Insurance customer renewal willingness prediction and management method, system and medium

Abstract

The invention relates to an insurance customer renewal willingness prediction and management method and system and a medium, and belongs to the technical field of big data analysis, and the method comprises the steps: collecting the historical interaction data, payment behavior data and recent activity data of a customer; generating a dynamic feature vector through feature extraction and conversion; inputting the dynamic feature vector into a pre-trained machine learning model, and outputting a customer renewal willingness probability value; performing risk grade division according to the customer renewal willingness probability value to generate a risk grade label; matching a preset reach strategy based on the risk level label, and generating a task instruction including a reach channel and an incentive measure; driving a corresponding communication interface to execute a touch operation according to the task instruction; and obtaining result feedback data of the touch operation, adding the result feedback data as a new label into the training data set, and updating the machine learning model. According to the method, potential lost customers can be accurately identified, and differential intervention measures can be taken in time.

Description

Technical Field

[0001] This application relates to the field of big data analytics, and in particular to a method, system, and medium for predicting and managing insurance customers’ renewal intentions. Background Technology

[0002] With the development of internet technology and the advancement of digital transformation, the insurance industry is gradually shifting from a traditional sales-oriented model to a customer-centric service-driven model. Predicting and managing customer renewal behavior has become a crucial aspect of measuring an insurance company's operational efficiency and service quality. In current business practices, insurance institutions generally face problems such as high customer churn rates and difficulties in converting policies into renewals. Especially in an increasingly competitive existing market environment, accurately identifying potential churned customers and taking timely, differentiated intervention measures has become key to improving customer retention rates and profitability.

[0003] However, in existing insurance customer management mechanisms, most insurance companies still rely on manual experience or simple rule engines to judge customers' renewal intentions. This makes it difficult to fully reflect the actual changing trends of customers' status, resulting in serious waste of outreach resources and poor conversion rates. Currently, some existing customer churn warning systems built on statistical models or simple classification algorithms often lack the ability for continuous iterative optimization. Due to the diverse sources, complex structure, and time-sensitive nature of customer data, existing systems often cannot effectively integrate multi-dimensional heterogeneous information, resulting in poor model generalization ability and low prediction accuracy, which in turn affects the scientific nature and effectiveness of subsequent strategy formulation. Summary of the Invention

[0004] To address the aforementioned technical issues, this application provides a method, system, and medium for predicting and managing insurance customers' renewal intentions.

[0005] Firstly, this application provides a method for predicting and managing insurance customers' renewal intentions, employing the following technical solution: A method for predicting and managing insurance customers' renewal intentions, the method comprising: Collect customers' historical interaction data, payment behavior data, and recent activity data; The historical interaction data, payment behavior data, and recent activity data are subjected to feature extraction and transformation to generate dynamic feature vectors; The dynamic feature vector is input into a pre-trained machine learning model, which outputs a probability value of the customer's willingness to renew. Risk levels are classified based on the probability values ​​of customer renewal intentions, and risk level labels are generated. Based on the risk level labels, a preset outreach strategy is matched to generate a task instruction that includes outreach channels and incentive measures; The corresponding communication interface is driven to perform a reach operation according to the task instruction; Obtain the result feedback data of the reach operation, add the result feedback data as a new label to the training dataset, and update the machine learning model.

[0006] By adopting the aforementioned technical solutions, refined governance practices are implemented around the entire customer lifecycle behavior trajectory, encompassing an integrated collaborative operation mechanism from the underlying perception layer to the top-level cognition layer and then to the feedback control layer. By integrating multiple key technologies such as big data collection, intelligent feature construction, accurate probability inference, dynamic risk classification, customized intervention scheduling, and autonomous evolution and updating, this not only significantly improves the level of refined operation and management of insurance institutions in the competitive landscape of the existing market, but also enhances the company's ability to withstand market fluctuations.

[0007] Optionally, the step of extracting and transforming features from the historical interaction data, payment behavior data, and recent activity data to generate dynamic feature vectors includes: Extract customer service call frequency and online login behavior from the historical interaction data, perform frequency statistics calculation according to a preset time window, and generate behavior frequency sequence; Extract the payment channel types from the payment behavior data and classify and aggregate them to generate a payment channel distribution vector; Extract page operation records from the recent activity data, compare the single dwell time of the relevant page with a threshold, and generate a high attention tag; The frequency sequence of behaviors, the distribution vector of payment channels, and the high attention markers are aligned and concatenated to form a dynamic feature vector.

[0008] By adopting the above technical solutions, and focusing on the goal of predicting insurance customers' renewal intentions, a feature representation framework that combines stability and flexibility is established through the effective extraction, in-depth processing, and organic integration of diverse and heterogeneous data. This framework not only fully explores the deep-seated correlation patterns among customers in terms of payment paths, communication behaviors, and browsing habits, but also enhances the model's robustness to external disturbances and its predictability of future trends by using multi-level time window divisions and multi-granularity evaluation criteria.

[0009] Optionally, the step of inputting the dynamic feature vector into a pre-trained machine learning model and outputting the probability value of customer renewal intention includes: Load a pre-trained machine learning model; wherein the machine learning model is generated through training using historical customer feature data and renewal result labels; The dynamic feature vector is input into the machine learning model to perform classification probability prediction calculation; Output the probability value corresponding to the customer's willingness to renew.

[0010] By adopting the above technical solutions, a closed-loop customer renewal intention prediction and risk assessment process was constructed. This process not only makes full use of the potential patterns in historical data, but also fully considers the dynamic changes in customer status. Through the enhanced interpretability of model output and the effective integration of business rules, the intelligence level and operational efficiency of insurance companies in customer lifecycle management are improved.

[0011] Optionally, the step of classifying risk levels and generating risk level labels based on the probability value of customer renewal intention includes: Probability of receiving customer renewal intentions; Load preset grading threshold parameters, the parameters including a first threshold and a second threshold; The customer's renewal intention probability value is compared with the first threshold and the second threshold to obtain the comparison result; Output a risk level label based on the comparison results; The risk level labels are associated and stored in the customer profile database.

[0012] By adopting the above technical solution, relying on scientifically sound probabilistic modeling results, and utilizing refined threshold control methods, the effective identification and classification management of renewal intention levels among a large number of customers has been achieved. This solution fully considers the complexity and diversity of real-world application scenarios, maintaining high versatility while also possessing strong scalability and intelligent potential, which is of great significance for promoting the digital transformation of the modern insurance industry.

[0013] Optionally, the step of generating a task instruction containing reach channels and incentives based on the risk level label and a preset reach strategy includes: Receive customer risk level labels, including three categories: low risk, medium risk, or high risk; Query the preset strategy rule base and match the reach strategy template corresponding to the risk level label; A structured task instruction is generated based on the reach strategy template, including the reach channel type and incentive identifier; The structured task instructions are written into the task queue database and associated with the customer's unique identification code.

[0014] By adopting the above technical solution and combining it with a strategy rule base to complete strategy matching, structured task instructions are generated and finally written into the task queue database, thus realizing a complete closed loop from customer risk identification to the execution of differentiated outreach strategies. This solution not only improves the efficiency of intervention in customer renewal intentions but also enhances the system's intelligence level and the efficiency of operational resource utilization through structured task management and differentiated strategy configuration.

[0015] Optionally, the step of driving the corresponding communication interface to perform the reach operation according to the task instruction includes: Read structured task instructions from the task queue database, the instructions containing a unique customer identifier, a reach channel type, and an incentive identifier; The corresponding communication interface protocol is invoked based on the type of the reach channel. Send an instruction execution request to the target communication interface, the instruction execution request carrying customer contact information and incentive identifier; Receive the operation status code returned by the target communication interface and update the task execution status log.

[0016] By adopting the above technical solutions, relying on a sophisticated task orchestration mechanism, and integrating diverse communication methods and personalized incentive combinations, the system maintains good robustness and scalability even under the pressure of large-scale concurrent access. Especially in the face of sudden network fluctuations or hardware anomalies, the multi-layered automatic degradation strategy effectively avoids service interruptions caused by the failure of a single node. At the same time, the introduction of a sophisticated operation log management system makes the entire customer interaction process transparent and controllable, enhancing the enterprise's operational efficiency and risk management level.

[0017] Optionally, the steps of obtaining the result feedback data of the outreach operation, adding the result feedback data as a new label to the training dataset, and updating the machine learning model include: Collect the operation status code returned after the task is executed, and the operation status code includes a renewal operation result identifier; Map the renewal operation result identifier to a binary classification label; The binary classification labels are associated with the dynamic feature vectors of the corresponding customers and stored in the training dataset; The pre-trained machine learning model is incrementally trained using the updated training dataset to generate a new version of the model; Load the new version of the model to overwrite the old model.

[0018] By adopting the above technical solution, a closed-loop control system was achieved from the business end to the algorithm end, enhancing the system's autonomous learning capability and intelligent adjustment level. Through refined state filtering and label definition, the quality and representativeness of training samples were ensured. Combined with dynamic feature management and incremental learning strategies, model maintenance costs were significantly reduced, and the iteration pace was accelerated. Simultaneously, a gradual upgrade path was adopted at the deployment level to ensure that online service quality remained unaffected. This technical solution not only enhances the predictive accuracy and intervention effectiveness of insurance companies in customer retention but also provides a valuable reference paradigm for other similar customer relationship management systems.

[0019] Optionally, before the step of inputting the dynamic feature vector into the pre-trained machine learning model, the method further includes: When the training dataset is detected to be empty, static snapshot data from the historical policy database is retrieved; Manual annotation is performed on the static snapshot data to generate an initial training label set; The initial machine learning model is trained based on the XGBoost basic parameters; The initial machine learning model is deployed to a pre-production environment to receive manual review and feedback. The initial machine learning model was adjusted based on the feedback from the manual review before being released.

[0020] By adopting the above technical solution, a complete model initialization and optimization process was constructed, which effectively solved the problem of lack of training data in the cold start stage and realized the rapid construction and optimization of the model from scratch. This not only improved the initial prediction ability of the model, but also laid a solid foundation for subsequent automated updates and strategy linkage.

[0021] Secondly, this application provides an insurance customer renewal intention prediction and management system, which adopts the following technical solution: An insurance customer renewal intention prediction and management system, the prediction and management system comprising: The data acquisition module is used to collect customers' historical interaction data, payment behavior data, and recent activity data; The data extraction and transformation module is used to extract and transform features from the historical interaction data, payment behavior data, and recent activity data to generate dynamic feature vectors. The data processing module is used to input the dynamic feature vector into a pre-trained machine learning model and output the probability value of the customer's renewal intention. The renewal intention prediction module is used to input the dynamic feature vector into a pre-trained machine learning model and output the customer's renewal intention probability value. The risk level matching module is used to classify risk levels based on the customer's renewal intention probability value and generate risk level labels. The task instruction generation module is used to generate task instructions that include reach channels and incentive measures by matching the risk level label with a preset reach strategy. The task instruction execution module is used to drive the corresponding communication interface to perform a reach operation according to the task instruction; The data update module is used to obtain the result feedback data of the outreach operation, add the result feedback data as a new label to the training dataset, and update the machine learning model.

[0022] Thirdly, this application provides a computer-readable storage medium, which adopts the following technical solution: A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the first process of an insurance customer renewal intention prediction and management method according to one embodiment of this application.

[0024] Figure 2 This is a schematic diagram of the second process of the insurance customer renewal intention prediction and management method according to one embodiment of this application.

[0025] Figure 3 This is a schematic diagram of the third process of the insurance customer renewal intention prediction and management method according to one embodiment of this application.

[0026] Figure 4 This is a schematic diagram of the fourth process of the insurance customer renewal intention prediction and management method according to one embodiment of this application.

[0027] Figure 5 This is a schematic diagram of the fifth process of the insurance customer renewal intention prediction and management method according to one embodiment of this application.

[0028] Figure 6 This is a schematic diagram of the sixth process of the insurance customer renewal intention prediction and management method according to one embodiment of this application.

[0029] Figure 7 This is a schematic diagram of the seventh process of the insurance customer renewal intention prediction and management method according to one embodiment of this application.

[0030] Figure 8 This is a schematic diagram of the eighth process of the insurance customer renewal intention prediction and management method according to one embodiment of this application. Detailed Implementation

[0031] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figure 1-8 The present application will be further described in detail below with reference to embodiments. 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 application.

[0032] This application discloses a method for predicting and managing insurance customers' willingness to renew their policies.

[0033] Reference Figure 1 A method for predicting and managing insurance customers' renewal intentions, the prediction and management method including: Step S101: Collect customers' historical interaction data, payment behavior data, and recent activity data; Historical interaction data typically includes a customer's past insurance records, whether they have ever had a lapse in coverage, the frequency of interactions with customer service, and their usage of online systems. These indicators, to some extent, reveal a customer's perception of the insurance company's service experience and their level of loyalty. Payment behavior data focuses on a customer's cash flow, such as preferred payment methods (e.g., bank card transfers, third-party payment tools), whether there have been any delays in payment, and the proportion of premium expenditure in their total income. These variables help assess a customer's financial stability and ability to fulfill their obligations.

[0034] Next, recent activity data focuses more on capturing changes in user activity patterns in the short term, such as whether a claim has been filed, whether a competitor's product details page has been viewed, and the length of time spent accessing the renewal application interface. These real-time signals can often effectively predict the strength of a customer's current interest in continuing to hold the policy.

[0035] The goal of the above steps is to form a basic material library for a comprehensive profile of the target individual through the integration of multi-source heterogeneous data, thereby providing sufficient and effective input support for subsequent modeling.

[0036] Step S102: Extract and transform features from historical interaction data, payment behavior data, and recent activity data to generate dynamic feature vectors; Feature extraction refers to constructing a more representative set of attributes by performing statistical operations, discretization mapping, or other mathematical transformations on the original fields.

[0037] In the embodiments of this application, for categorical variables such as "payment channel type", one-hot encoding can be used to convert them into a set of binary numerical vectors, so that nominal classifications that do not have a natural ordering relationship can be correctly identified and included in the scope of distance or weight calculation.

[0038] For continuous or count variables, the application of aggregate functions within a sliding time window may be involved, such as the total number of customer service calls made in the past N days or the average frequency of web page logins in the past month. This approach not only preserves the characteristics of time-series evolution but also enhances the ability to identify differences between samples.

[0039] In addition, some specific behavior marking mechanisms can be introduced. For example, if a customer stays on the renewal page for more than a preset threshold (such as 60 seconds), it can be considered that they have shown a high degree of attention to the matter, and a new Boolean indicator can be created as one of the additional features.

[0040] Step S103: Input the dynamic feature vector into the pre-trained machine learning model and output the probability value of the customer's renewal intention. In this context, a pre-trained machine learning model means that the parameters have already been initialized using a large set of labeled samples. Common implementation paths include, but are not limited to, classic classifier architectures such as random forest, support vector machine, or gradient boosting decision tree (GBDT). Once trained, the model has the ability to automatically derive the corresponding label distribution based on a given input, specifically referring to the probability that a customer will pay the next year's premium on time.

[0041] It should be noted that, unlike the traditional hard-decision output, the response here is presented as a floating-point number, a confidence score within the range [0,1]. The closer to 1, the more likely the system is that the customer is likely to renew their contract, and vice versa. This soft-decision mechanism gives the entire decision-making chain greater flexibility and controllability, and also provides a quantitative reference for subsequent refined operations.

[0042] Step S104: Classify the risk level based on the probability value of the customer's willingness to renew and generate a risk level label; The risk level labels include low risk, medium risk, or high risk labels; Specifically, by arranging the clients to be tested according to their probability scores from high to low, and setting several dividing points as the cutting criteria between different levels, the risk classification process can be automated.

[0043] For example, when a user's probability of renewing the policy is greater than or equal to 0.8, they are classified into the "low-risk" group, which means that they are likely to fulfill their obligations voluntarily without much intervention; the group between 0.4 and 0.8 is classified into the "medium-risk" category, which indicates that there is some uncertainty but it can still be improved through appropriate guidance; and those below 0.4 are classified as "high-risk" members, who urgently need more resources to focus on and attempt to recover their losses.

[0044] Understandably, the aforementioned tiered system essentially reflects the application of the principle of minimizing opportunity cost in risk management theory, that is, prioritizing the allocation of limited human and material resources to the objects most likely to generate positive returns, thereby maximizing overall benefits.

[0045] Step S105: Based on the risk level label, match the preset outreach strategy to generate a task instruction that includes outreach channels and incentive measures; Among them, the outreach strategy refers to a set of highly targeted communication solutions tailored to the risk attributes of different customers, covering multiple aspects such as the selection of contact media, content expression style, and even feedback and reward mechanisms.

[0046] For example, when dealing with "low-risk" users, since they already have a high degree of self-motivation, it is sufficient to simply use low-cost and efficient information push channels (such as SMS notifications) to inform them of relevant information, without any additional material incentives. In contrast, "medium-risk" customers, although still in a wait-and-see state, still have significant conversion potential. In this case, it is advisable to consider using a voice robot with some intelligent features to make outbound calls, combined with a small amount of points return policy to enhance attractiveness. As for those extreme cases that have been determined to be "high-risk," it is necessary to mobilize higher-level service resources, such as arranging experienced human agents to personally intervene in negotiations, and supplementing them with personalized discount packages or even value-added service upgrade options as strong bargaining chips to facilitate the transaction.

[0047] Step S106: Drive the corresponding communication interface to perform the reach operation according to the task instruction; Among them, communication interfaces refer to various technical channels that can be used for two-way information transmission between enterprises and consumers, including but not limited to SMS gateway APIs provided by telecom operators, instant messaging platform SDK encapsulation modules, CRM system internal work order flow queues, etc.

[0048] The entire triggering process typically relies on an automated dispatch center to send instruction packets in batches to the corresponding terminal nodes according to predetermined rules. The latter are then responsible for parsing the message format and delivering the final message to the designated recipient. Simultaneously, to ensure the entire closed-loop controllability of the chain, it is also necessary to synchronously record the exact time of each event, the identity of the participating roles, a summary of the specific content of the interaction, and other relevant metadata for later auditing and traceability.

[0049] Step S107: Obtain the result feedback data of the touch operation, add the result feedback data as a new label to the training dataset, and update the machine learning model.

[0050] In this embodiment of the application, the result feedback data mainly includes two types of core components: first, the customer's actual reaction after receiving the intervention, such as whether they clicked the link to view the renewal page, whether they actively contacted customer service to inquire about details, and whether they ultimately made a successful payment; second, the auxiliary observation indicators generated throughout the entire interaction process, such as semi-structured supplementary materials like transcripts of call recordings, sentiment analysis reports of email replies, and sentiment index calculation results of social media comments.

[0051] Understandably, once these new instances are aggregated and added to the database, they immediately participate in a new learning cycle. Especially under the premise of an incremental learning paradigm, this allows existing complex models such as GBDT to absorb new knowledge while maintaining their existing performance levels, gradually correcting biases and strengthening robustness. Furthermore, as the sample size continues to expand and accumulate, the importance ranking of various input factors can be periodically reviewed, and the previously set risk threshold boundaries can be fine-tuned accordingly, ensuring that the entire evaluation scale remains up-to-date.

[0052] In other words, this mechanism fundamentally breaks away from the static model's adherence to conventions and instead establishes an intelligent agent system architecture with self-awareness and environmental adaptability.

[0053] The above implementation methods focus on refined governance practices centered around the customer's entire lifecycle behavior trajectory, encompassing an integrated collaborative operation mechanism from the bottom-level perception layer to the top-level cognition layer and then to the feedback control layer. By integrating multiple key technologies such as big data collection, intelligent feature construction, accurate probability inference, dynamic risk classification, customized intervention scheduling, and autonomous evolution and updating, it not only significantly improves the level of refined operation and management of insurance institutions in the competitive landscape of the existing market, but also enhances the company's ability to withstand market fluctuations.

[0054] Reference Figure 2 As one implementation of step S102, the step of extracting and transforming features from historical interaction data, payment behavior data, and recent activity data to generate dynamic feature vectors includes: Step S201: Extract customer service call frequency and online login behavior from historical interaction data, perform frequency statistics calculation according to preset time windows, and generate behavior frequency sequence; This step aims to characterize the intensity of customer service requests and their participation in self-service platforms over a specific time period. Specifically, calculating the average daily number of calls over a 30-day period reflects customers' habit of seeking human assistance when facing problems; while accumulating the number of logins to the online system over a 90-day span reflects customers' trust in digital channels and their proactive use of them. Because these two behaviors involve different psychological motivations and social attributes, different time scales are used to differentiate them, thereby enhancing the model's understanding of the rhythm of changes in customer stickiness and loyalty. The resulting two-dimensional time-series features not only possess good interpretability but also adapt to the needs of various regression or classification tasks.

[0055] Step S202: Extract payment channel types from payment behavior data and classify and aggregate them to generate a payment channel distribution vector; In this method, different types of payment methods are categorized (such as credit card payment, debit card payment, and third-party payment), and one-hot encoding is used to map them into a three-dimensional binary vector, with each dimension representing whether a certain type of payment method has been used.

[0056] Furthermore, the concept of usage ratio can be introduced, which represents the proportion of each payment method in the entire payment record, quantified as a percentage to form another ternary real vector. The concatenation of these two sub-vectors forms a six-dimensional payment channel distribution vector, preserving both the presence and absence of payment methods and reflecting the changing trends in their usage intensity. This dual encoding strategy effectively enhances the information density of feature representation, enabling the model to capture both the static selection and dynamic evolution characteristics of users' payment preferences.

[0057] Step S203: Extract page operation records from recent activity data, compare the single dwell time of renewal-related pages with the threshold, and generate high attention markers; This step focuses on the user's response speed and level of focus at key points. By monitoring the duration and frequency of visits to specific pages (such as renewal reminder pages and price display pages), reasonable judgment criteria are set (e.g., a single visit exceeding 30 seconds is considered in-depth reading, or multiple repeated visits, even if the time limit is not met, indicate potential interest). This allows for the assessment of the user's level of attention to the matter. This labeling mechanism avoids the risk of misjudgment from a single indicator, improves the accuracy of identification, and helps to identify potentially at-risk users in advance and implement intervention measures.

[0058] Step S204 involves dimensional alignment and concatenation of the behavior frequency sequence, payment channel distribution vector, and high attention markers to form a dynamic feature vector.

[0059] To ensure that features from different sources with varying units and value ranges can work collaboratively within a unified space, they must be standardized and scaled to eliminate weight imbalances caused by differences in the dimensions of dependent variables. After standardization, the six-dimensional payment channel features, two-dimensional behavior frequency sequences, and one-dimensional high-attention markers are arranged and combined in a predefined order to form the initial nine-dimensional basic feature set. The resulting dynamic feature vector is not only suitable for training current supervised learning models but also reserves sufficient room for evolution to meet new challenges brought about by changes in the market environment, achieving a leap from static profiling to dynamic insights.

[0060] In addition, to further improve the model's generalization ability and adapt to complex scenario requirements, it can be extended to a higher-dimensional space by combining basic customer information (such as age, gender, occupation category, etc.) and other derived variables (such as claims status, complaint records, etc.) to form a comprehensive dynamic feature vector system.

[0061] In the above implementation, focusing on the goal of predicting insurance customers' renewal intentions, a feature representation framework that combines stability and flexibility is established through the effective extraction, in-depth processing, and organic integration of diverse and heterogeneous data. This framework not only fully explores the deep-seated correlation patterns among customers in terms of payment paths, communication behaviors, and browsing habits, but also enhances the model's robustness to external disturbances and its predictability of future trends by using multi-level time window divisions and multi-granularity evaluation criteria.

[0062] Reference Figure 3 As one implementation of step S103, the step of inputting the dynamic feature vector into a pre-trained machine learning model and outputting the probability value of customer renewal intention includes: Step S301: Load the pre-trained machine learning model; The machine learning model is generated by training historical customer characteristic data and renewal result labels; Specifically, the training process of this machine learning model relies on a large amount of historical customer data, which typically includes basic customer attributes (such as age, gender, and occupation), behavioral characteristics (such as payment frequency and number of consultations), and product-related characteristics (such as policy type and coverage period), collectively referred to as "customer feature vectors".

[0063] Simultaneously, each sample contains a clear label: whether the customer successfully renewed their subscription within a certain time period. This label constitutes the target variable in supervised learning. During training, the model models the mapping relationship between input features and output labels, optimizing parameters with the objective function of minimizing prediction error. The resulting model possesses the ability to generalize and predict new customer renewal behavior. Loading the pre-trained model essentially transfers previously accumulated knowledge to the current application scenario, enabling intelligent judgment of unknown samples.

[0064] Step S302: Input the dynamic feature vector into the machine learning model and perform classification probability prediction calculation; The dynamic feature vector refers to the latest set of features of a current customer that may change over time, such as the status of the most recent payment, recent interaction records, or policy changes. These features reflect the customer's current status and, compared to static historical data, better demonstrate the changing trends in their renewal intentions.

[0065] Next, the vector is input into the loaded model. The model performs forward propagation based on its internal parameter structure (such as the node decision path in the decision tree, the activation function in the neural network, etc.) to obtain the probability value that the customer belongs to the "renewal" category. This process is a typical probability output mechanism in classification tasks. Especially in models such as Gradient Boosting Decision Tree (GBDT), the output probability is usually based on the weighted voting or weighted averaging of multiple weak classifiers, which has strong non-linear fitting ability and robustness.

[0066] Step S303: Output the probability value of the customer's willingness to renew.

[0067] The probability value is a real number between 0 and 1, representing the model's confidence level in whether the customer will renew their policy at some point in the future. For example, an output value of 0.75 means the model believes there is a 75% chance the customer will choose to renew. This continuous output not only provides classification results but, more importantly, preserves uncertainty information, allowing business personnel to take differentiated intervention measures based on different confidence levels.

[0068] In the above implementation, a closed-loop customer renewal intention prediction and risk assessment process was constructed. This process not only makes full use of the potential patterns in historical data, but also fully considers the dynamic changes in customer status. By enhancing the interpretability of model output and effectively integrating business rules, the intelligence level and operational efficiency of insurance companies in customer lifecycle management are improved.

[0069] Reference Figure 4 As one implementation of step S104, the step of classifying risk levels based on the probability value of customer renewal intention and generating risk level labels includes: Step S401: Receive the customer's renewal intention probability value; The probability value is a continuous value within the interval [0,1]. Specifically, the probability value of a customer's willingness to renew is usually output by a pre-trained machine learning model (such as logistic regression, random forest, or deep neural network), representing the likelihood that a particular customer will continue to pay premiums in the future. This probability value belongs to a closed interval [0,1] in the real number field, where 0 represents no intention to renew, 1 represents a very high probability of renewing, and intermediate values ​​reflect different degrees of inclination.

[0070] Step S402: Load preset grading threshold parameters, which include a first threshold and a second threshold. The first and second thresholds are not fixed constants, but rather a set of critical points determined through preliminary data analysis, expert experience summarization, and multiple testing and verifications. These two thresholds function similarly to "decision boundaries," collectively defining the dividing lines between three different risk levels.

[0071] For example, assuming the current first threshold is 0.8 and the second threshold is 0.4, this means that any customer whose renewal intention score is 0.8 or higher can be considered a low-risk user; if the score is between 0.4 and 0.8, it is classified as medium-risk; and those below 0.4 are marked as high-risk. This three-tiered structure helps companies take differentiated intervention measures for different types of customers, improving resource utilization and service efficiency.

[0072] Step S403: Compare the customer's renewal intention probability value with the first threshold and the second threshold to obtain the comparison result; Specifically, by comparing the input single floating-point value (i.e. the probability of renewal intention) with its upper and lower thresholds in turn, it can be mapped to three mutually exclusive state categories.

[0073] In addition, considering the potential problems caused by floating-point precision, a certain tolerance mechanism needs to be introduced to deal with the fuzzy classification problem in boundary cases, such as how to reasonably define the category when a certain probability happens to fall around 0.8.

[0074] Step S404: Output risk level labels based on the comparison results; Specifically, if the probability value is greater than or equal to the first threshold, a low-risk label is output; if the probability value is less than the first threshold but greater than or equal to the second threshold, a medium-risk label is output; and if the probability value is less than the second threshold, a high-risk label is output. It's important to note that these three labels not only represent different levels of potential churn, but more importantly, they serve as the basis for selecting a series of differentiated operational strategies. For example, when a customer's renewal intention score is 0.76, because it's below the first threshold (0.8) but above the second threshold (0.4), it will be labeled "medium risk," triggering the AI ​​customer service module to proactively contact the customer to understand the reasons for the late renewal. Conversely, if the score is 0.38, it will be labeled "high risk" and prioritized for handling by a human agent. The advantage of this multi-level labeling system is that it can balance cost-effectiveness and service quality, enabling companies to implement precise and targeted strategies even when dealing with a massive customer base.

[0075] Step S405: Associate and store the risk level label in the customer profile database.

[0076] This associated storage refers to establishing structured links between newly generated risk tags and other existing fields (such as customer ID, contact information, and past spending records). This allows for quick location of relevant information during subsequent queries or report display, enabling cross-table fusion retrieval and other functions. Simultaneously, it provides strong support for building a complete customer view, helping businesses better understand the overall situation and evolving trends of each user.

[0077] For example, whenever a new customer enters the renewal cycle one month in advance, the system will automatically capture the most recently updated risk level label and combine it with other auxiliary indicators (such as age distribution, regional origin, etc.) to generate a personalized reminder plan, which can not only improve work efficiency, but also significantly enhance user experience satisfaction.

[0078] The above implementation scheme, relying on scientifically sound probabilistic modeling results and employing refined threshold control methods, effectively identifies and classifies the renewal intention levels of a large number of customers. This solution fully considers the complexity and diversity of real-world application scenarios, maintaining high versatility while also possessing strong scalability and intelligent potential, making it significant for promoting the digital transformation of the modern insurance industry.

[0079] Reference Figure 5 As one implementation of step S105, the step of generating a task instruction containing reach channels and incentives based on risk level labels and a preset reach strategy includes: Step S501: Receive customer risk level labels, including three categories: low risk, medium risk, or high risk. By classifying customers into three risk levels, the system can develop differentiated intervention strategies based on different risk levels, avoiding the inefficient "one-size-fits-all" management approach and thereby improving overall customer retention and operational efficiency.

[0080] Step S502: Query the preset strategy rule base and match the reach strategy template corresponding to the risk level label; The predictive strategy rule base is essentially a rule engine that stores a variety of preset strategy templates. Each template corresponds to a different customer risk level and defines elements such as the contact channels, incentive measures, and communication strategies to be adopted.

[0081] Specifically, the outreach strategy templates are based on statistical analysis of extensive historical data and accumulated operational experience. For example, SMS outreach is used for low-risk customers because these customers are typically highly sensitive to disruptions and have a strong willingness to renew their policies; a low-cost, non-intrusive approach can maintain customer relationships. AI-powered outbound calling combined with points incentives is used for medium-risk customers, considering they are on the verge of churn and require intervention through proactive communication and mild incentives. For high-risk customers, human agents are employed, and substantial discounts are offered to maximize customer retention. This step, through a rule-matching mechanism, automatically converts customer risk labels into specific execution strategies, providing a strategic basis for subsequent task generation.

[0082] Step S503: Generate structured task instructions based on the outreach strategy template, including outreach channel type and incentive identifier; The type of outreach channel determines the execution method of the task, such as SMS, AI outbound calling, and human agents. Different channels correspond to different costs, efficiency, and customer experience. The incentive identifier is used to identify the incentive methods that should be provided to customers during the task execution process, such as points, coupons, and upgraded protection.

[0083] For example, a task instruction might be: {Channel Type: AI Outbound Call, Incentive Identifier: Points +100, Customer ID: C12345}. This instruction clearly defines the execution method and incentive content, making it easy for the system to automatically call and execute it in the future.

[0084] Step S504: Write the structured task instructions into the task queue database and associate them with the customer's unique identification code.

[0085] Each task instruction must be associated with a unique customer identification code (such as a customer ID) when it is written, so as to accurately identify the target customer during the task execution process and to perform data backtracking and analysis in the subsequent execution result feedback.

[0086] Furthermore, the task queue design must support features such as task priority sorting, execution status updates, and failure retries to ensure tasks are executed efficiently as planned. For example, tasks from high-risk clients can be assigned higher priority to ensure they are processed first when resources are limited. By writing task instructions to the task queue database, the system achieves unified task scheduling and asynchronous execution, improving the overall system's execution efficiency and resource utilization.

[0087] In the above implementation, strategy matching is completed by combining a strategy rule base, generating structured task instructions, and finally writing them into the task queue database, thereby realizing a complete closed loop from customer risk identification to the execution of differentiated outreach strategies. This solution not only improves the efficiency of intervention in customer renewal intentions, but also enhances the system's intelligence level and the efficiency of operational resource utilization through structured task management and differentiated strategy configuration.

[0088] Reference Figure 6 As one implementation of step S106, the step of driving the corresponding communication interface to perform a reach operation according to the task instruction includes: Step S601: Read the structured task instruction from the task queue database. The instruction includes the customer's unique identifier, the contact channel type, and the incentive identifier. The unique customer identifier is used to accurately locate the identity information of the target audience; the contact channel type determines the interaction method (such as SMS, telephone, or human customer service); and the incentive identifier corresponds to the specific marketing tool (such as coupons, points rewards, or access to value-added services). This design not only improves data consistency and processability but also facilitates downstream modules in flexibly matching resources and service paths according to different business needs.

[0089] Step S602: Invoke the corresponding communication interface protocol according to the type of the reach channel; Among them, communication interface protocols refer to the standard API interface set for connecting to external communication service platforms, which are usually divided into various categories based on factors such as transmission method, security level and response time.

[0090] For example, SMS platforms use HTTP / HTTPS RESTful APIs, supporting message body encapsulation in JSON / XML formats; while AI voice robots tend to use the low-latency, high-concurrency WebSocket bidirectional communication protocol to achieve real-time audio stream delivery; and human agent systems may use browser-based message buses or enterprise internal instant messaging frameworks for task distribution. Because these interfaces differ significantly in their underlying implementation, an abstract adapter pattern is necessary for unified call control, thereby shielding the complexity caused by heterogeneous environments.

[0091] Step S603: Send an instruction execution request to the target communication interface. The instruction execution request carries customer contact information and incentive identifier. In addition to conveying basic customer contact information, the previously extracted incentive identifiers also need to be further transformed into usable content carriers. For example, if a customer's incentive is a "discount e-coupon," then a unique redemption code needs to be obtained through the relevant authentication interface of the payment settlement center before being embedded into the message template to be sent.

[0092] In addition, considering that some customers may refuse to accept the package or switch between multiple devices, it is also necessary to set up an error rollback plan in advance so that the system can quickly switch to the alternative solution when there is network jitter or third-party service unavailability, thereby improving the overall success rate and user experience satisfaction.

[0093] Step S604: Receive the operation status code returned by the target communication interface and update the task execution status log.

[0094] The numerical range of operation status codes often covers various scenarios such as success, timeout, authentication failure, and missing parameters. By collecting and classifying these status codes in real time, we can not only accurately grasp the actual effectiveness of each customer interaction, but also feed back into the upstream data analysis engine, forming a closed-loop optimization chain.

[0095] More importantly, to meet regulatory compliance and audit trail requirements, all relevant event nodes need to be fully recorded to form a detailed log archive, including but not limited to trigger timestamps, executor identities, communication quality indicators (such as RTT latency), and failure reason classifications. This not only helps operations and maintenance personnel quickly locate potential sources of failure but also provides strong support for future risk investigation.

[0096] The above implementation, relying on a sophisticated task orchestration mechanism and integrating diverse communication methods with personalized incentive combinations, maintains good robustness and scalability even under the pressure of large-scale concurrent access. Especially in the face of sudden network fluctuations or hardware anomalies, the multi-layered automatic degradation strategy effectively avoids service interruptions caused by the failure of a single node. Simultaneously, by introducing a sophisticated operation log management system, the entire customer interaction process becomes transparent and controllable, enhancing the enterprise's operational efficiency and risk management level.

[0097] Reference Figure 7 As one implementation of step S107, the steps of obtaining the result feedback data of the touch operation, adding the result feedback data as a new label to the training dataset, and updating the machine learning model include: Step S701: Collect the operation status code returned after the task is executed. The operation status code includes the renewal operation result identifier. In this context, "reaching out" refers to a series of proactive communication actions initiated by insurance companies to promote policy renewal, such as outbound phone calls, SMS push notifications, and email reminders. After these tasks are completed, the system records the final status of each interaction, such as whether the customer was contacted, whether the customer intends to continue the policy, and whether the payment has been completed. These statuses are often represented in coded form (i.e., "status codes"), which have good standardization characteristics and facilitate subsequent automated processing.

[0098] Step S702: Map the renewal operation result identifier to a binary classification label; Among them, successful renewals are marked as positive samples, and renewal failures or rejections are marked as negative samples. Specifically, in the insurance industry, a customer's willingness to renew their policy is essentially a binary decision process: either accept the continued coverage (success) or terminate the partnership (failure). Therefore, adopting a binary classification labeling system not only meets the essential needs of the business but also facilitates the selection of appropriate classification algorithms for modeling.

[0099] It's important to note that this mapping isn't a simple copy of the original status codes; it requires semantic understanding and normalization. For example, "paid" and "contract signed" should be categorized as "positive samples," while "rejected calls more than three times" and "explicitly stated no intention to renew" are classified as "negative samples." Furthermore, information in intermediate states, such as "no reply yet" and "under consideration," may be filtered out and excluded from training. This ensures that each data point in the training set represents a definitive conclusion, thereby improving the model's learning efficiency and generalization ability.

[0100] Step S703: Associate the binary classification labels with the corresponding dynamic feature vectors of customers and store them in the training dataset; By binding the previously extracted dynamic feature vectors with their corresponding binary classification labels, learning instances with explicit target values ​​can be constructed. The term "dynamic" refers to the fact that these features are not static; they evolve over time. For example, if a customer frequently inquires about insurance product information recently, their interest activity index will increase accordingly. This design helps capture trends in user behavior, enabling the model to adapt promptly to market fluctuations and changing customer needs.

[0101] Step S704: Perform incremental training on the pre-trained machine learning model using the updated training dataset to generate a new version of the model; Understandably, in real-world applications, if the model needs to be fully retrained using all historical data each time, it would not only be extremely time-consuming but also a serious waste of resources.

[0102] In contrast, incremental learning only requires fine-tuning for newly added valid samples, saving computational costs while maintaining the continuity and stability of model performance. This process typically does not change the original network architecture or hyperparameter configuration; model evolution is achieved simply by adjusting local weights along the gradient direction of the new samples. This approach is particularly suitable for rapid iterative optimization in online service scenarios, enabling rapid response to the latest business changes while ensuring model robustness.

[0103] Step S705: Load the new version of the model to overwrite the old model.

[0104] In a real-world production environment, once a new model version passes the evaluation and testing phase, it needs to be deployed and used as soon as possible to ensure that the latest predictive capabilities serve the next round of customer outreach initiatives. This "loading" involves more than just simple file replacement; it also includes a series of complex engineering operations such as memory cache refreshing, interface compatibility checks, and runtime context switching.

[0105] More importantly, due to the use of incremental training, the new and old models have a high degree of inheritance and smooth transition characteristics. Therefore, even if the model is replaced without stopping the system, it will not cause significant service interruption or sudden changes in prediction bias.

[0106] The above implementation achieves closed-loop control across the entire process from the business end to the algorithm end, enhancing the system's autonomous learning capability and intelligent adjustment level. Through refined state filtering and label definition, the quality and representativeness of training samples are ensured. Combined with dynamic feature management and incremental learning strategies, model maintenance costs are significantly reduced, and the iteration pace is accelerated. Simultaneously, a gradual upgrade path is adopted at the deployment level to ensure that online service quality remains unaffected. This technical solution not only enhances the predictive accuracy and intervention effectiveness of insurance companies in customer retention but also provides a valuable reference paradigm for other similar customer relationship management systems.

[0107] Reference Figure 8 As a further implementation of the method for predicting and managing insurance customers' renewal intentions, before the step of inputting the dynamic feature vector into the pre-trained machine learning model, the method further includes: Step S801: When the training dataset is detected to be empty, the static snapshot data of the historical policy database is called. This step aims to address the "cold start" problem in machine learning model training, which occurs when there are insufficient labeled samples for model training during the initial deployment of a system or the launch of a new business. An empty training dataset means that the dataset used for model training has not yet been fully constructed, or that effective training samples cannot be obtained due to the specific business scenario or the insufficient scale of the customer base. In this case, directly training the model will result in the model failing to converge or exhibiting extremely poor generalization ability.

[0108] To address this issue, the system accesses "static snapshot data" in the historical policy database to obtain historical information suitable for modeling. Static snapshot data typically refers to a complete record of a customer's policy status at a specific point in time, including but not limited to the customer's basic attributes (such as age, gender, and occupation), policy type, payment history, and historical claims. While this data lacks real-time dynamism, it possesses high stability and representativeness, reflecting a customer's behavioral patterns and risk characteristics at a particular stage.

[0109] Step S802: Perform manual annotation on the static snapshot data to generate an initial training label set; Since historical data itself does not contain explicit target variables (such as whether a customer will renew their policy or default on their policy), manual annotation is necessary to generate the labels required for supervised learning. This process involves business experts judging customer behavioral tendencies based on experience and assigning a corresponding category label (such as "will renew" or "will not renew") to each sample. Manual annotation not only improves data quality but also ensures the accuracy and consistency of the labels, thus providing a reliable target output for subsequent model training.

[0110] It's worth noting that stratified sampling strategies may be used during manual annotation to ensure representativeness across different customer groups; crowdsourcing mechanisms can also be introduced to improve annotation efficiency. The generated initial training label set, together with the original feature data, will constitute a complete training sample set, laying the foundation for subsequent modeling.

[0111] Step S803: Train the initial machine learning model based on the XGBoost basic parameters; XGBoost is a high-efficiency machine learning algorithm based on gradient boosting decision tree (GBDT). It has good generalization ability, strong robustness, and supports parallel computing. It is widely used in fields such as financial risk control and insurance pricing.

[0112] In this embodiment, XGBoost was chosen over other models because of its excellent performance in handling structured data and its strong tolerance for missing and outlier values, making it very suitable for modeling with limited samples in the initial stages. XGBoost's basic parameters refer to a set of preset hyperparameter configurations, such as the maximum tree depth, learning rate, and regularization coefficient. These parameters are typically set based on experience or determined through simple grid search, with the aim of quickly building a working initial model version. Although this initial machine learning model has not undergone fine-tuning, it still possesses a certain predictive ability, providing support for subsequent manual review and strategy iteration.

[0113] Step S804: Deploy the initial machine learning model to the pre-production environment and receive feedback from manual review. The pre-production environment is an intermediate deployment phase between the development and testing environment and the formal production environment. Its purpose is to conduct small-scale verification and evaluation of the model's predictive performance without affecting online business.

[0114] In this environment, the model will make predictions based on some real business data, and the prediction results will be manually reviewed by business experts or risk control personnel. The review includes, but is not limited to, whether the probabilities output by the model are reasonable, whether there are obvious biases, and whether they conform to business logic. This human-machine collaborative approach not only identifies potential problems with the model (such as overfitting, underfitting, and feature misuse), but also provides valuable experience feedback for subsequent model optimization.

[0115] In addition, the pre-production environment allows the system to simulate real business processes, verify the interface compatibility between the model and other modules, and the correctness of data flow, thereby reducing the risk of going live.

[0116] Step S805: Adjust the hyperparameters of the initial machine learning model based on the feedback from manual review and release it.

[0117] Hyperparameter tuning is a crucial step in the model optimization process, directly impacting model performance. Human feedback obtained in the pre-production environment provides important guidance for hyperparameter tuning. For example, if the model shows significant prediction bias for certain customer groups, its expressive power can be enhanced by adjusting tree depth or adding regularization terms. If the model's response speed is slow, inference efficiency can be improved by limiting the number of trees or reducing the number of iterations.

[0118] In some embodiments, commonly used hyperparameter optimization methods include Bayesian optimization, random search, and grid search, which can be flexibly selected according to actual needs. After adjustment, the system will retrain the model and evaluate its performance on the validation set. Once confirmed to be error-free, it will be released as a formal version for production use. This process embodies the closed-loop management concept from "cold start" to "continuous optimization," ensuring that the model is always in the best operating state.

[0119] The above implementation constructs a complete model initialization and optimization process, effectively solving the problem of lack of training data in the cold start stage, and realizing the rapid construction and optimization of the model from scratch. This not only improves the initial prediction ability of the model, but also lays a solid foundation for subsequent automated updates and policy linkage.

[0120] This application also discloses an insurance customer renewal intention prediction and management system.

[0121] An insurance customer renewal intention prediction and management system, the prediction and management system includes: The data acquisition module is used to collect customers' historical interaction data, payment behavior data, and recent activity data; The data extraction and transformation module is used to extract and transform features from historical interaction data, payment behavior data, and recent activity data to generate dynamic feature vectors. The data processing module is used to input dynamic feature vectors into a pre-trained machine learning model and output the probability value of customer renewal intention. The renewal intention prediction module is used to input dynamic feature vectors into a pre-trained machine learning model and output the probability value of the customer's renewal intention. The risk level matching module is used to classify risk levels based on the probability value of customers' willingness to renew and generate risk level labels. The task instruction generation module is used to match the preset outreach strategy based on the risk level label and generate task instructions that include outreach channels and incentive measures. The task instruction execution module is used to drive the corresponding communication interface to perform the reach operation according to the task instruction; The data update module is used to obtain the result feedback data of the outreach operation, add the result feedback data as a new label to the training dataset, and update the machine learning model.

[0122] An insurance customer renewal intention prediction and management system according to an embodiment of this application can implement any of the above prediction and management methods, and the specific working process of each module in the prediction and management system can refer to the corresponding process in the above method embodiments.

[0123] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0124] This application also discloses a computer-readable storage medium.

[0125] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the methods for predicting and managing insurance customer renewal intentions.

[0126] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0127] In this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0128] Although this application has been described herein in conjunction with various embodiments, those skilled in the art, by reviewing the accompanying drawings, disclosure, and appended claims, will understand and implement other variations of the disclosed embodiments in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce a good effect.

[0129] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for predicting and managing insurance customers' renewal intentions, characterized in that, The method includes: Collect customers' historical interaction data, payment behavior data, and recent activity data; The historical interaction data, payment behavior data, and recent activity data are subjected to feature extraction and transformation to generate dynamic feature vectors; The dynamic feature vector is input into a pre-trained machine learning model, which outputs a probability value of the customer's willingness to renew. Risk levels are classified based on the probability values ​​of customer renewal intentions, and risk level labels are generated. Based on the risk level labels, a preset outreach strategy is matched to generate a task instruction that includes outreach channels and incentive measures; The corresponding communication interface is driven to perform a reach operation according to the task instruction; Obtain the result feedback data of the reach operation, add the result feedback data as a new label to the training dataset, and update the machine learning model.

2. The method for predicting and managing insurance customer renewal intentions according to claim 1, characterized in that, The steps for extracting and transforming features from the historical interaction data, payment behavior data, and recent activity data to generate dynamic feature vectors include: Extract customer service call frequency and online login behavior from the historical interaction data, perform frequency statistics calculation according to a preset time window, and generate behavior frequency sequence; Extract the payment channel types from the payment behavior data and classify and aggregate them to generate a payment channel distribution vector; Extract page operation records from the recent activity data, compare the single dwell time of the relevant page with a threshold, and generate a high attention tag; The frequency sequence of behaviors, the distribution vector of payment channels, and the high attention markers are aligned and concatenated to form a dynamic feature vector.

3. The method for predicting and managing insurance customer renewal intentions according to claim 1, characterized in that, The steps of inputting the dynamic feature vector into a pre-trained machine learning model and outputting the probability value of customer renewal intention include: Load a pre-trained machine learning model; wherein the machine learning model is generated through training using historical customer feature data and renewal result labels; The dynamic feature vector is input into the machine learning model to perform classification probability prediction calculation; Output the probability value corresponding to the customer's willingness to renew.

4. The method for predicting and managing insurance customer renewal intentions according to claim 1, characterized in that, The steps for classifying risk levels and generating risk level labels based on the probability values ​​of customer renewal intentions include: Probability of receiving customer renewal intentions; Load preset grading threshold parameters, the parameters including a first threshold and a second threshold; The customer's renewal intention probability value is compared with the first threshold and the second threshold to obtain the comparison result; Output a risk level label based on the comparison results; The risk level labels are associated and stored in the customer profile database.

5. The method for predicting and managing insurance customer renewal intentions according to claim 1, characterized in that, The steps for generating task instructions that include reach channels and incentives based on the risk level labels and pre-defined reach strategies include: Receive customer risk level labels, including three categories: low risk, medium risk, or high risk; Query the preset strategy rule base and match the reach strategy template corresponding to the risk level label; A structured task instruction is generated based on the reach strategy template, including the reach channel type and incentive identifier; The structured task instructions are written into the task queue database and associated with the customer's unique identification code.

6. The method for predicting and managing insurance customer renewal intentions according to claim 5, characterized in that, The steps for executing a reach operation based on the corresponding communication interface driven by the task instruction include: Read structured task instructions from the task queue database, the instructions containing a unique customer identifier, a reach channel type, and an incentive identifier; The corresponding communication interface protocol is invoked based on the type of the reach channel. Send an instruction execution request to the target communication interface, the instruction execution request carrying customer contact information and incentive identifier; Receive the operation status code returned by the target communication interface and update the task execution status log.

7. The method for predicting and managing insurance customer renewal intentions according to claim 6, characterized in that, The steps of obtaining feedback data on the results of the outreach operation, adding the feedback data as a new label to the training dataset, and updating the machine learning model include: Collect the operation status code returned after the task is executed, and the operation status code includes a renewal operation result identifier; Map the renewal operation result identifier to a binary classification label; The binary classification labels are associated with the dynamic feature vectors of the corresponding customers and stored in the training dataset; The pre-trained machine learning model is incrementally trained using the updated training dataset to generate a new version of the model; Load the new version of the model to overwrite the old model.

8. A method for predicting and managing insurance customer renewal intentions according to any one of claims 1 to 7, characterized in that, Before the step of inputting the dynamic feature vector into the pre-trained machine learning model, the method further includes: When the training dataset is detected to be empty, static snapshot data from the historical policy database is retrieved; Manual annotation is performed on the static snapshot data to generate an initial training label set; The initial machine learning model is trained based on the XGBoost basic parameters; The initial machine learning model is deployed to a pre-production environment to receive manual review and feedback. The initial machine learning model was adjusted based on the feedback from the manual review before being released.

9. An insurance customer renewal intention prediction and management system, characterized in that, The system includes: The data acquisition module is used to collect customers' historical interaction data, payment behavior data, and recent activity data; The data extraction and transformation module is used to extract and transform features from the historical interaction data, payment behavior data, and recent activity data to generate dynamic feature vectors. The data processing module is used to input the dynamic feature vector into a pre-trained machine learning model and output the probability value of the customer's renewal intention. The renewal intention prediction module is used to input the dynamic feature vector into a pre-trained machine learning model and output the customer's renewal intention probability value. The risk level matching module is used to classify risk levels based on the customer's renewal intention probability value and generate risk level labels. The task instruction generation module is used to generate task instructions that include reach channels and incentive measures by matching the risk level label with a preset reach strategy. The task instruction execution module is used to drive the corresponding communication interface to perform a reach operation according to the task instruction; The data update module is used to obtain the result feedback data of the outreach operation, add the result feedback data as a new label to the training dataset, and update the machine learning model.

10. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 8.

Images