preset threshold.

6. A system for analyzing user behavior data in a BOSS system, characterized by, The method comprises the following steps: A data collection module is configured to collect multi-source behavior data of users in a BOSS system; the multi-source behavior data comprises operation logs, business database records, third-party system data, and user role attributes; A dynamic weight generation module is configured to generate a dynamic feature set comprising at least role weight features and behavior decay features according to user roles, business stages, and real-time behaviors; A fusion weighting module is configured to use a reinforcement learning framework to optimize the weighted fusion of dynamic features in real time according to user retention rates and path conversion rates based on an adaptive weight allocation algorithm; A collaborative analysis module is configured to perform user clustering, path optimization, and churn prediction based on fused features through a graph neural network and a multi-task learning model; A pushing module is configured to push analysis results to the BOSS system in real time through a stream processing engine to trigger interface permission adjustment or resource pushing.

Images