Method and system for analyzing user behavior data in BOSS system

Abstract

The invention relates to the technical field of data processing technologies, and discloses a method and system for analyzing user behavior data in a BOSS system, and the method comprises the steps: collecting multi-source behavior data of a user in the BOSS system; generating a dynamic feature set at least comprising role weight features and behavior attenuation features according to the user roles, the service stages and the real-time behaviors; according to a self-adaptive weight distribution algorithm based on a reinforcement learning framework, optimizing weight fusion of dynamic characteristics in real time according to a user retention rate and a path conversion rate; based on the fused features, performing collaborative analysis through a graph neural network and a multi-task learning model, and executing user grouping, path optimization and loss prediction; and the analysis result is pushed to the BOSS system in real time through the stream processing engine, and interface permission adjustment or resource pushing is triggered, so that the refinement degree of service decision is enhanced, and the user retention and conversion efficiency and the self-adaptive capability of system operation are remarkably improved through a real-time closed loop.

Description

Technical Field

[0001] This application relates to the technical field of data processing technology, and in particular, to a method and system for analyzing user behavior data in a BOSS system. Background Art

[0002] In a Business Operation Support System (BOSS system), the analysis of user behavior data is a core part of optimizing system performance, enhancing user experience, and ensuring business security. As a key infrastructure for enterprise operation, the BOSS system is responsible for handling all processes such as user registration, service subscription, bill management, and fault reporting. Its data flow is complex and large in scale, including multi-dimensional information such as user login frequency, operation path, service request pattern, and abnormal behavior. However, traditional user behavior analysis methods face many challenges when dealing with the complexity of the BOSS system.

[0003] First, traditional analysis methods mainly rely on log statistics and fixed rule engines. For example, user operation logs are extracted through ETL tools and static reports are generated. Although such methods can provide a basic data overview, they have significant limitations: First, the data processing process has a high latency and cannot achieve real-time analysis, resulting in a lag in response to sudden abnormalities such as account theft or batch fraud; Second, the analysis dimension is single, and it is difficult to capture the dynamic patterns of user behavior (such as the operation correlation across services), resulting in the lack of key business insights; Third, relying on manually defined rules, it is difficult to adapt to the rapid changes in business scenarios, such as the launch of new services or the migration of user behavior patterns.

[0004] Second, with the development of big data technology, some systems have begun to try to introduce machine learning and data mining technologies. For example, clustering analysis is used to identify user group characteristics, or a time series prediction model is used to predict the risk of user churn. However, the application of existing technologies in the BOSS system still has bottlenecks: on the one hand, data sources are scattered in multiple subsystems (such as the billing module, customer service system, and background management), and data integration is difficult, resulting in one-sided analysis results; on the other hand, model training relies on historical data and cannot quickly and accurately respond to the user's behavior sequence in a timely manner. Summary of the Invention

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

[0006] In the first aspect, this application provides a method for analyzing user behavior data in a BOSS system, adopting the following technical solution:

[0007] A method for analyzing user behavior data in a BOSS system includes:

[0008] Collect 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;

[0009] 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;

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

[0011] Based on the fused features, we conduct collaborative analysis through graph neural networks and multi-task learning models to perform user segmentation, path optimization, and churn prediction.

[0012] The analysis results are pushed to the BOSS system in real time through the stream processing engine, triggering interface permission adjustments or resource push.

[0013] By adopting the above technical solutions and integrating operation logs, business data and third-party information, the system can fully capture the multi-dimensional characteristics of user behavior, especially the dynamic changes in role attributes and business stages, making the feature description both accurate and timely, and 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 user payments fail, the system will automatically strengthen the weight of recent behaviors, or adjust the impact of role privileges on decision-making, so that the model's response at different business stages is more in line with actual needs and reduces the limitations of manually preset parameters.

[0014] By modeling user behavior sequences through graph neural networks, the system can uncover temporal correlations and business logic relationships between behaviors. For example, it can identify unusual paths like "multiple payment cancellations followed by returns to the shopping cart," thereby more accurately pinpointing user churn risk or optimizing recommended paths. The multi-task learning framework further integrates clustering, path optimization, and prediction tasks for collaborative training, avoiding the one-sidedness of a single model. For example, it can simultaneously identify high-value users while generating personalized interface adjustment strategies.

[0015] Finally, the stream processing engine enables real-time analysis and integration with business systems, enabling policy adjustments to be made almost simultaneously with user behavior. For example, when the model detects a user at risk of churn, it can immediately trigger coupon push or streamlined interface operations. This closed-loop mechanism not only enhances the user experience but also optimizes the model through continuous data feedback, forming a continuous iterative "analysis-decision-verification" process and reducing business intervention costs.

[0016] Overall, upgrading user behavior analysis from traditional offline statistics to an active policy system of dynamic perception and intelligent response not only enhances the refinement of business decisions but also significantly improves user retention, conversion efficiency, and the adaptive ability of system operation through real-time closed-loop control.

[0017] Optionally, the generation process of the role weight feature includes: ; ; ;

[0018] where is the adaptive role weight determined by the adaptive weight allocation algorithm, is the dynamic weight of the business stage, is the dynamic weight of user activity, is the total number of stages, is the stage factor of the th stage, is the user value, is the user churn rate, is the number of times the user's finger slides on the specified product interface within the specified duration, is the total duration of staying on the specified product interface, is the time taken by the user to complete the target behavior, is the quantization factor of the user portrait label.

[0019] Optionally, the generation process of the behavior attenuation feature includes: ; ;

[0020] where is the piecewise attenuation function, is the behavior type weight, is the time demarcation point, is the recent attenuation coefficient, is the long-term attenuation coefficient, and are the time attenuation coefficients determined by the adaptive weight allocation algorithm respectively, is the business type sensitivity coefficient, .

[0021] By adopting the above technical solutions, the limitations of traditional static role tags can be overcome. The weight allocation can be flexibly adjusted according to real-time business objectives (for example, during the promotion period, VIP user privileges are emphasized, and during the risk control period, the monitoring of high-risk roles is strengthened), enabling the model to focus on the key role attributes that truly affect user behavior. The synergistic effect of role weights and behavior decay characteristics effectively captures the impact of role changes on behavior.

[0022] Optionally, the adaptive weight allocation algorithm includes:

[0023] Taking the user role and business stage as state inputs, define the state space S;

[0024] Define the adjustment range of the weight parameter as the action space A;

[0025] Set the reward function : ;

[0026] Wherein, is the scenario factor related to the business scenario, is the user retention rate, indicating the probability that the user has not churned within TC hours after the weight adjustment, is the path conversion rate, representing the proportion of users who complete the target behavior in the optimized path, is the penalty coefficient weight, is the penalty coefficient.

[0027] Optionally, the graph neural network adopts a graph attention network, aggregates neighbor node information through the self-attention mechanism, and extracts high-order features of user behavior. The neighbor nodes are composed of time edges and logical edges.

[0028] Optionally, the multi-task learning model shares the underlying feature representation of the graph attention network and outputs user clustering, path optimization, and churn prediction simultaneously.

[0029] Optionally, it includes:

[0030] Each user behavior is a node;

[0031] Establishment of time edges: If two user behaviors occur continuously within a preset time interval, establish time edges;

[0032] Establishment of logical edges, based on the logical association of the business process;

[0033] The 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.

[0034] Optionally, the stream processing engine is implemented using Apache Flink, and the conditions for triggering policy adjustment include the user entering a high churn risk group or the business path conversion rate being lower than a preset threshold.

[0035] In a second aspect, the present application provides a system for analyzing user behavior data in a BOSS system, which adopts the following technical solutions:

[0036] A user behavior data analysis system in a BOSS system, comprising:

[0037] The data collection module is used to collect 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;

[0038] A dynamic weight generation module is used to generate a dynamic feature set including at least a role weight feature and a behavior decay feature based on the user role, business stage and real-time behavior;

[0039] The fusion weighting module is used to optimize the weighted fusion of dynamic features in real time based on user retention rate and path conversion rate using a reinforcement learning framework based on an adaptive weight allocation algorithm;

[0040] The collaborative analysis module is used to perform user segmentation, path optimization, and churn prediction based on the fused features through collaborative analysis of graph neural networks and multi-task learning models;

[0041] The push module is used to push analysis results to the BOSS system in real time through the stream processing engine, triggering interface permission adjustments or resource push.

[0042] In summary, the present application includes at least one of the following beneficial technical effects:

[0043] By integrating operation logs, business data, and third-party information, the system can fully capture the multi-dimensional characteristics of user behavior, especially the dynamic changes in role attributes and business stages, making the feature description both accurate and timely, 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 goals. By modeling user behavior sequences through graph neural networks, the system can explore the temporal correlation and business logic relationships between behaviors, thereby more accurately locating 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 one-sidedness of a single model; finally, the stream processing engine enables real-time analysis and business system linkage, which not only improves the user's immediate experience, but also optimizes the model through continuous data feedback, reducing business intervention costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic flow chart of the method for analyzing user behavior data in this application. Specific Embodiments

[0045] The following details the embodiments of this application, and the examples of the embodiments are shown in the accompanying drawings.

[0046] In the description of this specification, the description with reference to the terms "certain embodiments", "one embodiment", "some embodiments", "schematic embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in connection with the embodiments or examples are included in at least one embodiment or example of this application. In this specification, the schematic descriptions of the above terms do not necessarily refer to the same embodiments or examples. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0047] The embodiments of this application disclose a method for analyzing user behavior data in a BOSS system. Refer to Figure 1 , a method for analyzing user behavior data in a BOSS system, including:

[0048] Collect 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;

[0049] Generate a dynamic feature set that at least includes role weight features and behavior decay features according to user roles (such as VIP level, permission type), business stages (such as registration process, promotion period, system maintenance, payment confirmation, complaint handling, etc.), and real-time behaviors (such as specific operations and operation types, timestamps, page paths);

[0050] According to the adaptive weight allocation algorithm based on the reinforcement learning framework, optimize the weighted fusion of dynamic features in real time according to user retention rate and path conversion rate;

[0051] Based on the fused features, perform user clustering, path optimization, and churn prediction through the collaborative analysis of a graph neural network and a multi-task learning model;

[0052] Push the analysis results to the BOSS system in real time through a stream processing engine, triggering interface permission adjustment or resource push.

[0053] By adopting the above technical solutions and integrating operation logs, business data and third-party information, the system can fully capture the multi-dimensional characteristics of user behavior, especially the dynamic changes in role attributes and business stages, making the feature description both accurate and timely, and 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 user payments fail, the system will automatically strengthen the weight of recent behaviors, or adjust the impact of role privileges on decision-making, so that the model's response at different business stages is more in line with actual needs and reduces the limitations of manually preset parameters.

[0054] By modeling user behavior sequences through graph neural networks, the system can uncover temporal correlations and business logic relationships between behaviors. For example, it can identify unusual paths like "multiple payment cancellations followed by returns to the shopping cart," thereby more accurately pinpointing user churn risk or optimizing recommended paths. The multi-task learning framework further integrates clustering, path optimization, and prediction tasks for collaborative training, avoiding the one-sidedness of a single model. For example, it can simultaneously identify high-value users while generating personalized interface adjustment strategies.

[0055] Finally, the stream processing engine enables real-time analysis and integration with business systems, enabling policy adjustments to be made almost simultaneously with user behavior. For example, when the model detects a user at risk of churn, it can immediately trigger coupon push or streamlined interface operations. This closed-loop mechanism not only enhances the user experience but also optimizes the model through continuous data feedback, forming a continuous iterative "analysis-decision-verification" process and reducing business intervention costs.

[0056] Optionally, the role weight feature The generation process includes: ; ; ;

[0057] in, is the adaptive role weight determined by the adaptive weight allocation algorithm, is the dynamic weight of the business stage, Active dynamic weight for users, is the total number of stages, For the The stage factor of each stage (configurable to 1.2 during the promotion period and 1.0 during the non-promotion period), For user value, is user churn rate, user value It can be measured by ARPU (average revenue per user), and the churn rate can be obtained from historical data statistics. is the average user value, is the number of times the user's finger slides on the specified product interface within the specified duration (configurable to 30 seconds), is the total duration of staying on the specified product interface, is the time taken by the user to complete the target behavior (such as clicking on the product category bar - adding to the shopping cart - confirming the order), is the quantization factor of the user portrait label. The user portrait label can be obtained by combining third - party data that can be consulted, and the specific value of the corresponding quantization factor is determined according to the preset matching rules.

[0058] Optionally, the generation process of the behavior attenuation feature includes: ; ;

[0059] where is the piece - wise attenuation function, , which is used to distinguish recent behaviors and long - term behaviors, is the 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 views is 0.8, is the time demarcation point, which can generally be selected as 24 hours, is the recent attenuation coefficient, which can be 0.02, is the long - term attenuation coefficient, which can be 0.1, and are respectively the time attenuation coefficients determined by the adaptive weight allocation algorithm; is the business type sensitivity coefficient. Different business scenarios have different sensitivities to the timeliness of behaviors. For example, in the financial business, it can be 0.05 because of high sensitivity and the need to quickly respond to abnormal transactions. In the e - commerce promotion business, it can be 0.1 to focus on long - term user retention.

[0060] Optionally, the adaptive weight allocation algorithm includes:

[0061] Using the role vector, business stage ID, stage duration, and behavior sequence encoding as state inputs, define the state space S;

[0062] Define the adjustment range of the weight parameter as the action space A, and adjust the dynamic weights of the business stage and the dynamic weight of user activity to determine the contribution ratio of the adaptive role weight and obtain the value range , which is discretized with a step size of 0.01 through the continuous action space to form 70 possible action values; the attenuation coefficient The value range of is controlled within and and is composed of the value combinations of

[0063] The reward function takes the business goal as the core and balances the retention rate and the path conversion rate through dynamic weighting;

[0064] Set the reward function : ;

[0065] Among them, is the scenario factor related to the business scenario, is the user retention rate, indicating the probability that the user does not churn within the TC hours (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 parameters from fluctuating violently, is the adaptive role weight value of the j-th action, Similarly.

[0066] Specifically:

[0067] User A is currently a VIP user (role vector [0, 1, 0]), in the payment stage (stage ID = 2, duration = 5 minutes), and the last 5 behaviors include "product browsing - adding to cart - clicking to pay - canceling payment - returning to the home page". 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 churn within 1 hour (retention rate reward +0.8), but does not complete the payment (conversion rate reward -0.2). At the same time, the weight adjustment range is small (stability reward +0.1), and the final total reward Rt = 0.7; calculate the advantage function through the PPO algorithm and update the policy network parameters, making it more inclined to reduce (extend the behavior decay period) or adjust (balance the role permissions and influence coefficients) in a similar state to improve the subsequent rewards. Through the real-time update of the state space, the model automatically increases during the high-incidence period of payment failures (such as the promotion period), strengthens the impact of recent cancellation behaviors, and can trigger the intervention strategy in advance. The stability penalty term controls the weight parameter fluctuations within the specified range to avoid model collapse caused by excessive pursuit of short-term conversion rates

[0068] Optionally, the graph neural network adopts a graph attention network (GAT), aggregates neighbor node information through a self-attention mechanism, extracts high-order features of user behavior, and is used to output an aggregated feature vector for each node, which is used to represent the context relevance of the user's current behavior.

[0069] The neighbor nodes are composed of time edges and logical edges, and each user behavior is a node;

[0070] Establishment of time edges: If two user behaviors occur continuously within a preset time interval (such as within 2 hours), a time edge is established;

[0071] Establishment of logical edges is based on the logical association of the business process. For example, "browsing products - adding to the shopping cart" is a natural business step;

[0072] 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.

[0073] Optionally, the multi-task learning model shares the underlying feature representation of the graph attention network and simultaneously outputs user clustering (users are divided into predefined groups such as "high-value VIP users" and "users at risk of churn" through a Softmax classification layer), path optimization (predicting the optimal path of users in the business process, such as "registration - real-name authentication - payment", and generating path weights through reinforcement learning), and churn prediction (a binary classification task to predict the churn probability of users in the next 7 days).

[0074] Optionally, the stream processing engine uses Apache Flink to collect user behavior logs, business stage status, and role change events in real time. The conditions for triggering policy adjustments include users entering a high-churn risk group or the conversion rate of the business path being lower than a preset threshold;

[0075] If the user is grouped as a "high-risk churn user", a 50-yuan no-threshold coupon is issued by calling the BOSS system through an API; if the user fails to pay multiple times during the payment stage (the predicted conversion rate of the path optimization result is lower than the threshold), the BOSS system is triggered to adjust the interface and simplify the payment steps and UI display.

[0076] Model inference is optimized through TensorRT or ONNX and completed on a GPU server. The link stream processing task achieves millisecond-level latency through a state backend (such as RocksDB). The model update adopts shadow deployment, and the old and new models run in parallel and are switched after verifying the effect through A / B testing.

[0077] Full-process scenario example:

[0078] A VIP user clicks "cancel payment" three times consecutively during the payment stage.

[0079] is 0.6 (enhanced VIP privilege), is 0.03 (focus on short-term behavior during the promotion period), and construct a dynamic graph structure including cancellation behavior;

[0080] User segmentation result: adjusted from "high-value users" to "high-risk users at risk of churn";

[0081] Path optimization suggestion: shorten the steps of the payment process;

[0082] Churn prediction probability: increased from 15% to 60%.

[0083] Real-time push and execution:

[0084] Flink triggers to push a 50-yuan coupon to the user's account;

[0085] The BOSS system adjusts the payment interface to a simplified version of "one-click payment";

[0086] Real-time event tracking records the user's subsequent behavior (such as whether the coupon is received and the payment is successful), and the data is fed back to the reinforcement learning framework to optimize the weight parameters.

[0087] Through the above design, the system realizes the full-link automation from dynamic feature fusion to business strategy execution, significantly improves user retention and conversion efficiency, and ensures real-time performance and stability at the same time.

[0088] This application embodiment also discloses an analysis system for user behavior data in a BOSS system, including:

[0089] A data collection module for 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;

[0090] A dynamic weight generation module for generating a dynamic feature set including at least role weight features and behavior attenuation features according to user roles, business stages, and real-time behaviors;

[0091] A fusion weighting module for using a reinforcement learning framework according to an adaptive weight allocation algorithm to optimize the weighted fusion of dynamic features in real time according to user retention rate and path conversion rate;

[0092] A collaborative analysis module for performing user segmentation, path optimization, and churn prediction through collaborative analysis of a graph neural network and a multi-task learning model based on the fused features;

[0093] A push module for pushing the analysis results to the BOSS system in real time through a stream processing engine to trigger interface permission adjustment or resource push.

[0094] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions, 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 in that Including: 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; Generating a dynamic feature set including at least role weight features and behavior decay features according to user roles, business stages, and real-time behaviors; According to the adaptive weight allocation algorithm based on the reinforcement learning framework, optimizing the weighted fusion of dynamic features in real time according to user retention rate and path conversion rate; Based on the fused features, through the collaborative analysis of the graph neural network and the multi-task learning model, performing user clustering, path optimization, and churn prediction; Pushing the analysis results to the BOSS system in real time through the stream processing engine to trigger interface permission adjustment or resource push.

2. The analysis method of user behavior data in the BOSS system according to claim 1, characterized in that, The said character weight feature The generation process includes: ； ； ； in, is the adaptive role weight determined by the adaptive weight allocation algorithm, is the dynamic weight of the business stage, Active dynamic weight for users, is the total number of stages, For the The stage factor of each stage, For user value, is the user churn rate, is the average user value, The number of times a user slides their finger on a specified product interface within a specified time period. The total time spent on a specific product interface. The time it takes for users to complete the target behavior, Quantify factors for user portrait labels.

3. The analysis method of user behavior data in the BOSS system according to claim 1, wherein, The generation process of the behavior decay feature includes: ； ； Among them, is a piecewise decay function, is the weight of behavior type, is the time demarcation point, is the recent decay coefficient, is the long-term decay coefficient, and are the time decay coefficients determined by the adaptive weight allocation algorithm respectively, is the business type sensitivity coefficient, .

4. The analysis method of user behavior data in the BOSS system according to claim 1, characterized in that, The adaptive weight allocation algorithm includes: Taking user roles and business stages as state inputs and defining the state space S; Defining the adjustment range of weight parameters as the action space A; Set the reward function : ； Among them, is a scenario factor related to the business scenario, is the user retention rate, representing the probability that users do not churn within TC hours after adjusting the weights, is the path conversion rate, representing the proportion of users who complete the target behavior in the optimized path, is the penalty coefficient weight, is the penalty coefficient.

5. The method for analyzing user behavior data in the BOSS system according to claim 1, wherein The graph neural network adopts a graph attention network, aggregates neighbor node information through the self-attention mechanism, and extracts high-order features of user behaviors. The neighbor nodes are composed of time edges and logical edges.

6. The method for analyzing user behavior data in the BOSS system according to claim 5, characterized in that, The multi-task learning model shares the underlying feature representation of the graph attention network and outputs user clustering, path optimization, and churn prediction at the same time.

7. The method for analyzing user behavior data in the BOSS system according to claim 6, wherein Including: Each user behavior is a node; Establishment of time edges: If two user behaviors occur continuously within a preset time interval, time edges are established; Establishment of logical edges, based on the logical association of the business process; 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.

8. The method for analyzing user behavior data in the BOSS system according to claim 1, characterized in that, The stream processing engine is implemented by Apache Flink, and the conditions for triggering policy adjustment include that the user enters a high-churn risk group or the business path conversion rate is lower than a preset threshold.

9. An analysis system for user behavior data in a BOSS system, characterized in that, Including: A data collection module for 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; A dynamic weight generation module for generating a dynamic feature set including at least role weight features and behavior decay features according to user roles, business stages, and real-time behaviors; A fusion weighting module for optimizing the weighted fusion of dynamic features in real time according to the user retention rate and path conversion rate by using the adaptive weight allocation algorithm based on the reinforcement learning framework; A collaborative analysis module for performing user clustering, path optimization, and churn prediction through the collaborative analysis of the graph neural network and the multi-task learning model based on the fused features; A push module for pushing the analysis results to the BOSS system in real time through the stream processing engine to trigger interface permission adjustment or resource push.

Images