Behavior analysis driven churn retention method, apparatus, device, and medium

Abstract

This invention relates to the field of intelligent decision-making technology, and discloses a behavior analysis-driven churn and retention method, apparatus, device, and medium, comprising: acquiring interaction record data and object attribute data; generating a multi-source behavioral data set based on the interaction record data and object attribute data to form a churn signal representation and a hierarchical state representation; outputting a churn risk score, churn stage identification results, and a candidate set of churn causes through a fusion prediction model, and generating a target retention strategy by combining historical retention response records and a strategy knowledge graph; and updating the fusion prediction model and strategy knowledge graph based on feedback data. This invention can be applied to business scenarios such as fintech, enabling churn warning results to be transformed into target retention strategies through the continuous collaboration of multi-source behavioral data, a fusion prediction model, and a strategy knowledge graph, and updating the model and graph through feedback data, thereby improving the continuity of warning, intervention, and feedback updates.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, and in particular to a behavior analysis-driven churn retention method, apparatus, device, and medium. Background Technology

[0002] In customer continuity services, existing churn prediction technologies typically separate risk prediction, cause identification, strategy generation, and feedback updates into different processing stages, lacking a stable data connection between these stages. Even if the prediction model can output churn risk, it often struggles to establish a continuous link between churn risk, churn stage, churn cause, and subsequent retention strategies. Furthermore, feedback data after strategy execution is difficult to integrate into the prediction model and strategy knowledge structure in a timely manner. Therefore, the main shortcoming of existing technologies lies in the lack of a traceable, continuous processing chain between prediction judgment and retention intervention.

[0003] In the fintech business, across scenarios such as banking, insurance, securities, consumer credit, wealth management, and payment services, customers generate a large number of interaction records during account access, product browsing, transaction processing, communication with account managers, customer service inquiries, and complaint handling. Existing churn warning methods mostly rely on asset changes, transaction frequency, product holdings, or simple activity indicators. Even if a customer is identified as having churn risk, it is difficult to further correlate risk scores, churn stages, reasons for churn, and historical retention response records, resulting in a lack of clear evidence linking warning results to subsequent retention strategies.

[0004] Meanwhile, retention management in existing fintech businesses typically relies more on human experience or fixed strategy configurations. Feedback data such as resource usage, page revisits, service reach, message response, changes in complaints, and retention status after strategy execution are difficult to correlate with churn signals prior to strategy triggering. Due to the lack of a continuous chain from risk identification to strategy generation and feedback updates, existing technologies struggle to support the continuous updating of subsequent predictive models and strategy knowledge structures. Summary of the Invention

[0005] The main objective of this invention is to provide a behavior analysis-driven churn and retention method, apparatus, device, and storage medium, aiming to solve the technical problem in the prior art where there is a lack of continuous processing mechanism between churn risk identification, retention strategy generation, and strategy feedback update, resulting in insufficient connection between churn warning results and subsequent retention intervention processes.

[0006] To achieve the above objectives, the present invention provides a behavior analysis-driven churn and retention method, comprising: Acquire interaction record data and object attribute data, and then identify, bind, and integrate the interaction record data and object attribute data in a time sequence to generate a multi-source behavior data set. Based on the multi-source behavioral data set, a behavioral path representation, emotional interaction features, and hierarchical state representation are generated. The behavioral path representation and the emotional interaction features are fused to generate a loss signal representation, and early warning judgment parameters are configured in the hierarchical state representation. The churn signal representation and the hierarchical state representation are input into the fusion prediction model, which outputs a churn risk score, churn stage identification results, and a candidate set of churn reasons. Based on the churn risk score, the churn stage identification result, and the candidate set of churn reasons, a cause evidence vector is constructed. The cause evidence vector is matched with historical retention response records to generate a candidate retention case set. A strategy knowledge graph is obtained. Based on the hierarchical state representation, the cause evidence vector, the candidate retention case set, and the strategy knowledge graph, a candidate retention strategy set is generated. The feasibility of implementing the candidate retention strategy set is assessed, and response deduction is performed based on the assessment results to generate the target retention strategy; When the churn risk score reaches the warning judgment parameter, the target retention strategy is pushed to the corresponding execution end, the feedback data collected for the target retention strategy is converted into strategy response features, the strategy response features are associated with the churn signal representation to generate strategy execution results, and the fusion prediction model and the strategy knowledge graph are updated according to the strategy execution results.

[0007] Furthermore, to achieve the above objectives, the present invention provides a behavior analysis-driven churn retention device, comprising: The multi-source behavior data integration module is used to acquire interaction record data and object attribute data, and to identify, bind, and integrate the interaction record data and object attribute data in a time sequence to generate a multi-source behavior data set. The churn signal construction module is used to generate a behavior path representation, emotional interaction features and hierarchical state representation based on the multi-source behavior data set, fuse the behavior path representation and the emotional interaction features to generate a churn signal representation, and configure early warning judgment parameters in the hierarchical state representation; The fusion risk prediction module is used to input the churn signal representation and the hierarchical state representation into the fusion prediction model, and output the churn risk score, churn stage identification result and churn reason candidate set; The strategy knowledge mapping module is used to construct a cause evidence vector based on the churn risk score, the churn stage identification result and the churn cause candidate set, match the cause evidence vector with historical retention response records to generate a candidate retention case set, obtain a strategy knowledge graph, and generate a candidate retention strategy set based on the hierarchical state representation, the cause evidence vector, the candidate retention case set and the strategy knowledge graph. The retention strategy deduction module is used to evaluate the feasibility of implementing the candidate retention strategy set, perform response deduction based on the evaluation results, and generate the target retention strategy. The strategy feedback update module is used to push the target retention strategy to the corresponding execution end when the churn risk score reaches the warning judgment parameter, convert the feedback data collected for the target retention strategy into strategy response features, associate the strategy response features with the churn signal representation to generate strategy execution results, and update the fusion prediction model and the strategy knowledge graph according to the strategy execution results.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a behavior analysis-driven churn retention program stored in the memory and executable on the processor, wherein when the behavior analysis-driven churn retention program is executed by the processor, it implements the steps of the behavior analysis-driven churn retention method as described above.

[0009] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a behavior analysis-driven churn retention program, which, when executed by a processor, implements the steps of the behavior analysis-driven churn retention method described above.

[0010] Beneficial Effects: This invention relates to the field of intelligent decision-making technology, and discloses a behavior analysis-driven churn and retention method, apparatus, device, and medium, comprising: acquiring interaction record data and object attribute data, generating a multi-source behavior data set through identifier binding and time-series integration; generating a churn signal representation and a hierarchical state representation based on the multi-source behavior data set; inputting the churn signal representation and hierarchical state representation into a fusion prediction model, and outputting a churn risk score, churn stage identification results, and a candidate set of churn causes; constructing a cause evidence vector based on the above results, and generating a candidate retention strategy set by combining historical retention response records and a strategy knowledge graph; generating a target retention strategy through implementation feasibility assessment and response deduction, and generating strategy execution results based on feedback data, and updating the fusion prediction model and strategy knowledge graph. This invention can be applied to business scenarios such as fintech, forming a churn signal representation and hierarchical state representation through a multi-source behavior data set, and then using a fusion prediction model to generate risk and cause results, enabling the early warning results to be transformed into a cause evidence vector and a candidate retention strategy set; generating strategy execution results through feedback data and updating the fusion prediction model and strategy knowledge graph, thereby improving the continuity between churn early warning, retention strategy generation, and feedback updates. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for a behavior analysis-driven churn and retention method according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an embodiment of the behavior analysis-driven churn and retention method of the present invention. Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the behavior analysis-driven churn and retention device of the present invention. Figure 4 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0013] The behavior analysis-driven churn and retention method provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can obtain interaction record data and object attribute data from the client, and generate a multi-source behavioral data set through identifier binding and time-series integration. Based on the multi-source behavioral data set, a churn signal representation and a hierarchical state representation are generated. The churn signal representation and hierarchical state representation are input into a fusion prediction model, which outputs a churn risk score, churn stage identification results, and a candidate set of churn causes. Based on the above results, a cause evidence vector is constructed, and a candidate retention strategy set is generated by combining historical retention response records and a strategy knowledge graph. After implementation feasibility assessment and response deduction, a target retention strategy is generated, and the strategy execution results are generated based on feedback data, updating the fusion prediction model and the strategy knowledge graph. This invention can be applied to business scenarios such as fintech. By forming a churn signal representation and a hierarchical state representation through a multi-source behavioral data set, and then using a fusion prediction model to generate risk and cause results, the warning results can be transformed into a cause evidence vector and a candidate retention strategy set. By generating strategy execution results through feedback data and updating the fusion prediction model and the strategy knowledge graph, the continuity between churn warning, retention strategy generation, and feedback updates is improved. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.

[0014] Please see Figure 2 , Figure 2This is a flowchart illustrating an embodiment of the behavior analysis-driven churn and retention method provided by the present invention. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0015] like Figure 2 As shown, the behavior analysis-driven churn and retention method proposed in this invention includes the following steps: S10, acquire interaction record data and object attribute data, and perform identification binding and time-series integration on the interaction record data and the object attribute data to generate a multi-source behavior data set; In this embodiment, interaction record data is collected from the application terminal, web page terminal, customer service terminal, messaging terminal, and business processing terminal. Fields may include interaction entry point, session marker, event occurrence time, business object marker, interaction action name, page dwell time, text content, and speech-to-text content. Object attribute data is read from object archives, service records, product records, and channel records. Fields may include object basic markers, object channel attributes, object grouping attributes, object service status, and object usage period.

[0016] Identifier binding matches session tags, business object tags, interaction entry points, object base tags, and object channel attributes to generate object binding results. When multiple candidate identifiers exist, the candidate identifiers, data sources, occurrence times, and trusted tags are retained, enabling records generated from different channels to be merged into the same object.

[0017] Time-series integration establishes time anchors based on event occurrence time, data reception time, and business completion time, and writes object attribute data into the corresponding time range according to attribute update time and valid interval. Missing time, repeated interaction, abnormal pause, abnormal jump, and text transcription anomalies form abnormal fragment markers, which, together with interaction records, object attributes, and object binding results, generate a multi-source behavioral data set.

[0018] In one implementation, the data service reads account access records, product browsing records, customer service communication records, and object attribute records according to time windows, generates object binding results based on session tags, business object tags, and object base tags, and then forms a multi-source behavioral data set according to the event occurrence time and business completion time.

[0019] In another implementation, interaction records are entered into a message queue, and object attribute data is cached incrementally. The integration service writes new records to the corresponding time segments based on the object binding results, and generates exception segment markers for repeated clicks, abnormal pauses, session interruptions, and text transcription errors.

[0020] This embodiment integrates interaction record data and object attribute data scattered across different entry points and business systems into the same object through identifier binding and time sequence integration. Time anchors can unify the recording order, and abnormal fragment markers can retain data anomalies, thereby reducing data deviations caused by cross-channel record fragmentation, unclear object ownership, and disordered time sequence.

[0021] S20, Generate a behavior path representation, emotional interaction features and hierarchical state representation based on the multi-source behavior data set, fuse the behavior path representation and the emotional interaction features to generate a loss signal representation, and configure early warning judgment parameters in the hierarchical state representation; In this embodiment, the object binding tags, event occurrence time, interaction action name, page dwell time, function usage record, and repeated trigger record in the multi-source behavioral data set can be organized into behavioral event fragments. These behavioral event fragments are grouped according to the object binding tags and arranged according to the event occurrence time. Node jumps, dwell time changes, repeated triggers, and path interruptions are used to generate a behavioral path representation.

[0022] Emotional interaction features can be extracted from evaluation content, complaint content, customer service dialogue content, and voice semantic analysis results to form markers such as emotional tendency, service dissatisfaction, interaction blockage, and emotional fluctuation. These markers are matched with behavioral event segments according to the time of event occurrence, ensuring that behavioral changes and feedback changes occur within the same time segment.

[0023] The hierarchical status representation can be generated based on the object's basic tags, object grouping attributes, object service status, object usage cycle, historical contribution tags, and recent activity tags. Warning judgment parameters are configured according to the hierarchical status representation, allowing different trigger conditions to correspond to different object states. Behavioral path representation and emotional interaction features are fused according to time segments to generate a churn signal representation.

[0024] The event fragment approach can be used to read the interaction actions, dwell records, function usage records and repeated trigger records of the same object from a multi-source behavioral data set to form behavioral event fragments, and then generate behavioral path representations based on the jumps, dwells and interruptions between fragments.

[0025] Alternatively, a time-slice fusion approach can be used to convert evaluation content, complaint content, customer service dialogue content, and voice semantic analysis results within the same time range into emotional interaction features, which are then aligned with behavioral path representations to generate churn signal representations.

[0026] A hierarchical configuration approach can also be adopted, which generates hierarchical status representations based on object usage cycle, object service status, object grouping attributes, and recent activity markers, and configures warning judgment parameters for different hierarchical statuses.

[0027] This embodiment combines behavioral path representation, emotional interaction features, and hierarchical state representation. Continuous behavioral changes, feedback changes, and object states can all participate in the generation of churn signal representation, reducing judgment bias caused by relying solely on a single behavioral indicator, and enabling the warning judgment parameters to be adjusted according to changes in object state.

[0028] S30, input the churn signal representation and the hierarchical state representation into the fusion prediction model, and output the churn risk score, churn stage identification result and churn reason candidate set; In this embodiment, the churn signal representation may include signal segment markers, segment time spans, signal strength markers, and behavioral sentiment deviation markers, used to express the behavioral and feedback changes of an object during continuous interaction. The hierarchical status representation may include periodic stage markers, object level markers, service status markers, group attribute markers, and warning judgment parameters, used to express the current service status and judgment conditions of an object.

[0029] The fusion prediction model can be configured with segment time-series branches, stage identification branches, risk discrimination branches, and cause identification branches. The segment time-series branch extracts short-term fluctuation features, continuous decay features, and abnormally concentrated segment markers from the churn signal representation; the stage identification branch combines hierarchical state representation to generate churn stage identification results; the risk discrimination branch combines signal strength, time span, and hierarchical state to generate a churn risk score; and the cause identification branch combines behavioral and emotional deviation markers and churn stage identification results to generate a candidate set of churn causes.

[0030] Hierarchical status representation is used in fusion prediction models to calibrate risk assessment scales for different object states. For the same type of churn signal representation, different object levels, service states, or usage periods can correspond to different risk score outputs and stage identification results, enabling the model output to not only depend on behavioral changes but also to make judgments based on object states.

[0031] A multi-branch neural network structure can be adopted. The churn signal represents the temporal branch of the input segment, and the temporal branch of the segment can use a temporal coding network or an attention network to extract the changes in continuous segments. The hierarchical state represents the input hierarchical calibration branch, which generates hierarchical calibration parameters. The risk discrimination branch outputs a churn risk score based on the temporal features and hierarchical calibration parameters.

[0032] Alternatively, a tree-based discrimination structure can be used, where the fusion prediction model converts the churn signal representation into a sequence of signal segments, converts the hierarchical state representation into a hierarchical condition vector, and then performs branch discrimination according to stage type, risk interval, and cause type, outputting the churn stage identification result, churn risk score, and churn cause candidate set, respectively.

[0033] This embodiment inputs both the churn signal representation and the hierarchical state representation into the fusion prediction model, allowing behavioral changes, feedback changes, and object states to jointly participate in risk assessment. The segment time-series branch extracts continuous changes, the stage identification branch judges the churn process, the risk discrimination branch outputs a risk score, and the cause identification branch generates a set of cause candidates, thereby reducing the bias caused by single indicator judgment and ensuring that the subsequent construction of cause evidence has a clear data source.

[0034] S40, construct a cause evidence vector based on the churn risk score, the churn stage identification result and the churn cause candidate set, match the cause evidence vector with historical retention response records to generate a candidate retention case set, obtain a strategy knowledge graph, and generate a candidate retention strategy set based on the hierarchical state representation, the cause evidence vector, the candidate retention case set and the strategy knowledge graph; In this embodiment, the churn risk score represents the risk intensity, the churn stage identification result represents the stage position of the object in the churn formation process, and the churn cause candidate set represents the cause categories and corresponding signal segments that may lead to churn. These three types of results can be written into the cause evidence slot set to form a cause evidence vector. The cause evidence vector not only stores the risk level but also the stage status, cause type, signal segment, and hierarchical trigger information, enabling subsequent matching to make combined judgments based on risk, stage, and cause.

[0035] Historical retention response records are used to preserve feedback from past recipients who accepted retention strategies under similar churn conditions. Record content may include historical causal evidence vectors, historical stratified state representations, historical retention strategies, historical response states, historical feedback snippets, and case source tags. When matching causal evidence vectors with historical retention response records, similar records can be filtered first based on the historical causal evidence vectors, and then state overlap tags can be generated based on the stratified state representations and historical stratified state representations, forming a candidate retention case set. The candidate retention case set retains historical retention strategies, historical response evaluation tags, and case source tags, facilitating the reference of existing response results during subsequent strategy generation.

[0036] The strategy knowledge graph is used to organize the relationships between object types, churn reasons, retention resources, execution endpoints, resource usage, disabling conditions, historical response status, and case sources. Hierarchical state representations can be mapped to object type nodes, reason evidence vectors can be mapped to churn reason nodes, and the candidate retention case set can be mapped to historical response status nodes and case source nodes. Based on the set of strategy association edges between nodes, paths that simultaneously cover object status, churn reasons, historical response, and available resources can be filtered. Retention resource markers, execution endpoint markers, resource usage markers, disabling condition markers, historical response status markers, and case source markers are extracted from these paths to generate a candidate retention strategy set.

[0037] A vector matching approach can be used. The server writes the churn risk score, churn stage identification results, and candidate churn reasons into a set of reason evidence slots, generating a reason evidence vector. Historical retention response records are indexed according to historical reason evidence vectors. During matching, the similarity between the reason evidence vector and the historical reason evidence vector is compared, and then combined with hierarchical status representation to filter historical records, generating a set of candidate retention cases. This approach is suitable for scenarios in fintech businesses with a large number of customers and a sufficient accumulation of historical response records.

[0038] Alternatively, a graph-based path filtering method can be used. The strategy knowledge graph pre-defines nodes for object type, churn reason, retention resource, execution end, resource consumption, disabling conditions, historical response status, and case source. The system maps hierarchical status representations, causal evidence vectors, and candidate retention case sets to corresponding nodes. It then filters candidate strategy paths using a set of strategy-related edges, extracts corresponding tags from these paths, and forms a candidate retention strategy set. This approach is suitable for scenarios where multiple strategy resources coexist in banking services, insurance services, payment services, and customer operation platforms.

[0039] A hybrid screening approach can also be used. The causal evidence vector is first used to narrow down the range of historical retention response records, and the candidate retention case set is then entered into the strategy knowledge graph for path filtering. Historical response evaluation tags are used to exclude cases with poor feedback, disabled condition nodes are used to exclude strategy paths that do not meet the current reach conditions, and resource usage nodes are used to control the scope of strategy resource usage.

[0040] This embodiment converts churn risk scores, churn stage identification results, and a candidate set of churn causes into causal evidence vectors, enabling risk assessment results to form a matching data representation. By matching the causal evidence vectors with historical retention response records, a set of candidate retention cases with similar risk and causal characteristics can be obtained. Furthermore, by combining hierarchical state representation and strategy knowledge graphs, a set of candidate retention strategies is generated, creating a continuous association between churn causes, object status, historical responses, and strategy resources, thereby reducing the disconnect between early warning results and retention strategy generation.

[0041] S50, conduct an implementation feasibility assessment on the candidate retention strategy set, perform response deduction based on the assessment results, and generate a target retention strategy; In this embodiment, the candidate retention strategy set may include retention resource markers, execution terminal markers, strategy resource usage markers, disabling condition markers, historical response status markers, and case source markers. The implementation feasibility assessment is used to determine whether the candidate retention strategy can be pushed and executed under current conditions. The assessment content may include resource availability, execution terminal available time periods, reach frequency, reach interval, target audience preferred channels, service personnel availability status, and resource usage limits.

[0042] The evaluation results can be generated from an implementation feasibility matrix. This matrix includes resource availability markers, execution end reachability markers, resource occupancy markers, and disabled hit markers, used to filter the set of implementable retention strategies. Response simulation is performed within this set of implementable retention strategies. A set of strategy action sequences is generated based on retention resource markers, execution end markers, historical response status markers, and case source markers. These sequences are then matched with historical strategy response data to generate response simulation results and determine the target retention strategy.

[0043] A matrix evaluation approach can be used to compare the retention resource markers, execution terminal markers, strategy resource occupancy markers, and disabling condition markers in the candidate retention strategy set with resource inventory, execution terminal available time periods, reach frequency, reach interval, and channel preference to generate an implementation feasibility matrix and screen the set of implementable retention strategies.

[0044] Alternatively, a response trajectory matching method can be used to convert the retention resource markers and execution terminal markers in the set of implementable retention strategies into a set of strategy action types. This set of strategy action sequences is then generated by combining historical response status markers and case source markers, and finally matched with historical strategy response data to obtain response deduction results.

[0045] This embodiment, through feasibility assessment, can eliminate candidate strategies that are unavailable due to resource unavailability, unreachable execution end, or limited reach conditions before push; through response simulation, it can compare the response trends of different strategy action sequences within the feasible range, so that the target retention strategy can simultaneously meet the execution conditions and historical response basis.

[0046] S60, when the churn risk score reaches the warning judgment parameter, the target retention strategy is pushed to the corresponding execution end, the feedback data collected for the target retention strategy is converted into strategy response features, the strategy response features are associated with the churn signal representation to generate strategy execution results, and the fusion prediction model and the strategy knowledge graph are updated according to the strategy execution results.

[0047] In this embodiment, when the churn risk score reaches the warning judgment parameter, a trigger judgment flag can be generated to record the risk trigger status and trigger time. The target retention strategy can read the execution end flag, retention resource flag, and strategy action sequence. The execution end flag is used to identify the corresponding execution end, and the retention resource flag and strategy action sequence are used to generate the strategy push payload. The push content flag and reach time window can be configured according to the strategy push payload, so that the target retention strategy sends the content to the corresponding execution end according to the specified content and time range.

[0048] Feedback data can be collected based on policy push records and target retention policies. Content may include resource usage feedback, page revisit feedback, service reach feedback, message response feedback, complaint change feedback, login continuation feedback, and retention status feedback. Resource usage feedback can be converted into resource response tags, page revisit feedback and login continuation feedback can be converted into behavior recovery tags, service reach feedback and message response feedback can be converted into reach response tags, and complaint change feedback and retention status feedback can be converted into status change tags. These tags are written into the policy response features according to the time the feedback occurred.

[0049] When policy response characteristics are correlated with churn signal representations, attribution relationships can be established through policy push records, generating response attribution markers, signal change markers, and policy status markers. Policy execution results are formed by combining these markers, describing changes in object behavior, feedback, and status after the target retention policy is pushed. Model update data can update the segment time-series branch, risk discrimination branch, stage identification branch, and cause identification branch in the fusion prediction model based on policy execution results; graph update data can update the retention resource nodes, execution end nodes, historical response status nodes, case source nodes, disabling condition nodes, and policy association edge sets in the policy knowledge graph.

[0050] An event-triggered approach can be adopted. The financial service platform monitors churn risk scores and early warning parameters. Once the trigger conditions are met, a strategy push payload is generated, and the target retention strategy is sent to the account manager's terminal, messaging terminal, or business page terminal. After the corresponding execution terminal returns a reach confirmation message, a strategy push record is generated to constrain the scope of subsequent feedback collection.

[0051] Alternatively, a feedback aggregation approach can be used. The system collects customer usage of retention resources, page return visits, message responses, and changes in complaints based on policy push records, and converts different feedback into policy response characteristics. After correlating these policy response characteristics with the churn signals before triggering the policy, and then correlating them with the target audience, the policy execution result is formed.

[0052] This embodiment pushes a target retention strategy when the churn risk score reaches the warning judgment parameter, and converts the feedback data into strategy response features. This allows resource usage, behavior recovery, reach response, and status changes after strategy execution to be included in the same result. By associating the strategy response features with the churn signal representation to generate strategy execution results, and then updating the fusion prediction model and strategy knowledge graph, the warning, push, feedback, and update can form a continuous processing process, reducing the problem of strategy execution results being difficult to return.

[0053] In one embodiment, step S10 includes: S101, obtain interaction record data from different interaction channels, and extract the interaction entry point, session tag, event occurrence time, business object tag and interaction action name from the interaction record data to generate an interaction source signature; S102, extract the object basic marker, object channel attribute, object grouping attribute, object service status and object usage period from the object attribute data to generate an object attribute profile; S103, Based on the interaction source signature and the object attribute profile, perform identifier binding, generate object candidate identifier cluster, and record candidate identifier, data source, occurrence time and binding trusted marker in the object candidate identifier cluster; S104, generate an identifier verification fragment based on the candidate identifier, data source, occurrence time and binding trusted marker in the object candidate identifier cluster, and bind the identifier verification fragment with the corresponding interaction record data; S105, extract the event occurrence time, data reception time, and business completion time from the interaction record data, generate event time anchors, and perform time-series integration of the interaction record data and the object attribute data based on the event time anchors, the object candidate identifier cluster, and the identifier verification fragment to generate an object time-series fragment set; S106, merge the interaction source signature, the object attribute profile, the object candidate identifier cluster, the identifier verification fragment, and the object time sequence fragment set; generate an object binding tag based on the object candidate identifier cluster and the identifier verification fragment; and generate an abnormal fragment tag based on missing time, repeated interaction, abnormal stay, abnormal jump, and text transcription anomaly; write the object binding tag and the abnormal fragment tag into the merged result to generate a multi-source behavior data set.

[0054] In this embodiment, different interaction channels may include application terminals, web pages, customer service terminals, messaging terminals, and business processing terminals. Interaction records generated by each channel are written into a unified field format during collection. These fields include interaction entry point, session marker, event occurrence time, business object marker, and interaction action name. The interaction entry point distinguishes whether the data originates from page access, customer service communication, message delivery, or business processing. The session marker distinguishes a continuous interaction process. The event occurrence time determines the position of the interaction record on the timeline. The business object marker indicates the product, service, or business matter involved in the interaction. The interaction action name indicates the action type, such as browsing, clicking, submitting, consulting, complaining, or exiting. The combination of these fields forms an interaction source signature, giving interaction records generated by different channels an identifiable source structure, facilitating subsequent binding with object attribute data.

[0055] Object attribute data is extracted from object archives, channel records, group records, service status records, and usage period records. The object's basic identifier represents its fundamental identification information within the business system; the object's channel attribute indicates its primary access or reach channel; the object's group attribute indicates its group category; the object's service status indicates its current service progress or response status; and the object's usage period indicates the duration or phase interval between the object and the business service. These elements are combined into an object attribute profile, ensuring that object attributes are not stored as scattered fields but in a structured format that allows for binding, integration, and subsequent retrieval.

[0056] Identifier binding takes the interaction source signature and object attribute profile as input. The interaction entry point, session tag, business object tag, and interaction action name in the interaction source signature can be matched with the object base tag, object channel attribute, and object grouping attribute in the object attribute profile. The matching result is not directly fixed as a single object, but rather generates a candidate identifier cluster. The candidate identifier cluster records the candidate identifier, data source, occurrence time, and binding trust mark. The candidate identifier stores the possible corresponding object, the data source stores which channel or record the candidate identifier comes from, the occurrence time records the position of the candidate identifier in the data, and the binding trust mark indicates the matching strength between the candidate identifier and the current interaction record. This structure preserves multiple candidate cases in cross-channel binding and reduces erroneous merging caused by missing single identifiers or inconsistent channel identifiers.

[0057] The identifier verification fragment is generated based on candidate identifiers, data sources, occurrence times, and bound trusted markers from the object's candidate identifier cluster. During generation, the source trustworthiness, occurrence time continuity, and field overlap with interaction records of different candidate identifiers can be compared, and the verification results are recorded. After the identifier verification fragment is bound to the corresponding interaction record data, the verification basis for the interaction record's object attribution can be retained. In cases where multiple candidate identifiers exist simultaneously, the identifier verification fragment can retain conflict sources and priority binding criteria; in cases where candidate identifiers are identical, the identifier verification fragment can record stable binding criteria. This ensures that interaction records are not directly discarded due to identifier uncertainty during subsequent time-series integration.

[0058] Event time anchors are generated from the event occurrence time, data reception time, and business completion time. The event occurrence time indicates when the interaction occurred; the data reception time indicates when the interaction record entered the data processing environment; and the business completion time indicates when the business action was completed or the state changed. These three types of times are used together to correct the arrangement of interaction records. For records with delayed reporting, asynchronous feedback, or lagging business status updates, their position in the time series can be adjusted using event time anchors to avoid time misalignment caused by arranging them solely according to data reception time.

[0059] Time-series integration is based on event time anchors, object candidate identifier clusters, and identifier verification fragments. Interaction record data is arranged according to event time anchors, and object attribute data is written to the corresponding time range according to attribute generation time, update time, or valid interval. Object candidate identifier clusters are used to determine the attribution of interaction records, and identifier verification fragments are used to store the attribution verification basis. The integration result forms a set of object time-series fragments. Each object time-series fragment can contain interaction records, corresponding attribute states, time anchors, and binding basis, enabling data of the same object generated on different channels and at different times to be arranged under the same time structure.

[0060] Interaction source signatures, object attribute profiles, object candidate identifier clusters, identifier verification fragments, and object time-series fragment sets are merged into a unified result. Object binding tags, generated from object candidate identifier clusters and identifier verification fragments, represent the attribution identifier of the same object in the merged result. Abnormal fragment tags are generated based on missing time, repeated interactions, abnormal pauses, abnormal jumps, and text transcription anomalies. Missing time tags mark records with missing or unlocatable time fields; repeated interactions mark similar actions occurring repeatedly within a short period; abnormal pauses mark fragments with pause durations significantly different from adjacent records; abnormal jumps mark discontinuous transitions between pages or business nodes; and text transcription anomalies mark content with transcription omissions or identification errors in the text source. After object binding tags and abnormal fragment tags are written into the merged result, a multi-source behavioral data set is formed, enabling subsequent readings to simultaneously obtain interaction content, attribute status, object attribution, time sequence, and abnormal tags.

[0061] This embodiment preserves the source structure of interaction records from different channels through interaction source signatures, saves the object attribute status through object attribute profiles, preserves the basis for cross-channel identifier binding through object candidate identifier clusters and identifier verification fragments, unifies the time arrangement of interaction records and object attribute data through event time anchors and object time sequence fragment sets, and writes the merged result through object binding tags and abnormal fragment tags. This enables the multi-source behavioral data set to reduce data deviations caused by unclear object ownership, fragmented cross-channel records, misaligned time order, and mixed abnormal data, providing stable data input for subsequent analysis based on continuous behavior of the same object.

[0062] In one embodiment, step S20 above includes: S201, read object binding markers, interaction action names, business object markers, event occurrence times, page dwell records, function usage records, and repeated trigger records from the multi-source behavior data set, and generate a set of behavior event fragments based on the object binding markers, interaction action names, business object markers, event occurrence times, page dwell records, function usage records, and repeated trigger records; S202, Based on the jump relationship, dwell change, repeated triggering state and path interruption state between adjacent behavior event segments in the behavior event segment set, generate a path transfer mark set, and write the path transfer mark set into the behavior event segment set to generate a behavior path representation; S203, read the evaluation content, complaint content, customer service dialogue content and voice semantic analysis results from the multi-source behavioral data set, and generate emotion tendency markers, service dissatisfaction markers, interaction blockage markers and emotion fluctuation markers based on the evaluation content, complaint content, customer service dialogue content and voice semantic analysis results; S204, establish a segment correspondence between the emotion tendency marker, the service dissatisfaction marker, the interaction blockage marker, and the emotion fluctuation marker according to the corresponding event occurrence time and the set of behavioral event segments to generate emotional interaction features; S205, read the object basic tag, object grouping attribute, object service status, object usage period, historical contribution tag and recent activity tag from the multi-source behavior data set, and generate periodic stage tag, value level tag, service status tag and group attribute tag based on the object basic tag, the object grouping attribute, the object service status, the object usage period, the historical contribution tag and the recent activity tag; S206, combine the periodic stage marker, the value level marker, the service status marker, and the group attribute marker to generate a hierarchical status representation, and configure early warning judgment parameters in the hierarchical status representation based on the periodic stage marker, the value level marker, the service status marker, and the group attribute marker; S207, The behavioral path representation and the emotional interaction features are aligned at the fragment level according to the event occurrence time to generate a behavioral-emotional aligned fragment set, and a churn signal representation is generated based on the behavioral-emotional aligned fragment set.

[0063] In this embodiment, object binding tags are used to aggregate records from multiple sources for the same object; interaction action names are used to distinguish action types such as browsing, clicking, submitting, exiting, consulting, and complaining; business object tags are used to identify the business object involved in the action; event occurrence time is used to determine the arrangement position; and page dwell time records, function usage records, and repeated trigger records are used to reflect dwell time, invocation, and repeated operations. When generating a set of behavioral event fragments, records can be merged according to object binding tags, arranged according to event occurrence time, and the interaction action name, business object tag, page dwell time record, function usage record, and repeated trigger record can be written into the corresponding fragment.

[0064] Jump relationships, pause changes, repeated trigger states, and path interruption states are used to describe the changes between adjacent behavioral event segments. Based on this information, a set of path transition markers is generated and written into the set of behavioral event segments to form a behavioral path representation, allowing the transition direction, pause anomalies, repeated operations, and interruption locations in continuous interactions to be preserved.

[0065] Evaluation content, complaint content, customer service dialogue content, and voice semantic analysis results are used to generate emotion tendency markers, service dissatisfaction markers, interaction blockage markers, and emotion fluctuation markers. These markers establish a segment correspondence based on the event occurrence time and a set of behavioral event segments, forming emotional interaction features that allow feedback content to match corresponding behavioral segments.

[0066] Object base tags, object grouping attributes, object service status, object usage period, historical contribution tags, and recent activity tags are used to generate periodic stage tags, value level tags, service status tags, and group attribute tags. These tags combine to form a hierarchical status representation, and warning judgment parameters are configured within this hierarchical status representation, enabling different object states to correspond to different triggering conditions.

[0067] Behavioral path representations and emotional interaction features are aligned at the fragment level according to the event occurrence time, forming a set of behavioral-emotional aligned fragments. Churn signals are generated from the set of behavioral-emotional aligned fragments and can simultaneously retain information such as path changes, dwell time changes, repeated triggers, interaction interruptions, emotional tendencies, and service dissatisfaction.

[0068] This embodiment forms a behavioral path representation by using a set of behavioral event fragments and a set of path transition markers, which can preserve continuous interactive changes; by converting feedback content into emotional interaction features and corresponding them to behavioral event fragments, it can reduce the deviation caused by the separation of behavioral data and feedback content; by configuring early warning judgment parameters through hierarchical state representation, different object states can have corresponding trigger conditions; by generating a churn signal representation through fragment-level alignment, it can provide continuous behavioral changes, feedback changes, and object state differences for subsequent risk identification.

[0069] In one embodiment, step S30 above includes: S301, the lost signal representation and the hierarchical state representation are input into the fusion prediction model, and the signal segment marker, segment time span, signal strength marker and behavioral emotion deviation marker are parsed from the lost signal representation in the fusion prediction model to generate a signal segment sequence; S302, Read the cycle stage marker, value level marker, service status marker, group attribute marker and early warning judgment parameter from the hierarchical status representation, and generate hierarchical condition vector and hierarchical calibration parameter; S303, based on the signal segment sequence, the segment time span, the signal strength marker, and the hierarchical condition vector, generate short-term fluctuation features, continuous attenuation features, and abnormally concentrated segment markers in the segment time-series branch of the fusion prediction model; S304, Based on the hierarchical calibration parameters, the short-term fluctuation characteristics, the continuous decay characteristics, and the abnormal concentrated fragment markers, generate the loss stage identification result in the stage identification branch of the fusion prediction model; S305, Based on the hierarchical calibration parameters, the signal strength marker, the short-term fluctuation characteristics, the continuous attenuation characteristics, and the abnormal concentrated segment marker, a loss risk score is generated in the risk discrimination branch of the fusion prediction model; S306, Based on the behavioral and emotional deviation markers, the short-term fluctuation features, the continuous decay features, the abnormal concentrated segment markers, and the loss stage identification results, a set of candidate causes is generated in the cause identification branch of the fusion prediction model; S307, generate a hierarchical triggering flag based on the hierarchical condition vector, and bind a cause type flag, a signal segment flag corresponding to the cause candidate flag, and a hierarchical triggering flag to each cause candidate in the cause candidate set to generate a churn cause candidate set.

[0070] In this embodiment, the lost signal is represented in the fusion prediction model as a signal segment marker, segment time span, signal strength marker, and behavioral-emotional deviation marker. The signal segment marker distinguishes risk signals within different time segments, the segment time span represents the duration of the signal, the signal strength marker represents the degree of risk change within the segment, and the behavioral-emotional deviation marker represents the deviation between behavioral changes and emotional feedback. These elements are combined to form a signal segment sequence, enabling the model to read continuous changes segment by segment, rather than reading only a single risk input.

[0071] The periodic stage markers, value level markers, service status markers, group attribute markers, and early warning judgment parameters in the hierarchical state representation are converted into hierarchical condition vectors and hierarchical calibration parameters. The hierarchical condition vectors represent the current state category of an object, while the hierarchical calibration parameters adjust the stage identification and risk discrimination scales under different states. This allows the same signal segment sequence to be processed under different discrimination conditions at different periodic stages, service states, or group attributes.

[0072] The segment time-series branch generates short-term fluctuation features, continuous decay features, and anomaly concentration segment markers based on the signal segment sequence, segment time span, signal strength markers, and hierarchical condition vectors. Short-term fluctuation features are used to represent the fluctuations of risk signals within a short period of time, continuous decay features are used to represent the continuous decline in behavioral activity or feedback status, and anomaly concentration segment markers are used to represent the concentrated occurrence of multiple anomalous signals within adjacent time segments.

[0073] The stage identification branch generates churn stage identification results based on hierarchical calibration parameters, short-term fluctuation characteristics, continuous decay characteristics, and abnormal concentrated segment markers. The risk discrimination branch adds signal strength markers to the above inputs to generate a churn risk score. The stage identification result is biased towards judging the position of the churn process, while the churn risk score is biased towards outputting risk intensity. The two are generated by different branches, which can avoid the stage judgment and risk intensity judgment being mixed into a single output.

[0074] The cause identification branch generates a set of candidate causes based on behavioral and emotional deviation markers, short-term fluctuation characteristics, continuous decay characteristics, abnormal concentrated segment markers, and churn stage identification results. Each candidate cause in the set is bound to a cause type marker, a corresponding signal segment marker, and a hierarchical trigger marker. The hierarchical trigger marker is generated by a hierarchical condition vector and is used to record the triggering relationship between the candidate cause and the object state, enabling the churn cause candidate set to simultaneously retain cause category, signal location, and hierarchical state basis.

[0075] This embodiment represents the loss signal as a sequence of signal segments and combines it with the hierarchical state representation to generate hierarchical condition vectors and hierarchical calibration parameters. The fusion prediction model can simultaneously utilize continuous signal changes and object state differences. By extracting short-term fluctuations, continuous decay, and anomaly concentration information through segment time-series branches, it generates loss stage identification results, loss risk scores, and a set of candidate loss causes, respectively. This can reduce the problem that a single output cannot simultaneously express the risk intensity, loss process, and source of cause, and make the subsequent construction of cause evidence vectors have clearer data basis.

[0076] In one embodiment, step S40 above includes: S401, generate risk evidence markers based on the churn risk score, generate stage evidence markers based on the churn stage identification results, and read cause type markers, signal fragment markers, and hierarchical trigger markers from the churn cause candidate set; S402, the risk evidence marker, the stage evidence marker, the cause type marker, the signal fragment marker, and the hierarchical trigger marker are written into the cause evidence slot set, and a cause evidence vector is generated based on the cause evidence slot set; S403, read the historical cause evidence vector, historical hierarchical status representation, historical retention strategy, historical response status, historical feedback fragment and case source tag from the historical retention response record, generate a historical response evaluation tag based on the historical response status and the historical feedback fragment, and write the historical cause evidence vector, the historical hierarchical status representation, the historical retention strategy, the historical response evaluation tag and the case source tag into the historical retention response index; S404, match the cause evidence vector with the historical retention response index, generate a state overlap marker based on the hierarchical state representation and the historical hierarchical state representation, filter the historical retention response records corresponding to the historical retention response index based on the state overlap marker and the historical response evaluation marker, generate a candidate retention case set, and bind a case source marker, historical retention strategy and historical response evaluation marker to the candidate retention cases in the candidate retention case set; S405, obtain a preset strategy knowledge graph, and read object type nodes, churn reason nodes, retention resource nodes, execution end nodes, strategy cost nodes, disabling condition nodes, historical response status nodes, case source nodes and strategy association edge set from the strategy knowledge graph, and generate a candidate strategy path set based on the object type nodes, the churn reason nodes, the retention resource nodes, the execution end nodes, the strategy cost nodes, the disabling condition nodes, the historical response status nodes, the case source nodes and the strategy association edge set; S406, based on the hierarchical state representation, determine the target object type node in the object type node; based on the cause evidence vector, determine the target churn cause node in the churn cause node; based on the historical retention strategies in the candidate retention case set, determine the candidate resource execution node group in the retention resource node and the execution end node; based on the candidate resource execution node group, determine the target strategy cost node in the strategy cost node; based on the candidate resource execution node group, determine the target disabling condition node in the disabling condition node; based on the historical response evaluation flag and case source flag in the candidate retention case set, determine the target historical response status node and the target case source node respectively; and filter the target strategy path set from the candidate strategy path set that simultaneously connects the target object type node, the target churn cause node, the retention resource node and execution end node in the candidate resource execution node group, the target strategy cost node, the target disabling condition node, the target historical response status node, and the target case source node. S407, extract retention resource markers, execution end markers, policy cost markers, disabling condition markers, historical response status markers, and case source markers from the target policy path set to generate a candidate retention policy set.

[0077] In this embodiment, the churn risk score can be converted into risk evidence tags, which represent the risk intensity range and trigger level. The churn stage identification result can be converted into stage evidence tags, which represent the object's stage position, such as early fluctuation, continuous decline, or imminent churn. The churn cause candidate set can read cause type tags, signal fragment tags, and hierarchical trigger tags. Cause type tags distinguish cause categories such as service response, resource matching, operation interruption, and negative feedback. Signal fragment tags record the time segment in which the cause occurs. Hierarchical trigger tags record the triggering relationship between the cause and the object's state. After these tags are written into the cause evidence slot set, cause evidence vectors can be generated according to the five categories of slots: risk, stage, cause, fragment, and hierarchical, enabling the prediction output to form a matching data representation.

[0078] Historical response records can store historical cause evidence vectors, historical hierarchical status representations, historical retention strategies, historical response statuses, historical feedback snippets, and case source tags. Historical response statuses and historical feedback snippets can generate historical response evaluation tags, used to represent the response results after the historical retention strategy is implemented. After the historical cause evidence vectors, historical hierarchical status representations, historical retention strategies, historical response evaluation tags, and case source tags are written into the historical response index, they can be used to match with the current cause evidence vector. During matching, the cause evidence vector is used to filter records with similar causes and risks; the hierarchical status representation and the historical hierarchical status representation are used to generate status overlap tags; the status overlap tags and historical response evaluation tags jointly filter historical response records, generating a candidate retention case set. Candidate retention cases in the candidate retention case set are bound to case source tags, historical retention strategies, and historical response evaluation tags, ensuring that each candidate retention case retains its source, strategy content, and historical response basis.

[0079] The strategy knowledge graph can store nodes for object type, churn reason, retention resource, execution end, strategy cost, disabling condition, historical response status, case source, and a set of strategy-related edges. Object type nodes represent hierarchical statuses; churn reason nodes represent causal evidence vectors; retention resource and execution end nodes represent available resources and push terminals; strategy cost nodes represent resource usage; disabling condition nodes represent reach or execution restrictions; and historical response status and case source nodes represent historical response evaluation tags and case source tags from the candidate retention case set. The combination of nodes and the set of strategy-related edges generates a set of candidate strategy paths, creating a filterable path structure between different nodes.

[0080] The target strategy path set can be generated through filtering by multiple types of nodes. Hierarchical status representation is used to identify target object type nodes within object type nodes; cause evidence vectors are used to identify target churn cause nodes within churn cause nodes; and historical retention strategies in the candidate retention case set are used to identify candidate resource execution node groups within retention resource nodes and execution end nodes. Candidate resource execution node groups are further used to identify target strategy cost nodes and target disabling condition nodes. Historical response evaluation tags and case source tags are used to identify target historical response status nodes and target case source nodes, respectively. Paths in the candidate strategy path set that simultaneously connect the above nodes are filtered to form the target strategy path set. Then, retention resource tags, execution end tags, strategy cost tags, disabling condition tags, historical response status tags, and case source tags are extracted from the target strategy path set to generate the candidate retention strategy set.

[0081] This embodiment converts churn risk scores, churn stage identification results, and churn reason candidate sets into reason evidence vectors, enabling risk intensity, stage location, and reason source to enter a unified matching process. By filtering historical retention response records through historical retention response indexes and state overlap markers, a set of candidate retention cases similar to the current object state and churn reason can be obtained. By filtering the target strategy path set through the set of nodes and strategy association edges in the strategy knowledge graph, object state, churn reason, historical response, retention resources, execution end, and disabling conditions can all participate in the generation of candidate retention strategy set, thereby reducing the data gap between warning results and retention strategies.

[0082] In one embodiment, step S50 above includes: S501, read the retention resource marker, execution terminal marker, strategy cost marker, disabling condition marker, historical response status marker, and case source marker from the candidate retention strategy set, and generate a candidate strategy decomposition table based on the retention resource marker, the execution terminal marker, the strategy cost marker, the disabling condition marker, the historical response status marker, and the case source marker; S502, acquire implementation constraint data, and extract resource inventory markers, execution terminal available time periods, contact frequency records, contact interval markers, object preference channel markers, service personnel availability status and cost occupancy limits from the implementation constraint data; S503, generate a resource availability marker based on the retained resource marker and the resource inventory marker in the candidate strategy decomposition table, and generate an execution end reachability marker based on the execution end marker, the execution end available time period and the object preference channel marker in the candidate strategy decomposition table; S504, generate a cost occupancy marker based on the strategy cost marker in the candidate strategy decomposition table and the cost occupancy limit, and generate a disable hit marker based on the disable condition marker, reach frequency record, reach interval marker and service personnel availability status in the candidate strategy decomposition table; S505, the resource availability flag, the execution end reachability flag, the cost occupancy flag, and the disabled hit flag are written into the implementation feasibility matrix, an evaluation result is generated based on the implementation feasibility matrix, and an implementable retention strategy set is selected from the candidate retention strategy set based on the evaluation result; S506, Read the retention resource marker and execution terminal marker from the set of implementable retention strategies, and generate a set of strategy action types based on the retention resource marker and the execution terminal marker; S507, Based on the historical response status markers corresponding to the set of implementable retention strategies in the candidate strategy decomposition table, the case source markers corresponding to the set of implementable retention strategies in the candidate strategy decomposition table, and the set of strategy action types, a strategy response status chain is generated, and the action sequence of the set of strategy action types is arranged based on the strategy response status chain to generate a set of strategy action sequences. S508, acquire historical strategy response data, match the strategy action sequence set with the historical strategy response data to generate response deduction results for each strategy action sequence; S509, Based on the response deduction results, a target retention strategy is determined from the set of implementable retention strategies, and the retention resource marker, execution end marker, and strategy action sequence corresponding to the target retention strategy are bound to the target retention strategy.

[0083] In this embodiment, the retention resource marker, execution terminal marker, policy cost marker, disabling condition marker, historical response status marker, and case source marker in the candidate retention policy set are decomposed to form a candidate policy decomposition table. The retention resource marker records the type of resource the policy needs to invoke; the execution terminal marker records the endpoint location the policy needs to reach; the policy cost marker records the amount of resources the policy occupies; the disabling condition marker records the conditions under which the policy cannot be triggered; the historical response status marker records the historical feedback status of similar policies; and the case source marker records the basis for the policy's origin. The candidate policy decomposition table aggregates these markers item by item according to the candidate policy, enabling subsequent evaluation to judge each candidate policy separately.

[0084] The constraint data provides the execution conditions under the current environment. Resource inventory flags determine the availability of retained resources; execution end availability time periods determine whether the execution end can receive the policy; reach frequency records and reach interval flags determine if there is overreach; object preference channel flags determine if the execution end and object receiving habits match; service personnel availability status determines if human service resources are schedulable; and cost occupancy limit determines if policy occupancy exceeds a preset range. After comparing the above data with the corresponding flags in the candidate policy decomposition table, resource availability flags, execution end reachability flags, cost occupancy flags, and disable hit flags are generated.

[0085] The feasibility matrix stores the evaluation results of each candidate strategy in terms of resources, execution endpoint, cost, and disabling conditions. After resource availability, execution endpoint reachability, cost occupancy, and disabling hit conditions are written into the same matrix, the executable status of the candidate strategies can be determined. The evaluation results are generated based on the feasibility matrix and used to filter the set of feasible retention strategies from the candidate retention strategy set, ensuring that subsequent response simulations only revolve around the strategies that have passed the evaluation.

[0086] The retention resource markers and execution terminal markers in the set of implementable retention strategies are converted into a set of strategy action types. The strategy action type set represents orchestratable actions such as resource invocation actions, terminal outreach actions, and service guidance actions. The historical response status markers and case source markers corresponding to the set of implementable retention strategies in the candidate strategy decomposition table are combined with the strategy action type set to generate a strategy response status chain. The strategy response status chain records the response status of different action types in similar cases and is used to determine the action order, generating a set of strategy action sequences.

[0087] Historical strategy response data is used to store existing strategy action sequences and their feedback results. When matching the strategy action sequence set with the historical strategy response data, matching can be performed by action type, execution end, resource type, historical response status, and case source to generate response deduction results for each strategy action sequence. The response deduction results are used to determine the target retention strategy from the set of implementable retention strategies, and bind the corresponding retention resource tags, execution end tags, and strategy action sequences to the target retention strategy, enabling the target retention strategy to directly enter subsequent push and feedback collection.

[0088] This embodiment uses a candidate strategy decomposition table to break down the resources, execution end, cost, disabling conditions, and historical response information in the candidate retention strategy set into evaluable data. Then, it uses an implementation feasibility matrix to filter the set of implementable retention strategies, which can reduce the number of unexecutable strategies entering the subsequent simulation. By generating response simulation results through the strategy action type set, strategy response state chain, and historical strategy response data, it is possible to compare the response trends of different action sequences within the feasible range, so that the target retention strategy has both execution conditions and historical response basis.

[0089] In one embodiment, step S60 above includes: S601, monitor the trigger status between the churn risk score and the early warning judgment parameter, and generate a trigger judgment flag when the trigger status is reached. S602, read the execution end marker, retention resource marker and policy action sequence from the target retention policy, and determine the corresponding execution end based on the execution end marker; S603, generate a strategy push payload based on the trigger determination flag, the retention resource flag, the strategy action sequence and the execution end flag, configure a push content flag and a reach time window for the strategy push payload, and push the strategy push payload to the corresponding execution end; S604, receive the reach confirmation information returned by the corresponding execution terminal, and generate a strategy push record based on the reach confirmation information, the strategy push payload and the target retention strategy; S605, Based on the strategy push records and the target retention strategy, collect resource usage feedback, page return feedback, service reach feedback, message response feedback, complaint change feedback, login continuation feedback and retention status feedback, and generate feedback data; S606, convert the resource usage feedback in the feedback data into a resource response flag, convert the page return feedback and login continuation feedback in the feedback data into a behavior recovery flag, convert the service reach feedback and message response feedback in the feedback data into a reach response flag, and convert the complaint change feedback and retention status feedback in the feedback data into a status change flag. S607, the resource response flag, the behavior recovery flag, the reach response flag, and the state change flag are written into the policy response feature according to the feedback occurrence time; S608, based on the strategy push record, associate the strategy response feature with the churn signal representation to generate a response attribution flag, a signal change flag and a strategy status flag, and generate a strategy execution result based on the response attribution flag, the signal change flag and the strategy status flag; S609, Based on the strategy execution result, generate model update data and graph update data. Based on the model update data, update the branch parameters of the segment time-series branch, risk discrimination branch, stage identification branch and cause identification branch in the fusion prediction model. Based on the graph update data, update the node attributes of the retention resource node, execution end node, historical response status node, case source node and disabling condition node in the strategy knowledge graph, as well as the association edge attributes of the strategy association edge set.

[0090] In this embodiment, the trigger state between the churn risk score and the early warning judgment parameter can be obtained through comparison and judgment. The trigger state is used to record whether the risk score has entered the policy push range. When the state is reached to generate a trigger judgment flag, the score source, judgment time, and hierarchical status identifier can be recorded simultaneously, so that subsequent push actions can be consistent with the corresponding risk judgment results. The trigger judgment flag is not retained as a separate output result, but participates in the generation of the policy push payload to distinguish the object status of triggered pushes and non-triggered pushes.

[0091] The target retention strategy reads the execution terminal flag, retention resource flag, and strategy action sequence. The execution terminal flag determines the strategy push location, which can correspond to the messaging terminal, page terminal, customer service terminal, or business processing terminal. The retention resource flag indicates the resource content that the strategy needs to call, and the strategy action sequence indicates the arrangement of actions such as strategy push, page guidance, service outreach, and resource allocation. After determining the corresponding execution terminal based on the execution terminal flag, the trigger judgment flag, retention resource flag, strategy action sequence, and execution terminal flag can be combined into a strategy push payload, and a push content flag and an outreach time window can be configured for the strategy push payload. The push content flag determines the actual content sent, and the outreach time window controls the time range in which the push occurs, preventing the target retention strategy from deviating from the trigger time and execution terminal conditions.

[0092] After the corresponding execution end returns the reach confirmation information, the reach confirmation information, the policy push payload, and the target retention policy can be combined to generate a policy push record. The policy push record records the push recipient, push time, push terminal, push content, reach status, and policy action sequence. Subsequent feedback collection is based on the policy push record. Resource usage feedback, page revisit feedback, service reach feedback, message response feedback, complaint change feedback, login continuation feedback, and retention status feedback can be filtered based on the policy push record and the target retention policy to reduce irrelevant feedback from being mixed into the same policy execution result.

[0093] When feedback data is converted into policy response features, different response tags can be generated according to the feedback type. Resource usage feedback is converted into resource response tags, page revisit feedback and login continuation feedback are converted into behavior recovery tags, service reach feedback and message response feedback are converted into reach response tags, and complaint change feedback and retention status feedback are converted into status change tags. These tags are written into the policy response features according to the time of feedback occurrence, so that the policy response features simultaneously retain the feedback category, feedback time, and feedback change direction.

[0094] Policy push records are used to establish the correlation between policy response characteristics and churn signal representations. During correlation, a response attribution marker can be determined based on the object identifier, push time, feedback occurrence time, and policy action sequence; a signal change marker can be generated based on the difference in churn signals before and after the push; and a policy status marker can be generated based on reach confirmation, resource usage, behavior recovery, and status changes. The response attribution marker, signal change marker, and policy status marker are combined to generate the policy execution result, ensuring that the feedback changes after policy execution correspond to the churn signal representation before triggering.

[0095] The strategy execution results generate model update data and graph update data. Model update data can be categorized into segment time-series branches, risk assessment branches, stage identification branches, and cause identification branches to form parameter update content, used to adjust the signal segment identification, risk scoring, stage judgment, and cause identification processes. Graph update data can update the node attributes of retained resource nodes, execution end nodes, historical response status nodes, case source nodes, and disabling condition nodes in the strategy knowledge graph, and update the association edge attributes of the strategy association edge set, enabling changes in resource response, execution end reach, historical response status, and disabling conditions to be written into the knowledge structure.

[0096] For example, a fintech platform collects customer interaction logs and object attribute data for banking account services, wealth management product services, insurance renewal services, and consumer credit services. Interaction log data includes client clickstreams, page dwell time, product browsing paths, loan application interruption records, insurance policy inquiry records, complaint content, customer service dialogue content, and voice and semantic analysis results. Object attribute data includes customer level, product holding status, channel preference, service status, and usage period. The platform identifies, binds, and integrates data generated from different interaction channels in a chronological order, forming a multi-source behavioral data set, enabling the chronological aggregation of behavioral records of the same customer across different entry points.

[0097] The platform generates behavioral path representations, emotional interaction features, and hierarchical state representations based on multi-source behavioral data sets. When a customer repeatedly views a financial product page and then exits, and later complains about slow service response in customer service conversations, related click paths, page dwell times, product browsing interruptions, complaint content, and customer service semantics are combined to form a churn signal. Customer level, product holding status, service status, and recent activity are combined to form a hierarchical state representation, with corresponding warning and judgment parameters configured.

[0098] After receiving churn signal representations and hierarchical state representations, the fusion prediction model outputs a churn risk score, churn stage identification results, and a candidate set of churn reasons. For example, the model can identify customers in a decline stage and include unsatisfactory service response, decreased product suitability, and service interruption in the candidate set of churn reasons. Based on the churn risk score, churn stage identification results, and candidate set of churn reasons, the platform constructs a causal evidence vector and matches it with historical retention response records to obtain a set of candidate retention cases for similar customers under strategies such as dedicated customer service, product feature guidance, renewal reminders, and benefit activities.

[0099] The platform acquires a strategy knowledge graph and, combined with hierarchical state representations, causal evidence vectors, and a set of candidate retention cases, generates a candidate retention strategy set. This set can include client-side product feature guidance, customer profile prompts for account managers, dedicated customer service outreach, automated message reminders, and promotional benefit notifications. The platform further assesses implementation feasibility based on resource inventory, available execution time slots, outreach frequency, outreach intervals, customer preferred channels, and service personnel availability. It also uses historical strategy response data to perform response simulations and generate a target retention strategy.

[0100] When the churn risk score reaches the warning threshold, the platform pushes the target retention strategy to the corresponding execution endpoint. For example, the account manager receives customer profile prompts and communication suggestions, while the client receives product feature guidance and exclusive activity reminders. The platform collects feedback data such as resource usage, page visits, message responses, changes in complaints, and retention status, converts the feedback data into strategy response features, and correlates them with the churn signals before the trigger to generate strategy execution results. This updates the fusion prediction model and strategy knowledge graph, enabling subsequent risk assessment, cause identification, and retention strategy generation for similar customers to reference the new execution feedback.

[0101] This embodiment establishes a traceable relationship between risk triggering, policy content, and execution-end push by using trigger judgment markers, policy push payloads, and policy push records. By converting feedback data into resource response markers, behavior recovery markers, reach response markers, and state change markers, different feedbacks after policy execution can be unified into processable policy response features. By associating policy response features with churn signal representations to generate policy execution results, the signal change relationship before and after policy push can be preserved. By updating the branch parameters of the fusion prediction model and the node attributes and associated edge attributes of the policy knowledge graph through policy execution results, the problem of policy feedback not being able to flow back to the prediction and policy generation process can be reduced.

[0102] In one embodiment, a behavior analysis-driven churn retention device is provided, which corresponds one-to-one with the behavior analysis-driven churn retention method in the above embodiments. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the functional modules of a preferred embodiment of the behavior analysis-driven churn and retention device of the present invention. The modules include a multi-source behavioral data integration module 10, a churn signal construction module 20, a fusion risk prediction module 30, a strategy knowledge mapping module 40, a retention strategy deduction module 50, and a strategy feedback update module 60. Detailed descriptions of each functional module are as follows: The multi-source behavior data integration module 10 is used to acquire interaction record data and object attribute data, and to identify, bind, and integrate the interaction record data and object attribute data in a time sequence to generate a multi-source behavior data set. The churn signal construction module 20 is used to generate a behavior path representation, emotional interaction features and hierarchical state representation based on the multi-source behavior data set, fuse the behavior path representation and the emotional interaction features to generate a churn signal representation, and configure early warning judgment parameters in the hierarchical state representation; The fusion risk prediction module 30 is used to input the churn signal representation and the hierarchical state representation into the fusion prediction model, and output the churn risk score, churn stage identification result and churn reason candidate set; The strategy knowledge mapping module 40 is used to construct a cause evidence vector based on the churn risk score, the churn stage identification result and the churn cause candidate set, match the cause evidence vector with historical retention response records to generate a candidate retention case set, obtain a strategy knowledge graph, and generate a candidate retention strategy set based on the hierarchical state representation, the cause evidence vector, the candidate retention case set and the strategy knowledge graph; The retention strategy deduction module 50 is used to conduct an implementation feasibility assessment of the candidate retention strategy set, perform response deduction based on the assessment results, and generate a target retention strategy. The strategy feedback update module 60 is used to push the target retention strategy to the corresponding execution end when the churn risk score reaches the warning judgment parameter, convert the feedback data collected for the target retention strategy into strategy response features, associate the strategy response features with the churn signal representation to generate a strategy execution result, and update the fusion prediction model and the strategy knowledge graph according to the strategy execution result.

[0103] In one embodiment, the multi-source behavioral data integration module 10 is specifically used for: Interaction record data is obtained from different interaction channels, and the interaction entry point, session tag, event occurrence time, business object tag and interaction action name are extracted from the interaction record data to generate an interaction source signature; Extract the object's basic identifier, object's channel attribute, object's grouping attribute, object's service status, and object's usage period from the object attribute data to generate an object attribute profile; Based on the interaction source signature and the object attribute profile, an identifier binding is performed to generate an object candidate identifier cluster, and the candidate identifier, data source, occurrence time and binding trusted marker are recorded in the object candidate identifier cluster; Based on the candidate identifiers, data sources, occurrence times, and binding trust tags in the object candidate identifier cluster, an identifier verification fragment is generated, and the identifier verification fragment is bound to the corresponding interaction record data; The event occurrence time, data reception time, and business completion time are extracted from the interaction record data to generate event time anchors. Based on the event time anchors, the object candidate identifier cluster, and the identifier verification fragment, the interaction record data and the object attribute data are integrated in time sequence to generate an object time sequence fragment set. The interaction source signature, the object attribute profile, the object candidate identifier cluster, the identifier verification fragment, and the object time sequence fragment set are merged. An object binding tag is generated based on the object candidate identifier cluster and the identifier verification fragment. An abnormal fragment tag is generated based on missing time, repeated interaction, abnormal stay, abnormal jump, and text transcription abnormality. The object binding tag and the abnormal fragment tag are written into the merged result to generate a multi-source behavior data set.

[0104] In one embodiment, the lost signal construction module 20 is specifically used for: Read object binding tags, interaction action names, business object tags, event occurrence times, page dwell times, function usage times, and repeated trigger records from the multi-source behavior data set, and generate a set of behavior event fragments based on the object binding tags, interaction action names, business object tags, event occurrence times, page dwell times, function usage times, and repeated trigger records; Based on the jump relationship, dwell change, repeated trigger state and path interruption state between adjacent behavior event segments in the behavior event segment set, a path transfer mark set is generated, and the path transfer mark set is written into the behavior event segment set to generate a behavior path representation; The evaluation content, complaint content, customer service dialogue content, and voice semantic analysis results are read from the multi-source behavioral data set, and emotion tendency markers, service dissatisfaction markers, interaction blockage markers, and emotion fluctuation markers are generated based on the evaluation content, complaint content, customer service dialogue content, and voice semantic analysis results. The emotional tendency marker, the service dissatisfaction marker, the interaction blockage marker, and the emotional fluctuation marker are associated with the set of behavioral event fragments according to the corresponding event occurrence time to generate emotional interaction features; Read object base tags, object grouping attributes, object service status, object usage period, historical contribution tags, and recent activity tags from the multi-source behavioral data set, and generate periodic stage tags, value level tags, service status tags, and group attribute tags based on the object base tags, object grouping attributes, object service status, object usage period, historical contribution tags, and recent activity tags; The periodic stage marker, the value level marker, the service status marker, and the group attribute marker are combined to generate a hierarchical status representation, and warning judgment parameters are configured in the hierarchical status representation based on the periodic stage marker, the value level marker, the service status marker, and the group attribute marker. The behavioral path representation and the emotional interaction features are aligned at the fragment level according to the event occurrence time to generate a behavioral-emotional aligned fragment set, and a churn signal representation is generated based on the behavioral-emotional aligned fragment set.

[0105] In one embodiment, the fusion risk prediction module 30 is specifically used for: The lost signal representation and the hierarchical state representation are input into the fusion prediction model, and the signal segment marker, segment time span, signal strength marker and behavioral emotion deviation marker are parsed from the lost signal representation in the fusion prediction model to generate a signal segment sequence. Read the cycle stage marker, value level marker, service status marker, group attribute marker, and early warning judgment parameter from the hierarchical status representation, and generate a hierarchical condition vector and hierarchical calibration parameters; Based on the signal segment sequence, the segment time span, the signal strength marker, and the hierarchical condition vector, short-term fluctuation features, continuous decay features, and abnormally concentrated segment markers are generated in the segment time-series branch of the fusion prediction model. Based on the hierarchical calibration parameters, the short-term fluctuation characteristics, the continuous decay characteristics, and the abnormal concentrated fragment markers, the loss stage identification results are generated in the stage identification branch of the fusion prediction model; Based on the hierarchical calibration parameters, the signal strength marker, the short-term fluctuation characteristics, the continuous attenuation characteristics, and the abnormal concentrated segment marker, a loss risk score is generated in the risk discrimination branch of the fusion prediction model; Based on the behavioral and emotional deviation markers, the short-term fluctuation features, the continuous decay features, the abnormal concentrated segment markers, and the loss stage identification results, a set of candidate causes is generated in the cause identification branch of the fusion prediction model; Based on the hierarchical condition vector, a hierarchical triggering flag is generated, and a cause type flag, a signal segment flag corresponding to the cause candidate flag, and a hierarchical triggering flag are bound to each cause candidate in the cause candidate set to generate a churn cause candidate set.

[0106] In one embodiment, the policy knowledge mapping module 40 is specifically used for: Risk evidence markers are generated based on the churn risk score, stage evidence markers are generated based on the churn stage identification results, and cause type markers, signal fragment markers, and hierarchical trigger markers are read from the churn cause candidate set. Write the risk evidence marker, the stage evidence marker, the cause type marker, the signal fragment marker, and the hierarchical trigger marker into the cause evidence slot set, and generate a cause evidence vector based on the cause evidence slot set; Read the historical cause evidence vector, historical hierarchical status representation, historical retention strategy, historical response status, historical feedback fragment, and case source tag from the historical retention response record; generate a historical response evaluation tag based on the historical response status and the historical feedback fragment; and write the historical cause evidence vector, the historical hierarchical status representation, the historical retention strategy, the historical response evaluation tag, and the case source tag into the historical retention response index. The cause evidence vector is matched with the historical retention response index. A state overlap marker is generated based on the hierarchical state representation and the historical hierarchical state representation. The historical retention response records corresponding to the historical retention response index are filtered based on the state overlap marker and the historical response evaluation marker to generate a candidate retention case set. The candidate retention cases in the candidate retention case set are bound with a case source marker, a historical retention strategy and a historical response evaluation marker. Obtain a preset strategy knowledge graph, and read object type nodes, churn reason nodes, retention resource nodes, execution end nodes, strategy cost nodes, disabling condition nodes, historical response status nodes, case source nodes, and strategy association edge sets from the strategy knowledge graph. Generate a candidate strategy path set based on the object type nodes, churn reason nodes, retention resource nodes, execution end nodes, strategy cost nodes, disabling condition nodes, historical response status nodes, case source nodes, and strategy association edge sets. Based on the hierarchical state representation, a target object type node is determined in the object type node; based on the cause evidence vector, a target churn cause node is determined in the churn cause node; based on the historical retention strategies in the candidate retention case set, a candidate resource execution node group is determined in the retention resource node and the execution end node; based on the candidate resource execution node group, a target strategy cost node is determined in the strategy cost node; based on the candidate resource execution node group, a target disabling condition node is determined in the disabling condition node; based on the historical response evaluation flag and case source flag in the candidate retention case set, a target historical response status node and a target case source node are determined respectively; and a target strategy path set that simultaneously connects the target object type node, the target churn cause node, the retention resource node and execution end node in the candidate resource execution node group, the target strategy cost node, the target disabling condition node, the target historical response status node, and the target case source node is selected from the candidate strategy path set. Extract retention resource markers, execution end markers, policy cost markers, disabling condition markers, historical response status markers, and case source markers from the target policy path set to generate a candidate retention policy set.

[0107] In one embodiment, the retention strategy deduction module 50 is specifically used for: Retention resource markers, execution terminal markers, policy cost markers, disabling condition markers, historical response status markers, and case source markers are read from the candidate retention policy set, and a candidate policy decomposition table is generated based on the retention resource markers, execution terminal markers, policy cost markers, disabling condition markers, historical response status markers, and case source markers; Obtain implementation constraint data, and extract resource inventory markers, execution terminal available time periods, reach frequency records, reach interval markers, object preference channel markers, service personnel availability status, and cost occupancy limits from the implementation constraint data; Based on the retained resource markers and resource inventory markers in the candidate strategy decomposition table, a resource availability marker is generated; based on the execution end markers, execution end availability time periods, and object preference channel markers in the candidate strategy decomposition table, an execution end reachability marker is generated. A cost occupancy marker is generated based on the strategy cost marker in the candidate strategy decomposition table and the cost occupancy limit, and a disable hit marker is generated based on the disable condition marker, reach frequency record, reach interval marker and service personnel availability status in the candidate strategy decomposition table. The resource availability flag, the execution end reachability flag, the cost occupancy flag, and the disabled hit flag are written into the implementation feasibility matrix. An evaluation result is generated based on the implementation feasibility matrix, and an implementable retention strategy set is selected from the candidate retention strategy set based on the evaluation result. Read retention resource tags and execution terminal tags from the set of implementable retention strategies, and generate a set of strategy action types based on the retention resource tags and the execution terminal tags; Based on the historical response status markers corresponding to the set of implementable retention strategies in the candidate strategy decomposition table, the case source markers corresponding to the set of implementable retention strategies in the candidate strategy decomposition table, and the set of strategy action types, a strategy response status chain is generated. Based on the strategy response status chain, the action sequence of the set of strategy action types is arranged to generate a set of strategy action sequences. Acquire historical strategy response data, match the strategy action sequence set with the historical strategy response data to generate response deduction results for each strategy action sequence; Based on the response simulation results, a target retention strategy is determined from the set of feasible retention strategies, and the retention resource marker, execution terminal marker, and strategy action sequence corresponding to the target retention strategy are bound to the target retention strategy.

[0108] In one embodiment, the policy feedback update module 60 is specifically used for: Monitor the trigger status between the churn risk score and the early warning judgment parameter, and generate a trigger judgment flag when the trigger status is reached; Read the execution terminal marker, retention resource marker, and policy action sequence from the target retention policy, and determine the corresponding execution terminal based on the execution terminal marker; Based on the trigger determination flag, the retention resource flag, the strategy action sequence, and the execution end flag, a strategy push payload is generated, and a push content flag and a reach time window are configured for the strategy push payload, and the strategy push payload is pushed to the corresponding execution end; Receive the reach confirmation information returned by the corresponding execution terminal, and generate a strategy push record based on the reach confirmation information, the strategy push payload, and the target retention strategy; Based on the strategy push records and the target retention strategy, feedback data is generated by collecting resource usage feedback, page return feedback, service reach feedback, message response feedback, complaint change feedback, login continuation feedback and retention status feedback. The resource usage feedback in the feedback data is converted into a resource response tag; the page return feedback and login continuation feedback in the feedback data are converted into behavior recovery tags; the service reach feedback and message response feedback in the feedback data are converted into reach response tags; and the complaint change feedback and retention status feedback in the feedback data are converted into status change tags. The resource response flag, the behavior recovery flag, the reach response flag, and the state change flag are written into the policy response features according to the feedback occurrence time; Based on the policy push records, the policy response features are associated with the churn signal representation to generate a response attribution flag, a signal change flag, and a policy status flag. Based on the response attribution flag, the signal change flag, and the policy status flag, a policy execution result is generated. Based on the strategy execution results, model update data and graph update data are generated. Based on the model update data, the branch parameters of the segment time-series branch, risk discrimination branch, stage identification branch, and cause identification branch in the fusion prediction model are updated. Based on the graph update data, the node attributes of the retention resource node, execution end node, historical response status node, case source node, and disabled condition node in the strategy knowledge graph, as well as the association edge attributes of the strategy association edge set, are updated.

[0109] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When executed by the processor, the computer program implements server-side functions or steps of a behavior analysis-driven churn retention method.

[0110] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps of a behavior analysis-driven churn retention method.

[0111] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Acquire interaction record data and object attribute data, and then identify, bind, and integrate the interaction record data and object attribute data in a time sequence to generate a multi-source behavior data set. Based on the multi-source behavioral data set, a behavioral path representation, emotional interaction features, and hierarchical state representation are generated. The behavioral path representation and the emotional interaction features are fused to generate a loss signal representation, and early warning judgment parameters are configured in the hierarchical state representation. The churn signal representation and the hierarchical state representation are input into the fusion prediction model, which outputs a churn risk score, churn stage identification results, and a candidate set of churn reasons. Based on the churn risk score, the churn stage identification result, and the candidate set of churn reasons, a cause evidence vector is constructed. The cause evidence vector is matched with historical retention response records to generate a candidate retention case set. A strategy knowledge graph is obtained. Based on the hierarchical state representation, the cause evidence vector, the candidate retention case set, and the strategy knowledge graph, a candidate retention strategy set is generated. The feasibility of implementing the candidate retention strategy set is assessed, and response deduction is performed based on the assessment results to generate the target retention strategy; When the churn risk score reaches the warning judgment parameter, the target retention strategy is pushed to the corresponding execution end, the feedback data collected for the target retention strategy is converted into strategy response features, the strategy response features are associated with the churn signal representation to generate strategy execution results, and the fusion prediction model and the strategy knowledge graph are updated according to the strategy execution results.

[0112] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, performs the following steps: Acquire interaction record data and object attribute data, and then identify, bind, and integrate the interaction record data and object attribute data in a time sequence to generate a multi-source behavior data set. Based on the multi-source behavioral data set, a behavioral path representation, emotional interaction features, and hierarchical state representation are generated. The behavioral path representation and the emotional interaction features are fused to generate a loss signal representation, and early warning judgment parameters are configured in the hierarchical state representation. The churn signal representation and the hierarchical state representation are input into the fusion prediction model, which outputs a churn risk score, churn stage identification results, and a candidate set of churn reasons. Based on the churn risk score, the churn stage identification result, and the candidate set of churn reasons, a cause evidence vector is constructed. The cause evidence vector is matched with historical retention response records to generate a candidate retention case set. A strategy knowledge graph is obtained. Based on the hierarchical state representation, the cause evidence vector, the candidate retention case set, and the strategy knowledge graph, a candidate retention strategy set is generated. The feasibility of implementing the candidate retention strategy set is assessed, and response deduction is performed based on the assessment results to generate the target retention strategy; When the churn risk score reaches the warning judgment parameter, the target retention strategy is pushed to the corresponding execution end, the feedback data collected for the target retention strategy is converted into strategy response features, the strategy response features are associated with the churn signal representation to generate strategy execution results, and the fusion prediction model and the strategy knowledge graph are updated according to the strategy execution results.

[0113] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0114] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0115] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0116] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0117] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A behavior analysis-driven churn and retention method, characterized in that, Includes the following steps: Acquire interaction record data and object attribute data, and then identify, bind, and integrate the interaction record data and object attribute data in a time sequence to generate a multi-source behavior data set. Based on the multi-source behavioral data set, a behavioral path representation, emotional interaction features, and hierarchical state representation are generated. The behavioral path representation and the emotional interaction features are fused to generate a loss signal representation, and early warning judgment parameters are configured in the hierarchical state representation. The churn signal representation and the hierarchical state representation are input into the fusion prediction model, which outputs a churn risk score, churn stage identification results, and a candidate set of churn reasons. Based on the churn risk score, the churn stage identification result, and the candidate set of churn reasons, a cause evidence vector is constructed. The cause evidence vector is matched with historical retention response records to generate a candidate retention case set. A strategy knowledge graph is obtained. Based on the hierarchical state representation, the cause evidence vector, the candidate retention case set, and the strategy knowledge graph, a candidate retention strategy set is generated. The feasibility of implementing the candidate retention strategy set is assessed, and response deduction is performed based on the assessment results to generate the target retention strategy; When the churn risk score reaches the warning judgment parameter, the target retention strategy is pushed to the corresponding execution end, the feedback data collected for the target retention strategy is converted into strategy response features, the strategy response features are associated with the churn signal representation to generate strategy execution results, and the fusion prediction model and the strategy knowledge graph are updated according to the strategy execution results.

2. The behavior analysis-driven churn and retention method as described in claim 1, characterized in that, Acquire interaction record data and object attribute data, and perform identification binding and time-series integration on the interaction record data and object attribute data to generate a multi-source behavior data set, including: Interaction record data is obtained from different interaction channels, and the interaction entry point, session tag, event occurrence time, business object tag and interaction action name are extracted from the interaction record data to generate an interaction source signature; Extract the object's basic identifier, object's channel attribute, object's grouping attribute, object's service status, and object's usage period from the object attribute data to generate an object attribute profile; Based on the interaction source signature and the object attribute profile, an identifier binding is performed to generate an object candidate identifier cluster, and the candidate identifier, data source, occurrence time and binding trusted marker are recorded in the object candidate identifier cluster; Based on the candidate identifiers, data sources, occurrence times, and binding trust tags in the object candidate identifier cluster, an identifier verification fragment is generated, and the identifier verification fragment is bound to the corresponding interaction record data; The event occurrence time, data reception time, and business completion time are extracted from the interaction record data to generate event time anchors. Based on the event time anchors, the object candidate identifier cluster, and the identifier verification fragment, the interaction record data and the object attribute data are integrated in time sequence to generate an object time sequence fragment set. The interaction source signature, the object attribute profile, the object candidate identifier cluster, the identifier verification fragment, and the object time sequence fragment set are merged. An object binding tag is generated based on the object candidate identifier cluster and the identifier verification fragment. An abnormal fragment tag is generated based on missing time, repeated interaction, abnormal stay, abnormal jump, and text transcription abnormality. The object binding tag and the abnormal fragment tag are written into the merged result to generate a multi-source behavior data set.

3. The behavior analysis-driven churn and retention method as described in claim 1, characterized in that, Based on the multi-source behavioral data set, a behavioral path representation, emotional interaction features, and hierarchical state representation are generated. The behavioral path representation and the emotional interaction features are then fused to generate a churn signal representation. Furthermore, warning judgment parameters are configured in the hierarchical state representation, including: Read object binding tags, interaction action names, business object tags, event occurrence times, page dwell times, function usage times, and repeated trigger records from the multi-source behavior data set, and generate a set of behavior event fragments based on the object binding tags, interaction action names, business object tags, event occurrence times, page dwell times, function usage times, and repeated trigger records; Based on the jump relationship, dwell change, repeated trigger state and path interruption state between adjacent behavior event segments in the behavior event segment set, a path transfer mark set is generated, and the path transfer mark set is written into the behavior event segment set to generate a behavior path representation; The evaluation content, complaint content, customer service dialogue content, and voice semantic analysis results are read from the multi-source behavioral data set, and emotion tendency markers, service dissatisfaction markers, interaction blockage markers, and emotion fluctuation markers are generated based on the evaluation content, complaint content, customer service dialogue content, and voice semantic analysis results. The emotional tendency marker, the service dissatisfaction marker, the interaction blockage marker, and the emotional fluctuation marker are associated with the set of behavioral event fragments according to the corresponding event occurrence time to generate emotional interaction features; Read object base tags, object grouping attributes, object service status, object usage period, historical contribution tags, and recent activity tags from the multi-source behavioral data set, and generate periodic stage tags, value level tags, service status tags, and group attribute tags based on the object base tags, object grouping attributes, object service status, object usage period, historical contribution tags, and recent activity tags; The periodic stage marker, the value level marker, the service status marker, and the group attribute marker are combined to generate a hierarchical status representation, and warning judgment parameters are configured in the hierarchical status representation based on the periodic stage marker, the value level marker, the service status marker, and the group attribute marker. The behavioral path representation and the emotional interaction features are aligned at the fragment level according to the event occurrence time to generate a behavioral-emotional aligned fragment set, and a churn signal representation is generated based on the behavioral-emotional aligned fragment set.

4. The behavior analysis-driven churn and retention method as described in claim 1, characterized in that, The churn signal representation and the hierarchical state representation are input into a fusion prediction model, which outputs a churn risk score, churn stage identification results, and a candidate set of churn causes, including: The lost signal representation and the hierarchical state representation are input into the fusion prediction model, and the signal segment marker, segment time span, signal strength marker and behavioral emotion deviation marker are parsed from the lost signal representation in the fusion prediction model to generate a signal segment sequence. Read the cycle stage marker, value level marker, service status marker, group attribute marker, and early warning judgment parameter from the hierarchical status representation, and generate a hierarchical condition vector and hierarchical calibration parameters; Based on the signal segment sequence, the segment time span, the signal strength marker, and the hierarchical condition vector, short-term fluctuation features, continuous decay features, and abnormally concentrated segment markers are generated in the segment time-series branch of the fusion prediction model. Based on the hierarchical calibration parameters, the short-term fluctuation characteristics, the continuous decay characteristics, and the abnormal concentrated fragment markers, the loss stage identification results are generated in the stage identification branch of the fusion prediction model; Based on the hierarchical calibration parameters, the signal strength marker, the short-term fluctuation characteristics, the continuous attenuation characteristics, and the abnormal concentrated segment marker, a loss risk score is generated in the risk discrimination branch of the fusion prediction model; Based on the behavioral and emotional deviation markers, the short-term fluctuation features, the continuous decay features, the abnormal concentrated segment markers, and the loss stage identification results, a set of candidate causes is generated in the cause identification branch of the fusion prediction model; Based on the hierarchical condition vector, a hierarchical triggering flag is generated, and a cause type flag, a signal segment flag corresponding to the cause candidate flag, and a hierarchical triggering flag are bound to each cause candidate in the cause candidate set to generate a churn cause candidate set.

5. The behavior analysis-driven churn and retention method as described in claim 1, characterized in that, Based on the churn risk score, the churn stage identification results, and the candidate set of churn reasons, a cause evidence vector is constructed. This cause evidence vector is then matched with historical retention response records to generate a candidate retention case set. A strategy knowledge graph is obtained, and a candidate retention strategy set is generated based on the hierarchical state representation, the cause evidence vector, the candidate retention case set, and the strategy knowledge graph, including: Risk evidence markers are generated based on the churn risk score, stage evidence markers are generated based on the churn stage identification results, and cause type markers, signal fragment markers, and hierarchical trigger markers are read from the churn cause candidate set. Write the risk evidence marker, the stage evidence marker, the cause type marker, the signal fragment marker, and the hierarchical trigger marker into the cause evidence slot set, and generate a cause evidence vector based on the cause evidence slot set; Read the historical cause evidence vector, historical hierarchical status representation, historical retention strategy, historical response status, historical feedback fragment, and case source tag from the historical retention response record; generate a historical response evaluation tag based on the historical response status and the historical feedback fragment; and write the historical cause evidence vector, the historical hierarchical status representation, the historical retention strategy, the historical response evaluation tag, and the case source tag into the historical retention response index. The cause evidence vector is matched with the historical retention response index. A state overlap marker is generated based on the hierarchical state representation and the historical hierarchical state representation. The historical retention response records corresponding to the historical retention response index are filtered based on the state overlap marker and the historical response evaluation marker to generate a candidate retention case set. The candidate retention cases in the candidate retention case set are bound with a case source marker, a historical retention strategy and a historical response evaluation marker. Obtain a preset strategy knowledge graph, and read object type nodes, churn reason nodes, retention resource nodes, execution end nodes, strategy cost nodes, disabling condition nodes, historical response status nodes, case source nodes, and strategy association edge sets from the strategy knowledge graph. Generate a candidate strategy path set based on the object type nodes, churn reason nodes, retention resource nodes, execution end nodes, strategy cost nodes, disabling condition nodes, historical response status nodes, case source nodes, and strategy association edge sets. Based on the hierarchical state representation, a target object type node is determined in the object type node; based on the cause evidence vector, a target churn cause node is determined in the churn cause node; based on the historical retention strategies in the candidate retention case set, a candidate resource execution node group is determined in the retention resource node and the execution end node; based on the candidate resource execution node group, a target strategy cost node is determined in the strategy cost node; based on the candidate resource execution node group, a target disabling condition node is determined in the disabling condition node; based on the historical response evaluation flag and case source flag in the candidate retention case set, a target historical response status node and a target case source node are determined respectively; and a target strategy path set that simultaneously connects the target object type node, the target churn cause node, the retention resource node and execution end node in the candidate resource execution node group, the target strategy cost node, the target disabling condition node, the target historical response status node, and the target case source node is selected from the candidate strategy path set. Extract retention resource markers, execution end markers, policy cost markers, disabling condition markers, historical response status markers, and case source markers from the target policy path set to generate a candidate retention policy set.

6. The behavior analysis-driven churn and retention method as described in claim 1, characterized in that, A feasibility assessment is performed on the candidate retention strategy set, and a response simulation is conducted based on the assessment results to generate a target retention strategy, including: Retention resource markers, execution terminal markers, policy cost markers, disabling condition markers, historical response status markers, and case source markers are read from the candidate retention policy set, and a candidate policy decomposition table is generated based on the retention resource markers, execution terminal markers, policy cost markers, disabling condition markers, historical response status markers, and case source markers; Obtain implementation constraint data, and extract resource inventory markers, execution terminal available time periods, reach frequency records, reach interval markers, object preference channel markers, service personnel availability status, and cost occupancy limits from the implementation constraint data; Based on the retained resource markers and resource inventory markers in the candidate strategy decomposition table, a resource availability marker is generated; based on the execution end markers, execution end availability time periods, and object preference channel markers in the candidate strategy decomposition table, an execution end reachability marker is generated. A cost occupancy marker is generated based on the strategy cost marker in the candidate strategy decomposition table and the cost occupancy limit, and a disable hit marker is generated based on the disable condition marker, reach frequency record, reach interval marker and service personnel availability status in the candidate strategy decomposition table. The resource availability flag, the execution end reachability flag, the cost occupancy flag, and the disabled hit flag are written into the implementation feasibility matrix. An evaluation result is generated based on the implementation feasibility matrix, and an implementable retention strategy set is selected from the candidate retention strategy set based on the evaluation result. Read retention resource tags and execution terminal tags from the set of implementable retention strategies, and generate a set of strategy action types based on the retention resource tags and the execution terminal tags; Based on the historical response status markers corresponding to the set of implementable retention strategies in the candidate strategy decomposition table, the case source markers corresponding to the set of implementable retention strategies in the candidate strategy decomposition table, and the set of strategy action types, a strategy response status chain is generated. Based on the strategy response status chain, the action sequence of the set of strategy action types is arranged to generate a set of strategy action sequences. Acquire historical strategy response data, match the strategy action sequence set with the historical strategy response data to generate response deduction results for each strategy action sequence; Based on the response simulation results, a target retention strategy is determined from the set of feasible retention strategies, and the retention resource marker, execution terminal marker, and strategy action sequence corresponding to the target retention strategy are bound to the target retention strategy.

7. The behavior analysis-driven churn and retention method as described in claim 1, characterized in that, When the churn risk score reaches the warning judgment parameter, the target retention strategy is pushed to the corresponding execution end. The feedback data collected for the target retention strategy is converted into strategy response features. The strategy response features are correlated with the churn signal representation to generate a strategy execution result. The fusion prediction model and the strategy knowledge graph are updated based on the strategy execution result, including: Monitor the trigger status between the churn risk score and the early warning judgment parameter, and generate a trigger judgment flag when the trigger status is reached; Read the execution terminal marker, retention resource marker, and policy action sequence from the target retention policy, and determine the corresponding execution terminal based on the execution terminal marker; Based on the trigger determination flag, the retention resource flag, the strategy action sequence, and the execution end flag, a strategy push payload is generated, and a push content flag and a reach time window are configured for the strategy push payload, and the strategy push payload is pushed to the corresponding execution end; Receive the reach confirmation information returned by the corresponding execution terminal, and generate a strategy push record based on the reach confirmation information, the strategy push payload, and the target retention strategy; Based on the strategy push records and the target retention strategy, feedback data is generated by collecting resource usage feedback, page return feedback, service reach feedback, message response feedback, complaint change feedback, login continuation feedback and retention status feedback. The resource usage feedback in the feedback data is converted into a resource response tag; the page return feedback and login continuation feedback in the feedback data are converted into behavior recovery tags; the service reach feedback and message response feedback in the feedback data are converted into reach response tags; and the complaint change feedback and retention status feedback in the feedback data are converted into status change tags. The resource response flag, the behavior recovery flag, the reach response flag, and the state change flag are written into the policy response features according to the feedback occurrence time; Based on the policy push records, the policy response features are associated with the churn signal representation to generate a response attribution flag, a signal change flag, and a policy status flag. Based on the response attribution flag, the signal change flag, and the policy status flag, a policy execution result is generated. Based on the strategy execution results, model update data and graph update data are generated. Based on the model update data, the branch parameters of the segment time-series branch, risk discrimination branch, stage identification branch, and cause identification branch in the fusion prediction model are updated. Based on the graph update data, the node attributes of the retention resource node, execution end node, historical response status node, case source node, and disabled condition node in the strategy knowledge graph, as well as the association edge attributes of the strategy association edge set, are updated.

8. A behavior analysis-driven churn and retention device, characterized in that, The behavior analysis-driven churn retention device includes: The multi-source behavior data integration module is used to acquire interaction record data and object attribute data, and to identify, bind, and integrate the interaction record data and object attribute data in a time sequence to generate a multi-source behavior data set. The churn signal construction module is used to generate a behavior path representation, emotional interaction features and hierarchical state representation based on the multi-source behavior data set, fuse the behavior path representation and the emotional interaction features to generate a churn signal representation, and configure early warning judgment parameters in the hierarchical state representation; The fusion risk prediction module is used to input the churn signal representation and the hierarchical state representation into the fusion prediction model, and output the churn risk score, churn stage identification result and churn reason candidate set; The strategy knowledge mapping module is used to construct a cause evidence vector based on the churn risk score, the churn stage identification result and the churn cause candidate set, match the cause evidence vector with historical retention response records to generate a candidate retention case set, obtain a strategy knowledge graph, and generate a candidate retention strategy set based on the hierarchical state representation, the cause evidence vector, the candidate retention case set and the strategy knowledge graph. The retention strategy deduction module is used to evaluate the feasibility of implementing the candidate retention strategy set, perform response deduction based on the evaluation results, and generate the target retention strategy. The strategy feedback update module is used to push the target retention strategy to the corresponding execution end when the churn risk score reaches the warning judgment parameter, convert the feedback data collected for the target retention strategy into strategy response features, associate the strategy response features with the churn signal representation to generate strategy execution results, and update the fusion prediction model and the strategy knowledge graph according to the strategy execution results.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a behavior analysis-driven churn retention program stored in the memory and executable on the processor, wherein the behavior analysis-driven churn retention program, when executed by the processor, implements the steps of the behavior analysis-driven churn retention method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a behavior analysis-driven churn retention program, which, when executed by a processor, implements the steps of the behavior analysis-driven churn retention method as described in any one of claims 1-7.

Images