An insurance customer loss early warning method and system fusing causal inference large model learning

Abstract

The present application belongs to the technical field of insurance data intelligent research and judgment, and particularly relates to an insurance customer loss early warning method and system fusing a causal inference large model learning. The method comprises: collecting customer behavior data, policy information, claim records and other multi-source data and standardizing processing; constructing a causal graph based on a structural equation model, and extracting a high causal degree path set; constructing a multi-layer recurrent neural network model for supervised training with the path structure as a priori; optimizing the model structure through path stability analysis; and finally combining the prediction probability with the path backtracking result to output the customer risk level and reversible intervention node. The method realizes causal interpretable prediction and operable path intervention of customer loss, and is suitable for intelligent customer management and personalized operation scenarios in the insurance industry.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent analysis of insurance data, specifically involving a method and system for early warning of insurance customer churn that integrates causal inference and large-scale model learning. Background Technology

[0002] Against the backdrop of deepening digital transformation and intelligent services, the insurance industry is facing multiple challenges, including diversified customer needs, refined services, and the maintenance of customer loyalty. Customer churn, as a crucial indicator of corporate stability and service quality, is receiving increasing attention from insurance companies. This is especially true in long-term product areas such as life and health insurance, where customer engagement has a profound impact on revenue structure and risk assessment systems. Therefore, leveraging advanced data analytics and artificial intelligence technologies to proactively identify and effectively intervene in customer churn has become a key direction for insurance technology innovation.

[0003] Existing insurance customer churn prediction methods primarily rely on statistical modeling and machine learning techniques to construct scoring models, predicting churn risk through variables such as historical insurance data, customer behavioral characteristics, and policy fulfillment status. Models such as logistic regression, decision trees, random forests, and XGBoost are widely used, often combined with customer segmentation strategies like RFM and lifecycle models to build churn scoring systems. However, these methods are mostly based on correlation analysis and fail to fully explore the causal relationships between variables, exhibiting limitations in addressing variable collinearity, confounding factors, and evaluating intervention effectiveness. Furthermore, existing models often rely on large amounts of labeled data for supervised training, requiring high data quality, sample balance, and feature engineering capabilities. They are also difficult to transfer to new customer groups or different insurance scenarios, resulting in insufficient model generalization ability and an inability to support personalized early warning and intervention decisions.

[0004] Driven by the dual demands of refined insurance operations and intelligent decision-making, there is an urgent need to introduce technological approaches with greater causal explanatory power and generalization ability to construct customer churn early warning methods applicable to complex and dynamic environments, thereby enhancing the foresight of insurance customer management and the scientific nature of intervention strategies. Summary of the Invention

[0005] To address the aforementioned problems, the present invention aims to propose an insurance customer churn early warning method that integrates causal analysis and sequential behavior determination, comprising the following steps:

[0006] S1. Collection of customer characteristic data and construction of behavioral indicators: Collect multi-source customer behavior data, including login behavior, policy renewal behavior, claims records, payment operations, online interactions, etc., clean them to form a set of structured customer characteristic data, and construct a subset of behavioral indicators including interaction frequency, dwell time, and operation activity.

[0007] S2. Construction of Causal Behavior Paths and Screening of Risk-Driven Paths: Based on customer behavior indicator data, analyze the causal relationship of key variables affecting changes in customer policy status, construct several causal path diagrams, apply behavioral perturbation tests to each path, and screen paths with significant differences before and after changes in behavioral status as a set of high-risk paths.

[0008] S3. Customer churn risk status inference and behavior trend determination: Combining the high-risk path set and the customer's historical behavior trajectory, infer the churn trend of each customer's current behavior status, identify the core influencing nodes in the path, and combine the customer's behavior evolution direction in the renewal period to determine the policy status risk level in the next period.

[0009] S4. Early Warning Level Classification and Behavioral Intervention Point Identification: The customer churn trend assessment results are divided into non-risk, medium-risk, and high-risk levels. For customers in the risk state, key behavioral nodes and turning variables are identified and marked as "interventional path nodes" or "irreversible behavioral nodes". Differentiated early warning suggestions are generated based on rule templates.

[0010] S5. Risk Customer Output and Manual Intervention Suggestion Push: Identified risk level customers are grouped and managed, and each customer's risk level, impact path and node description are output. Intervention suggestion documents for customer service or channels are generated, including renewal communication reminders, personalized marketing reminders or claims processing priority suggestions, etc.

[0011] As a preferred technical solution, the construction of behavioral indicators in S1 includes:

[0012] S11. Extract customer access logs and clickstream records from the past three months by timestamp to construct user behavior sequences;

[0013] S12. Based on the time span and behavior distribution pattern, the behavior sequence is standardized to form a data set with a balanced time step.

[0014] S13. Perform interval judgment and filling on the abnormal behavior segments. The filled data is used as the main input vector. The original abnormal interval is used to construct additional scoring factors.

[0015] S14. Scoring the stability of behavioral activity using a scoring function to determine whether there are gaps or fluctuations in the customer's current behavioral status.

[0016] As a preferred technical solution, the construction of the causal path in S2 includes:

[0017] Using insurance customer behavior variables as nodes, causal edges are filtered based on structural correlation to form a directed behavior path graph;

[0018] The behavioral perturbation test method is used to simulate the frequency changes of variables in the path and calculate the changes in the churn probability distribution before and after the change.

[0019] Using the change in churn probability caused by behavioral disturbances as the criterion for path effectiveness, highly sensitive paths are selected as the core early warning path set.

[0020] As a preferred technical solution, the determination of customer churn risk in S3 includes:

[0021] Based on the current node state of the customer in a high-causality path, combined with the historical evolution trend of the path, predict the possibility of its development into a non-renewal state.

[0022] If a customer's current behavior is more similar to the behavior of a typical churned customer in the warning path than a set threshold, it is presumed that there is a potential risk of churn.

[0023] As a preferred technical solution, the risk level determination rules in S4 include:

[0024] If the predicted churn probability is greater than the preset second threshold, the customer is marked as a high-risk customer.

[0025] If the prediction result is between the first threshold and the second threshold, it is marked as a medium-risk customer;

[0026] Otherwise, mark them as non-risk customers;

[0027] The threshold can be automatically calibrated based on historical data or set by operation and management rules.

[0028] As a preferred technical solution, the calibration of behavioral intervention points in S4 includes:

[0029] Based on the comparison between the customer's current behavior nodes and the early warning path structure, path nodes with the potential for behavior reversal are identified and marked as interventionable nodes;

[0030] If a certain behavior sequence is interrupted and the customer has no further active signals or the access path is completely consistent with that of a risky customer, it is marked as an irreversible node.

[0031] The two types of nodes are output separately for personalized intervention recommendations and risk closed-loop management.

[0032] As a preferred technical solution, the manual intervention suggestion push in S5 includes:

[0033] Generate customer-level risk profiles, including customer identifiers, behavioral tags, risk levels, and core path node information;

[0034] Based on the customer's historical renewal status and current behavior, corresponding intervention suggestion texts are generated for each type of customer;

[0035] The list of high-risk customers is synchronized to the sales, customer service, and agent channel systems to trigger dedicated communication or task reminder processes.

[0036] This invention also provides an insurance customer churn early warning system that integrates a causal inference learning mechanism, used to implement the method, characterized in that it includes:

[0037] Customer data acquisition module: used to synchronize customer behavior data from multiple sources such as customer management system (CRM), policy management system, and online platform;

[0038] Causal Path Analysis Module: Used to build a path structure diagram based on behavioral variables and calculate causal scores;

[0039] Risk Status Reasoning Module: Used to predict a customer's future policy renewal status and classify risk levels;

[0040] Node identification and intervention suggestion module: used to generate key behavioral nodes and corresponding early warning suggestion information;

[0041] Risk output and task linkage module: Export high-risk customer data to external intervention systems or manual intervention platforms.

[0042] Beneficial effects:

[0043] This invention introduces structural equation modeling and path scoring mechanisms to achieve causal structure identification and inference path control, which are difficult to obtain in traditional churn prediction. It can structurally model the potential causal relationships between high-dimensional heterogeneous customer characteristics, clarifying the transmission logic chain between behavioral variables. Compared to shallow learning models based solely on correlation, this method can not only predict whether customers will churn, but also output a quantitative explanation of "why they churn" based on a graph, making the prediction results highly traceable and providing a clear path basis for subsequent business interventions.

[0044] The deep reasoning model constructed in this invention, which integrates high-causal paths and multi-layer recurrent neural networks, possesses dual structural characteristics: on the one hand, it retains causal paths as semantic priors to define the structural boundaries of information propagation; on the other hand, it dynamically models behavioral transitions over time through recursive learning. This structure significantly enhances the model's ability to perceive behavioral disturbances and capture long-term fluctuation trends, achieving model stability control based on path fluctuation analysis, thereby ensuring robust reasoning performance of the system under non-static scenarios such as business cycle fluctuations and user behavior evolution.

[0045] This invention establishes a closed-loop system from prediction to intervention by identifying "reversible causal disturbance nodes." This allows the system to not only determine whether a customer is at risk of churn, but also to clearly identify which behavioral nodes remain adjustable. Furthermore, a path backtracking mechanism outputs the set of variables with the greatest intervention potential. This mechanism provides customer operations teams with targeted and actionable trigger strategies, enabling a shift from a "static classification" to a "dynamic reversal" paradigm in churn risk management. It possesses high practical value and system integration advantages in insurance customer lifecycle operations. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0047] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0048] Example 1

[0049] like Figure 1 As shown, this embodiment proposes an insurance customer churn early warning method that integrates causal inference large model learning. Based on multi-source heterogeneous customer data, structured causal inference graph and deep recursive learning model, an intelligent insurance decision support process oriented towards prediction, analysis and intervention is established.

[0050] The method described is applicable to the integrated deployment of customer service systems, marketing strategy systems, and risk control systems in the insurance industry, and is scalable and platform compatible.

[0051] This method includes the following main steps:

[0052] S1. Collection and Standardization Processing of Customer Feature Data:

[0053] This step aims to construct a customer feature dataset with a unified structure and consistent dimensions, serving as the input foundation for the model. The data acquisition module integrates access to multiple business systems, and the main sources of the collected data are as follows:

[0054] Customer behavior data is derived from client interaction logs, including access records from PCs and mobile devices, clickstream logs, page navigation behavior, dwell time, mouse paths, and form completion information. These logs are recorded by the Web Behavior Analytics SDK and uploaded to a log database for daily batch processing.

[0055] Historical policy information comes from the core policy system, including fields such as each customer's policy number, product type, coverage amount, payment frequency, policy start and end dates, whether it is a renewal, and cumulative payment duration.

[0056] Claim records are sourced from the claims subsystem, including application time, payout amount, payout status (under review, rejected, closed), number of claims, classification of reasons for compensation, and whether it is a critical illness.

[0057] Online interaction data comes from the customer service center and user feedback system, including text chat logs, telephone service tags, satisfaction ratings, complaint types, and customer-initiated feedback.

[0058] Marketing response behavior data comes from the marketing campaign system: such as whether customers clicked on SMS links, opened marketing emails, participated in promotional activities, or claimed coupons.

[0059] After data collection is complete, follow the standardized procedure below:

[0060] Time alignment processing: unifies the features of different time zones and time granularities into a time series in days;

[0061] Missing value handling: For time series data, a sliding window average interpolation is used; for non-time data, the mean of the same type of customer is used for filling.

[0062] Feature normalization: Min-Max normalization is used to compress numerical features into 0-1 values, and one-hot encoding is used for categorical variables.

[0063] Behavioral segmentation: Customer behavior throughout the entire lifecycle is divided into three levels: the past 7 days, the past 30 days, and the past 90 days, and high-frequency, medium-term, and long-term behavioral trend features are extracted respectively.

[0064] Anomaly detection: An algorithm based on the Local Outlier Factor (LOF) is introduced to perform quality verification on the feature data and remove obviously erroneous records.

[0065] Ultimately, a unified set of customer feature vectors is obtained, which serves as the basic input for subsequent causal modeling and prediction models.

[0066] S2. Causal Graph Construction and High Causality Path Selection:

[0067] After standardizing customer feature data, the process moves on to constructing a causal graph. This is the first core innovation of this method, aiming to extract paths with clear causal relationships from a large number of feature variables and use them as prior knowledge for subsequent inference models.

[0068] Structural Equation Modeling (SEM): This step uses structural equation modeling as the basis for causal inference, constructing a linear equation of the following form for each pair of variables:

[0069]

[0070] Where y is the dependent variable, x is the independent variable, β is the dependency coefficient, and ε is the error term. The parameter matrix of the model is solved using the maximum likelihood estimation method.

[0071] Construct a causal graph structure: Treat all variables as nodes in the graph, and retain edges with dependency coefficients greater than a threshold (such as 0.3) to form a preliminary directed graph. The edges in the graph represent the causal directionality between variables.

[0072] Calculate path score: Construct a multi-hop path from the cause variable to the target variable (e.g., "whether to renew insurance"), and calculate the path score for each path:

[0073] Dependency score (weighted average of edge coefficients in the path);

[0074] Relevance score (cooperation variance between path nodes);

[0075] Path entropy (reflects the amount of information in a path);

[0076] Path compression (the ratio of non-redundant nodes in a path).

[0077] Path filtering and pruning: Paths with scores below a set threshold (e.g., path entropy < 0.5, dependency score < 0.4) are deleted, while paths with high structural redundancy are merged.

[0078] Causal intervention simulation: Bayesian networks are used to conduct intervention tests on each path, simulating the distribution of customer behavior before and after the change of path node state, and KL divergence is used to evaluate the degree of impact of path intervention.

[0079] Finally, the set of paths that rank in the top 10% and have significant intervention effects (KL divergence) are selected as the "high causal path set" to be input into the next step of the modeling process.

[0080] S3. Recursive Structure Learning and Customer Intent Reasoning Modeling:

[0081] This step, based on an RNN-like deep learning structure, jointly models customer behavior time series and high-causality path sets to achieve recursive reasoning of customer behavior states towards future churn probabilities.

[0082] Input data construction: customer behavior time series tensor, path structure matrix, label vector;

[0083] Model architecture design:

[0084] The first layer is the path structure encoding layer, which uses a graph embedding algorithm to map the path matrix into a vector representation;

[0085] The second layer is the BiGRU time series processing layer, which performs forward and backward analysis on the customer behavior sequence.

[0086] The third layer is the fusion layer, which integrates the path vector and the behavior vector with attention weighting.

[0087] The fourth layer is a fully connected sigmoid output layer that outputs the dropout probability.

[0088] Training details:

[0089] Loss function: Binary cross-entropy is used;

[0090] Optimizer: Adam, initial learning rate 0.001;

[0091] Training rounds: 30 rounds, with each round having a Mini-Batch size of 64;

[0092] Overfitting prevention strategy: Dropout=0.2, EarlyStopping tolerance step count is 5;

[0093] Model evaluation metrics: AUC, F1-Score, Precision@Top5.

[0094] This model can not only output the churn probability of each customer, but also trace the high causal path on which its prediction results depend, thus possessing good interpretability.

[0095] S4. Path stability analysis and model optimization:

[0096] This step aims to improve the inference model's adaptability to customer behavior drift by identifying unstable paths and performing targeted training optimization through multi-time period validation and path disturbance response analysis.

[0097] Customer grouping: The sample customers are divided into multiple 30-day sliding windows according to time, and each group of data is used as a stability assessment unit;

[0098] In-window inference simulation: Perform a prediction independently within each time window and record the average participation frequency, prediction accuracy, and output probability volatility for each path;

[0099] Constructing a fluctuation score matrix: Construct a matrix for the performance of each path under different time windows, and calculate its mean square deviation as the path fluctuation score;

[0100] Screening unstable paths: If the mean squared deviation of a path exceeds twice the mean score of the overall path, the path is marked as an unstable path and added to the training reinforcement list.

[0101] Path weight reallocation: The sampling weights of each path sample in the training dataset are readjusted, maintaining the default sampling probability of 1.0 for stable paths and increasing the sampling weight of unstable paths by 1.5 times;

[0102] Optimize model retraining: Re-execute 10 rounds of iterative training to improve the overall robustness of the model under the optimized path sample distribution.

[0103] The final result is a set of deep inference models with higher stability, which have time transfer capabilities and cross-cycle generalization capabilities.

[0104] S5. Customer Risk Level Assessment and Reversible Causal Path Output:

[0105] After the model is deployed, the system can make real-time predictions for any one or more target customers and return the risk level and reversible intervention points. The specific process is as follows:

[0106] Predictive Invocation: Input the target customer's behavioral data and static attributes over the past 90 days into the model inference interface to obtain the output probability P;

[0107] Risk classification assessment:

[0108] If P ≥ 0.75, the customer is marked as "high risk";

[0109] If 0.5 ≤ P < 0.75, then it is considered "medium risk";

[0110] If P < 0.5, it is considered "low risk".

[0111] Path contribution tracking: By backtracking the attention mechanism and path weights in the model, the path structure that contributes most to the current prediction (Top 3 paths) is determined.

[0112] Determining the reversibility of causal nodes:

[0113] Check if there are any variable nodes in the path that can be affected by marketing, service, policy changes, etc. (such as "number of customer service interactions in the past 30 days", "frequency of coupon usage", etc.).

[0114] If these nodes have experienced significant fluctuations in the past and are currently in an "adjustable state," they are marked as "reversible intervention nodes."

[0115] Intervention suggestion generation: Combining path analysis and node behavior values, specific intervention suggestions are automatically generated, such as "proactively initiate policy renewal communication", "send customized discount reminders", and "push successful claims cases" to improve customer retention probability.

[0116] Output structure: Customer ID; Current risk level; Predicted probability value; Main causal path name (variable sequence); List of reversible intervention nodes; Intervention strategy suggestion text.

[0117] The above outputs can be automatically synchronized to the CRM system, allowing customer operations specialists or marketing teams to implement differentiated relationship maintenance.

[0118] Example 2

[0119] This embodiment provides an insurance customer churn early warning device that integrates causal inference large model learning. It is applicable to business scenarios such as insurance customer lifecycle management, renewal strategy formulation, and intelligent customer service operation. It is particularly suitable for deployment in the intelligent operation platform, data middle platform, or private cloud environment of insurance companies.

[0120] This device is designed for real-time processing of large-scale customer behavior data, causal structure learning, and high-dimensional reasoning tasks. It combines causal graph modeling and deep learning models, which can effectively improve the accuracy of customer churn prediction and the targeting of intervention strategies.

[0121] The device includes the following core modules:

[0122] 1. Customer Data Acquisition and Preprocessing Module:

[0123] This module is used to collect multi-source customer data and perform format standardization, missing data filling, and feature standardization processing. It mainly includes the following sub-components:

[0124] Data access submodule: It uses Kafka data bus to connect with the policy data table, claims database, interaction log system and marketing management platform of the insurance business system; it supports structured data (SQL tables), semi-structured data (JSON, XML) and unstructured data (customer service voice text) input; it is compatible with scheduled batch retrieval and real-time streaming synchronization, and the data synchronization cycle is configured to be 10 minutes.

[0125] Feature standardization submodule: performs unified field mapping on data from different sources, including field naming standardization, unit conversion (e.g., converting ten thousand yuan to yuan), and type forced conversion; uses the Pandas data framework to complete segmented normalization processing, such as Z-score standardization of click count, login frequency, etc. according to behavioral distribution; null value handling strategies include: forward filling of the most recent time point, filling with the mean of similar customers, and interpolation completion (e.g., behavioral time series).

[0126] Customer behavior vector generation submodule: Constructs fixed-length feature vectors (e.g., 300-dimensional) containing behavioral feature summaries for the last 7, 30, and 90 days; supports generating window sliding behavior fragments on a user-by-user basis, which can then be input into subsequent recursive models.

[0127] This module outputs a standardized customer feature dataset, providing a unified data foundation for the causal graph construction module.

[0128] 2. Causal Graph Construction and Path Selection Module:

[0129] This module is used to build structural equation models based on customer characteristic data, form a causal relationship graph between variables, and extract a set of high causal paths, mainly including:

[0130] The causal relationship identification submodule adopts structural equation modeling (SEM) as the causal learning framework; it sequentially regresses all latent independent variables for each target variable (such as "whether it is lost"), calculates the path dependence coefficient and the fitting residual; and uses maximum likelihood estimation to solve for the strength of directed edges between nodes.

[0131] The causal graph structure generation submodule treats all variables as nodes and retains edges with causal strength greater than a set threshold (e.g., 0.3) to form a preliminary causal network; the NetworkX graphics computing library is used to store the directed graph structure.

[0132] The path search and combination submodule: Based on the DFS algorithm, it searches for all causal paths from "key upstream variables" to "whether customers are churned"; for each path, it calculates indicators such as path length, cumulative causal strength, and node information entropy.

[0133] The path scoring and screening submodule performs a simulated intervention experiment for each path and calculates the distribution of customer states before and after the intervention based on Bayesian networks. KL divergence is used as the basis for path scoring, with higher scores indicating a greater causal intervention impact. Paths with scores in the top 10% and path difference not equal to 0 are selected to form a set of high causal paths.

[0134] This module outputs a path structure matrix and a score vector, which serve as structural priors for model training input.

[0135] 3. Inference Model Construction and Training Module:

[0136] This module is used to construct a recurrent neural network structure based on high-causal paths as priors, and to perform time-series learning of customer status and churn intention prediction, including:

[0137] The path embedding encoding submodule represents each causal path as a sequence of nodes and uses a Node2Vec or graph neural network encoder to embed the path structure into a low-dimensional dense vector (e.g., 64-dimensional); all paths constitute a path vector set.

[0138] The behavior sequence recursive processing submodule: Customer behavior vectors are arranged into a sequence according to time windows and input into the BiGRU structure; the hidden state sequence H of each time step is output, representing the potential state evolution of the customer at each time node.

[0139] The path behavior fusion and attention mechanism submodule utilizes a weighted attention mechanism between the path vector and the sequence state to calculate the interpretive contribution of the path to the current behavior state; the fused state representation is then input to the fully connected layer.

[0140] The classification output and optimization submodule outputs the probability value of whether a customer will churn in the next cycle (e.g., 30 days) through the Sigmoid activation function; it uses cross-entropy as the loss function and adopts the Adam optimizer for training; it records AUC, F1, and accuracy in each training round for model tuning and evaluation.

[0141] This module supports model saving, loading, fine-tuning, and batch training, and can be deployed independently on GPU-accelerated nodes or AI model services.

[0142] 4. Path stability analysis and model iteration module:

[0143] This module analyzes the contribution of paths in the model to the stability of the results, eliminates highly volatile paths, and enhances the model's generalization ability. Its structure includes:

[0144] Time window grouping submodule: Divides the training data into multiple time periods according to the customer's active time, with a typical window length of 30 days; each window serves as an independent evaluation unit.

[0145] Model Inference Simulation Submodule: Runs model inference independently in each window, records the deviation between predicted values ​​and actual results, and marks the participation of each path in inference.

[0146] The path fluctuation scoring submodule: statistically analyzes the performance of each path in inference across all time windows; calculates its standard deviation, variance, and mean error, and constructs the path fluctuation matrix.

[0147] The path weight adjustment submodule sets high training weights for paths with fluctuating scores significantly higher than the average to enhance the model's adaptability to them; and sets low sampling rates for paths with stable scores but low information contribution to reduce interference.

[0148] Retraining control submodule: Retrains the model based on the adjusted weights; supports two strategies: local path freezing or full path retraining.

[0149] This module outputs a set of updated and optimized model structures, further improving prediction accuracy and path reliability.

[0150] 5. Risk Prediction and Reversible Path Output Module:

[0151] This module is business-oriented, providing a real-time prediction interface and interpretable path output, facilitating manual or automated execution of customer retention strategies. It includes:

[0152] Customer input receiving submodule: Receives customer identifiers and behavioral data from the past 90 days transmitted from the front end; parses them into a unified format input vector.

[0153] Risk level determination submodule: Based on the probability value output by the model, two thresholds are set (such as 0.5 and 0.75): greater than 0.75 is high risk; between the two is medium risk; less than 0.5 is low risk.

[0154] The path interpretation and reversible node identification submodule: backtracks the attention mechanism results and extracts the Top-N key causal paths; marks the node variables in the path that can be manually or systematically intervened, such as interaction frequency and claim status; determines whether these nodes have historical evidence of changing their status, and if so, marks them as reversible perturbation nodes.

[0155] The intervention suggestion generation submodule generates text-based operational suggestions based on model analysis results and node values, such as: "It is recommended to stimulate the willingness to renew the policy through coupons"; "It is recommended to arrange customer service representatives to proactively contact and communicate about claims disputes"; "It is recommended to add the customer to the short-term recall marketing list".

[0156] The Results Display and Output Interface submodule supports front-end display of risk scoring cards, path maps, and intervention suggestions; and outputs results to customer relationship management systems (CRM) or business operation dashboards via REST API.

[0157] This device achieves a closed-loop process from data acquisition, causal relationship modeling, deep reasoning to risk output and path interpretation through multi-module collaboration. Compared to traditional prediction methods based on shallow models, this device has the following advantages:

[0158] Structural interpretability: Introducing a causal graph structure improves model transparency;

[0159] Results are traceable: outputs critical path and node variables to assist business intervention;

[0160] High performance: It adopts BiGRU and attention mechanism to achieve compressed learning of high-dimensional sequences;

[0161] Flexible deployment: Supports deployment on private clouds, local servers, and model service platforms.

[0162] This embodiment can be widely applied to insurance technology scenarios such as customer retention analysis, personalized operations, risk stratification, and product pricing.

[0163] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for early warning of insurance customer churn that integrates causal analysis and sequential behavior determination, characterized in that, Includes the following steps: S1. Collection of customer characteristic data and construction of behavioral indicators: Collect multi-source customer behavior data, including login behavior, policy renewal behavior, claims records, payment operations, and online interactions. After cleaning, form a structured customer feature data set and construct a subset of behavioral indicators, including interaction frequency, dwell time, and operation activity. S2. Construction of causal behavioral pathways and screening of risk-driven pathways: Based on customer behavior indicator data, we analyze the causal relationships of key variables that affect changes in customer policy status, construct several causal path diagrams, apply behavioral perturbation tests to each path, and select paths with significant differences before and after changes in behavioral status as a set of high-risk paths. S3. Customer Churn Risk Status Reasoning and Behavioral Trend Determination: By combining a set of high-risk paths with customers' historical behavior trajectories, we can infer the churn trend of each customer's current behavior status, identify the core influencing nodes in the path, and determine the risk level of the policy status in the next cycle by combining the customer's behavioral evolution direction during the renewal period. Step S3 is based on an RNN-like deep learning structure to jointly model customer behavior time series and high-causality path sets, realizing recursive reasoning from customer behavior state to future churn probability; input data construction: customer behavior time series tensor, path structure matrix, label vector; output churn probability for each customer; S4. Early Warning Level Classification and Behavioral Intervention Point Marking: The customer churn trend assessment results are divided into non-risk, medium-risk and high-risk levels. For customers in the risk state, key behavioral nodes and turning variables are identified and labeled as "interventional path nodes" or "irreversible behavioral nodes". Differentiated early warning suggestions are generated based on rule templates. The identification of behavioral intervention points includes: Based on the comparison between the customer's current behavior nodes and the early warning path structure, path nodes with the potential for behavior reversal are identified and marked as interventionable nodes; If a certain behavior sequence is interrupted and the customer has no further active signals or the access path is completely consistent with that of a risky customer, it is marked as an irreversible node. The two types of nodes are output separately for personalized intervention recommendations and risk closed-loop management; S5. Risk customer output and manual intervention suggestion push: The identified risk-level customers are grouped and managed, and each customer's risk level, impact path and node description are output. Intervention suggestion documents for customer service or channels are generated, including renewal communication reminders, personalized marketing reminders or claims processing priority suggestions.

2. The method according to claim 1, characterized in that, The construction of behavioral indicators in S1 includes: S11. Extract customer access logs and clickstream records from the past three months by timestamp to construct user behavior sequences; S12. Based on the time span and behavior distribution pattern, the behavior sequence is standardized to form a data set with a balanced time step. S13. Perform interval judgment and filling on the abnormal behavior segments. The filled data is used as the main input vector. The original abnormal interval is used to construct additional scoring factors. S14. Scoring the stability of behavioral activity using a scoring function to determine whether there are gaps or fluctuations in the customer's current behavioral status.

3. The method according to claim 1, characterized in that, The construction of the causal path in S2 includes: Using insurance customer behavior variables as nodes, causal edges are filtered based on structural correlation to form a directed behavior path graph; The behavioral perturbation test method is used to simulate the frequency changes of variables in the path and calculate the changes in the churn probability distribution before and after the change. Using the change in churn probability caused by behavioral disturbances as the criterion for path effectiveness, highly sensitive paths are selected as the core early warning path set.

4. The method according to claim 1, characterized in that, The determination of customer churn risk in S3 includes: Based on the current node state of the customer in a high-causality path, combined with the historical evolution trend of the path, predict the possibility of its development into a non-renewal state. If a customer's current behavior is more similar to the behavior of a typical churned customer in the warning path than a set threshold, it is presumed that there is a potential risk of churn.

5. The method according to claim 1, characterized in that, The risk level determination rules in S4 include: If the predicted churn probability is greater than the preset second threshold, the customer is marked as a high-risk customer. If the prediction result is between the first threshold and the second threshold, it is marked as a medium-risk customer; Otherwise, mark them as non-risk customers; The threshold can be automatically calibrated based on historical data or set by operation and management rules.

6. The method according to claim 1, characterized in that, The manual intervention suggestion push in S5 includes: generating a customer-level risk profile, which includes customer identification, behavioral tags, risk level and core path node information.

7. The method according to claim 6, characterized in that, The manual intervention suggestion push in S5 also includes: Based on customers' historical renewal status and current behavior, corresponding intervention suggestions are generated for each type of customer; the list of high-risk customers is synchronized to the sales, customer service, and agent channel systems to trigger dedicated communication or task reminder processes.

8. An insurance customer churn early warning system integrating a causal inference learning mechanism, used to implement the method described in any one of claims 1-7, characterized in that, include: Customer data collection module: used to synchronize customer behavior data from the customer management system, policy management system, and online platform; Causal Path Analysis Module: Used to build a path structure diagram based on behavioral variables and calculate causal scores; Risk Status Reasoning Module: Used to predict a customer's future policy renewal status and classify risk levels; Node identification and intervention suggestion module: used to identify key behavioral nodes and generate corresponding early warning suggestion information; Risk output and task linkage module: Export high-risk customer data to external intervention systems or manual intervention platforms.

Images