Customer personalized retention strategy determination method and device, equipment and storage medium

Abstract

The invention provides a customer personalized retention strategy determination method and device, equipment and a storage medium, and the method comprises the steps: collecting multi-modal data of a target customer; constructing and updating a dynamic customer portrait of the target customer based on the multi-modal data; extracting customer features from the dynamic customer portrait, and inputting the customer features into a dual-channel insurance renewal prediction model to obtain an insurance policy insurance renewal probability; performing interpretability analysis on the insurance renewal probability of the insurance policy, and determining key features influencing the insurance renewal probability; and generating a personalized retention strategy of the target customer based on the key features. By adopting the scheme, the scientificity and effectiveness of insurance policy renewal prediction and customer retention work can be comprehensively improved.

Description

Technical Field

[0001] This application relates to the field of smart insurance, specifically a method, apparatus, device, and storage medium for determining personalized customer retention strategies. Background Technology

[0002] In insurance business operations, the policy renewal rate is a key indicator for insurance companies to maintain customer base and ensure business stability. Accurately judging customers' intention to renew their policies and taking effective retention measures is of great significance to enhancing industry competitiveness.

[0003] In current insurance industry business, existing technical solutions mainly rely on manual determination: staff extract structured data such as basic customer attributes, policy information and historical claims records from core business systems, filter key information based on business experience, set judgment rules, manually assess the customer's renewal risk level, and formulate unified standardized retention strategies for different levels (such as uniformly sending reminder text messages and arranging batch telephone follow-ups).

[0004] The existing technology has obvious drawbacks: the manual determination process relies on historical static data and personal experience, which cannot capture the real-time dynamic changes in customer behavior and demand during the policy period. This results in a lack of timeliness and accuracy in judging customers' intention to renew their policies, making it difficult to match retention strategies with the actual situation of customers and affecting customer retention. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method, device, equipment and storage medium for determining personalized customer retention strategies, which can comprehensively improve the scientificity and effectiveness of policy renewal forecasting and customer retention work.

[0006] In a first aspect, embodiments of this application provide a method for determining a personalized customer retention strategy, the method comprising: Collect multimodal data from target customers; Based on the multimodal data, construct and update dynamic customer profiles of target customers; Customer features are extracted from the dynamic customer profile, and the customer features are input into the dual-channel renewal prediction model to obtain the policy renewal probability; An interpretability analysis was performed on the policy renewal probability to identify the key features affecting the renewal probability. Based on the aforementioned key features, a personalized retention strategy for the target customer is generated.

[0007] Optionally, the multimodal data includes static attribute data, time-series behavior log data, unstructured text data, and macro-environmental data; The process of constructing and updating a dynamic customer profile for the target customer based on the multimodal data includes: The static attribute data, the time-series behavior log data, the unstructured text data, and the macro-environment data are subjected to multimodal data fusion processing to construct the dynamic customer profile, wherein the dynamic customer profile includes customer attribute dimensions, behavioral trend dimensions, emotional tendency dimensions, and environmental association dimensions. The dynamic customer profile is continuously evolved using an event-triggered mechanism and a periodic incremental update mechanism, wherein: The event triggering mechanism immediately updates the corresponding dimension of the dynamic customer profile when a specific business event occurs; The periodic incremental update mechanism performs sliding window calculations on the time-series behavior log data and the sentiment sequence data extracted from unstructured text data according to a fixed time window, so as to update the trend indicators in the dynamic customer profile.

[0008] Optionally, customer features are extracted from the dynamic customer profile, including: Static attribute features and policy features are extracted from the customer attribute dimensions of the dynamic customer profile. From the behavioral trend dimension of the dynamic customer profile, extract the statistical features, trend features, and sequence pattern features of the dynamic behavior sequence; From the emotional tendency dimension of the dynamic customer profile, extract emotional score features and interest topic features; Environmental association features are extracted from the environmental association dimension of the dynamic customer profile.

[0009] Optionally, inputting the customer characteristics into the dual-channel renewal prediction model to obtain the policy renewal probability includes: The static attribute features, the policy features, and the environmental association features are input into the static feature channel of the dual-channel renewal prediction model and processed through a fully connected network to obtain a static feature vector. The statistical features, trend features, sequence pattern features, sentiment score features, and topic of concern features are input into the dynamic time-series channel of the dual-channel insurance renewal prediction model. The time-series dependencies are modeled through gated recurrent units or long short-term memory networks, and an attention mechanism is used to assign weights to features at different time steps to obtain a dynamic time-series feature vector. The static feature vector and the dynamic time-series feature vector are fused together, and the policy renewal probability is output by calculating through a fully connected layer and the Softmax function.

[0010] Optionally, an interpretability analysis is performed on the policy renewal probability to determine key features affecting the renewal probability, including: The decision attribution of the policy renewal probability is performed, and SHAP or LIME interpretability analysis technology is applied to perform reverse analysis of the model decision process that generates the probability. Based on the reverse analysis, the contribution of each input feature of the dual-channel renewal prediction model to the policy renewal probability prediction result is calculated. Based on a preset contribution threshold, input features whose absolute contribution value exceeds the threshold are identified as key features, and a key influence factor report containing each key feature and its corresponding contribution value is generated.

[0011] Optionally, generating a personalized retention strategy for the target customer based on the key features includes: The key features are matched with a preset feature strategy mapping rule library, wherein the feature strategy mapping rule library predefines the mapping relationship between key feature combinations and retention intervention actions; Based on the matching results, generate personalized retention strategy suggestions that include at least one specific intervention action.

[0012] Optionally, the method further includes: After implementing the personalized retention strategy, collect customer feedback data and renewal result data; The customer feedback data and the renewal result data are used as new multimodal data and fed back into the dynamic customer profile update process; Using the renewal result data as labels, the dual-channel renewal prediction model is retrained to achieve closed-loop optimization.

[0013] Secondly, embodiments of this application provide a device for determining personalized customer retention strategies, the device comprising: The multimodal data acquisition module is used to collect multimodal data from target customers. The dynamic customer profile building module is used to build and update dynamic customer profiles of target customers based on the multimodal data. The policy renewal probability determination module is used to extract customer features from the dynamic customer profile and input the customer features into the dual-channel renewal prediction model to obtain the policy renewal probability. The key feature determination module is used to perform interpretability analysis on the policy renewal probability and determine the key features that affect the renewal probability. The retention strategy generation module is used to generate personalized retention strategies for the target customers based on the key features.

[0014] Optionally, the multimodal data includes static attribute data, time-series behavior log data, unstructured text data, and macro-environmental data; The process of constructing and updating a dynamic customer profile for the target customer based on the multimodal data includes: The static attribute data, the time-series behavior log data, the unstructured text data, and the macro-environment data are subjected to multimodal data fusion processing to construct the dynamic customer profile, wherein the dynamic customer profile includes customer attribute dimensions, behavioral trend dimensions, emotional tendency dimensions, and environmental association dimensions. The dynamic customer profile is continuously evolved using an event-triggered mechanism and a periodic incremental update mechanism, wherein: The event triggering mechanism immediately updates the corresponding dimension of the dynamic customer profile when a specific business event occurs; The periodic incremental update mechanism performs sliding window calculations on the time-series behavior log data and the sentiment sequence data extracted from unstructured text data according to a fixed time window, so as to update the trend indicators in the dynamic customer profile.

[0015] Optionally, customer features are extracted from the dynamic customer profile, including: Static attribute features and policy features are extracted from the customer attribute dimensions of the dynamic customer profile. From the behavioral trend dimension of the dynamic customer profile, extract the statistical features, trend features, and sequence pattern features of the dynamic behavior sequence; From the emotional tendency dimension of the dynamic customer profile, extract emotional score features and interest topic features; Environmental association features are extracted from the environmental association dimension of the dynamic customer profile.

[0016] Optionally, inputting the customer characteristics into the dual-channel renewal prediction model to obtain the policy renewal probability includes: The static attribute features, the policy features, and the environmental association features are input into the static feature channel of the dual-channel renewal prediction model and processed through a fully connected network to obtain a static feature vector. The statistical features, trend features, sequence pattern features, sentiment score features, and topic of concern features are input into the dynamic time-series channel of the dual-channel insurance renewal prediction model. The time-series dependencies are modeled through gated recurrent units or long short-term memory networks, and an attention mechanism is used to assign weights to features at different time steps to obtain a dynamic time-series feature vector. The static feature vector and the dynamic time-series feature vector are fused together, and the policy renewal probability is output by calculating through a fully connected layer and the Softmax function.

[0017] Optionally, an interpretability analysis is performed on the policy renewal probability to determine key features affecting the renewal probability, including: The decision attribution of the policy renewal probability is performed, and SHAP or LIME interpretability analysis technology is applied to perform reverse analysis of the model decision process that generates the probability. Based on the reverse analysis, the contribution of each input feature of the dual-channel renewal prediction model to the policy renewal probability prediction result is calculated. Based on a preset contribution threshold, input features whose absolute contribution value exceeds the threshold are identified as key features, and a key influence factor report containing each key feature and its corresponding contribution value is generated.

[0018] Optionally, generating a personalized retention strategy for the target customer based on the key features includes: The key features are matched with a preset feature strategy mapping rule library, wherein the feature strategy mapping rule library predefines the mapping relationship between key feature combinations and retention intervention actions; Based on the matching results, generate personalized retention strategy suggestions that include at least one specific intervention action.

[0019] Optionally, the device further includes a closed-loop optimization module for: After implementing the personalized retention strategy, collect customer feedback data and renewal result data; The customer feedback data and the renewal result data are used as new multimodal data and fed back into the dynamic customer profile update process; Using the renewal result data as labels, the dual-channel renewal prediction model is retrained to achieve closed-loop optimization.

[0020] Thirdly, embodiments of this application provide a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the customer personalized retention strategy determination method described in any of the optional embodiments of the first aspect above are performed.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the customer personalized retention strategy determination method described in any of the optional embodiments of the first aspect.

[0022] The technical solution provided in this application includes, but is not limited to, the following beneficial effects: Collecting multimodal data from target customers breaks through the limitations of traditional methods that rely solely on single structured static data. By integrating multidimensional data, it comprehensively covers basic customer attributes, dynamic behaviors, interaction information, and related external data, providing a rich and comprehensive data foundation for subsequent dynamic customer profile construction and renewal probability prediction. This ensures that each subsequent step can be carried out based on complete data and avoids analytical biases caused by missing data.

[0023] Building and updating dynamic customer profiles based on multimodal data transforms customer profiles from static tag sets to dynamic evolution models. Through a continuous update mechanism, it captures changes in customer status in real time during the policy period, ensuring that customer profiles always remain consistent with the actual situation of customers. This effectively solves the problem of traditional static profiles being out of touch with real-time customer status, providing accurate profile support for accurately extracting customer characteristics and predicting renewal intentions.

[0024] By extracting customer features from dynamic customer profiles and inputting these features into a dual-channel renewal prediction model to obtain the policy renewal probability, this step extracts multi-dimensional features from dynamic profiles and combines them with targeted processing of static and dynamic features by the dual-channel model. This approach not only makes full use of fixed customer attributes and environmental information but also accurately captures the temporal dependencies of dynamic customer behavior, significantly improving the accuracy of policy renewal probability prediction and providing a reliable probabilistic basis for developing effective retention strategies.

[0025] Interpretability analysis of policy renewal probabilities identifies key characteristics that influence renewal probabilities. This step clarifies the core factors affecting customers' willingness to renew, breaking the "black box" dilemma of traditional prediction models. It enables sales personnel to clearly understand the generation logic of renewal probabilities, avoiding the problem of blindly formulating retention strategies, and providing a clear and implementable decision-making basis for the generation of subsequent personalized retention strategies.

[0026] This step generates personalized retention strategies for target customers based on key characteristics. It abandons the traditional standardized, one-size-fits-all retention model and formulates retention strategies that are adapted to individual needs based on the key influencing characteristics of customers. This significantly improves the targeting and effectiveness of the strategies, can accurately respond to the core needs of customers, thereby increasing customer acceptance and effectively reducing the risk of customer churn.

[0027] The invention of this application achieves comprehensiveness of data dimensions, real-time status of profile, accuracy of prediction results, clarity of influencing factors, and personalization of retention strategies through a complete process of "data collection - profile construction - probability prediction - feature analysis - strategy generation". Each step is progressive and mutually supportive, comprehensively improving the scientificity and effectiveness of policy renewal prediction and customer retention work, and effectively solving the core problems existing in traditional related technologies.

[0028] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 The flowchart of a method for determining a personalized customer retention strategy provided in Embodiment 1 of this application is shown; Figure 2 The flowchart of a dynamic customer profile construction method provided in Embodiment 1 of this application is shown; Figure 3 A flowchart of a method for determining the probability of policy renewal provided in Embodiment 1 of this application is shown; Figure 4 A flowchart of a key feature determination method provided in Embodiment 1 of this application is shown; Figure 5 The flowchart of a personalized retention strategy generation method provided in Embodiment 1 of this application is shown; Figure 6 A flowchart of a closed-loop optimization method provided in Embodiment 1 of this application is shown; Figure 7 This paper illustrates the architecture of a personalized customer retention strategy technology process provided in Embodiment 1 of this application; Figure 8 This illustration shows a schematic diagram of a customer personalized retention strategy determination device provided in Embodiment 2 of this application; Figure 9 A schematic diagram of the structure of a computer device provided in Embodiment 3 of this application is shown. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0032] Example 1 To facilitate understanding of this application, the following is combined with... Figure 1 The flowchart illustrating a method for determining a personalized customer retention strategy provided in Embodiment 1 of this application provides a detailed description of Embodiment 1 of this application.

[0033] Referring to Figure 1, Figure 1 shows a flowchart of a method for determining a personalized customer retention strategy provided in Embodiment 1 of this application, wherein the method includes steps S101 to S103: S101: Collect multimodal data from target customers. Specifically, the collected multimodal data includes four main categories, all acquired in real-time or near real-time to ensure accurate reflection of the customer's latest status and environmental dynamics. Static data originates from the insurance company's core business system, covering basic customer attributes (age, gender, occupation, geographic information), basic policy information (type of insurance, insurance period, premium amount, payment method, etc.), and historical claims records (number of claims in the past 1-3 years, total compensation amount, date of the last claim, etc.). Dynamic behavioral data comes from user interaction channels such as apps, CRM systems, and the insurance company's official website, recording customer behavioral event sequences on a timeline, such as app login frequency, policy details viewing operations, health task completion status, and marketing activity participation. Unstructured text data comes from call center recordings (processed using speech-to-text technology), online customer service chat logs, ticket systems, and other scenarios, requiring the extraction of sentiment scores and topic distributions using natural language processing technology.

[0034] Sentiment analysis employs a customized model fine-tuned for the insurance field based on general pre-trained models such as BERT (Bidirectional Encoder Representations from Transformers). Because insurance customer service dialogues contain professional terms such as premium waiver, cash value, and deductible, general models struggle to accurately understand the context. Sentiment scores are quantified using a discrete five-category classification (very positive, positive, neutral, negative, very negative) + [-1, 1] continuous values, where -1 represents extreme negative and 1 represents extreme positive. Macroeconomic data includes macroeconomic indices (such as the consumer confidence index), industry competition dynamics (such as competitor premium adjustments and the launch of new insurance products), and seasonal factors (such as peak / off-peak seasons for auto insurance), supporting cross-scenario decision analysis.

[0035] S102: Construct and update dynamic customer profiles of target customers based on the aforementioned multimodal data. Specifically, the construction of dynamic customer profiles relies on a multimodal data fusion module, employing a hierarchical fusion strategy to break down heterogeneous data barriers. First, feature-level fusion is achieved through a parallel feature extraction sub-network: static data is processed using an embedding layer, text data is processed using a BERT (Bidirectional Encoder Representations from Transformers) model fine-tuned for the insurance domain, and time-series data is processed using a GRU (Gated Recurrent Unit) layer. The four types of data are mapped to feature vectors of a unified dimension (e.g., 64 dimensions) and then concatenated.

[0036] The fused features are then input into a prediction model incorporating attention mechanisms and multilayer perceptrons. Through model training, the model automatically learns the nonlinear interactions and semantic alignment between features of different modalities, achieving deep fusion at the model level. Ultimately, a unique profile is generated for each customer, stored in a dynamic profile library rather than a simple collection of static labels. The profile update mechanism is divided into two categories: an event-triggered update mechanism that instantly locates and updates relevant dimensions of the profile when key business events such as customer complaints, claims settlements, policy information changes, or participation in / abandonment of marketing activities occur, ensuring that the profile reflects key changes in the customer's status in real time; and a periodic incremental update mechanism that performs sliding window calculations on time-series data such as dynamic behavior sequences and sentiment curves at fixed intervals (e.g., daily). The window parameters (size, step size) are initialized based on business experience (e.g., a 30-day window + a 1-day step size) and can be adaptively adjusted according to the density of customer events or the policy lifecycle stage (e.g., switching to a 7-day window for complaints to improve the sensitivity of capturing negative states). By incrementally calculating and updating trend indicators, the profile ensures accurate capture of the evolution trend of customer behavior.

[0037] S103: Extract customer features from the dynamic customer profile and input these features into the dual-channel renewal prediction model to obtain the policy renewal probability. Specifically, customer features are extracted from the four dimensions of the dynamic customer profile to achieve a multi-dimensional and comprehensive representation of customer status. The customer attribute dimension extracts static attribute features and policy features: Static attribute features are obtained by processing customer demographic attributes, such as age segmentation (18-25, 26-35, 36-45, 46-55, 56 and above), occupation categorization (company employee, civil servant, freelancer, student, retiree), and region grading (first-tier city, new first-tier city, second-tier city, third- and fourth-tier city, other regions), and also includes whether family members have purchased the company's products, etc. The features include characteristics derived from the company's VIP clients; policy features focus on attributes strongly correlated with renewal intentions, such as the type of insurance being divided into car insurance, health insurance, life insurance, and property insurance categories; the insurance period being divided into 1-year, 2-5-year, and over 5-year ranges; the premium amount being divided into the ranges below 5,000 yuan, 5,000-10,000 yuan, 10,000-20,000 yuan, and over 20,000 yuan; and derivative features such as the policy's effective period and whether additional value-added services (such as free roadside assistance and green channels for medical treatment) are included.

[0038] Statistical features, trend features, and sequence pattern features are extracted from the behavioral trend dimension: Statistical features are quantitative summaries of customer behavior within fixed time windows (7 days, 15 days, 30 days), such as the number of times the APP has logged in in the past 7 days and the number of times policy details have been viewed in the past 15 days; Trend features are calculated by comparing statistical features in different time windows, such as the week-on-week change rate of login frequency, calculated as (current window statistical value - historical window statistical value) / historical window statistical value × 100%; Sequence pattern features are identified by frequent sequence pattern mining algorithms such as PrefixSpan (PrefixSpan Algorithm), such as repeated behavior combinations like APP login → viewing policy → viewing claim progress → logging out.

[0039] The sentiment tendency dimension extracts sentiment score features and attention topic features: Sentiment score features are calculated based on sentiment analysis results, using multi-dimensional statistical indicators such as the average sentiment score of the last 10 interactions and the number of negative sentiments in the last 30 days; Attention topic features are obtained by mining themes in text interaction content using LDA (Latent Dirichlet Allocation) to statistically analyze the frequency and percentage of customers mentioning preset business themes such as claims efficiency and premium discounts. The environmental association dimension extracts environmental association features, such as the average premium price of the same insurance product from competitors in the customer's region and the local auto insurance claim rate, to achieve accurate quantification of the impact of the external environment.

[0040] The dual-channel renewal prediction model employs a heterogeneous feature separation processing architecture to maximize the representational capabilities of different feature types. Static attribute features, policy features, and environmental correlation features are input into the static feature channel and, after preprocessing, are fed into a fully connected network. Categorical features are processed using one-hot encoding or embedding encoding (mapping high-dimensional categorical features into low-dimensional dense vectors), while numerical features are processed using Z-Score (Z-Score Standardization) to ensure that the feature mean is 0 and the standard deviation is 1, avoiding interference from features of different magnitudes during model training. The fully connected network contains 2-3 hidden layers, using ReLU or LeakyReLU activation functions to extract the correlation between features through nonlinear transformations, outputting a 32-dimensional or 64-dimensional static feature vector.

[0041] Statistical features, trend features, sequence pattern features, sentiment score features, and focus topic features are input into the dynamic time-series channel. First, a time-series feature sequence of nearly 30 days is constructed with daily time steps. Then, it is input into GRU (Gated Recurrent Unit) or LSTM (Long Short-Term Memory) to model the temporal dependencies. GRU controls information forgetting and updating through update and reset gates, while LSTM captures long-term dependencies through input, forget, and output gates. Both can effectively solve the gradient vanishing problem of traditional recurrent neural networks. On this basis, an attention mechanism is introduced to calculate the similarity (dot product or cosine similarity) between the hidden state vector and the query vector (the hidden state vector of the last time step of GRU or LSTM) at each time step. The attention weights are obtained by normalization through the Softmax function. Higher weights are assigned to features at key time steps (such as 7 days before the policy expires). After weighted summation, a dynamic time-series feature vector with the same dimension as the static feature vector is output.

[0042] Finally, the two types of feature vectors are fused by vector concatenation, and the probability of renewal in the range of 0-1 is calculated by a fully connected layer (containing one hidden layer and one output layer) and the Softmax function. The closer the probability is to 1, the higher the likelihood of the customer renewing the policy.

[0043] S104: Perform interpretability analysis on the policy renewal probability to identify key features affecting the renewal probability. Specifically, interpretability analysis is achieved through SHAP or LIME techniques. The core objective is to break the black box of model decision-making and establish a clear mapping between prediction results and business characteristics. When applying SHAP (SHapley Additive exPlanations) technology, based on the Shapley value principle in game theory, each input feature is regarded as a game participant, and the renewal probability is regarded as a cooperative payoff. By calculating the average contribution of each feature across all possible feature subsets, its final contribution to the probability is determined. During the reverse analysis, a SHAP value map is generated, showing the SHAP value (contribution) of the feature, the original value, and the direction of influence (positive values ​​increase probability, negative values ​​decrease probability). It can also present the interaction between features (such as the synergistic contribution of the age segment 26-35 years old and those with high login frequency in the past 7 days).

[0044] For abstract features of dynamic time-series channels (such as behavioral sequence vectors), key time windows (such as 7 days before policy expiration) are first located using attention weights. Then, SHAP values ​​are calculated for the original business features within the window (such as daily login count and daily sentiment score), transforming the importance of the abstract vectors into business-understandable dimensional information such as a 70% month-on-month decrease in login frequency. When applying LIME (Local Interpretable Model-agnostic Explanations) technology, based on the principle of local approximation, a large number of similar samples are generated in the neighborhood of the current customer feature samples (fine-tuning the login count and sentiment score in the past 7 days). The prediction results of similar samples are fitted with a simple linear model, and the contribution of each feature in the original model is approximated by the linear model coefficients. Inverse analysis generates a local interpretation map, showing the feature coefficients, values, and the linear model fitting effect. For abstract time-series features, the focus is also on the original business features of the key window, ensuring that the interpretation results fit the actual business scenario.

[0045] The contribution of each input feature is calculated based on the results of reverse analysis: In SHAP technology, the contribution is the SHAP value, which typically ranges from -1 to 1. It focuses on calculating the original business features within the key time window, avoiding the incomprehensibility of abstract vectors; In LIME technology, the contribution is the coefficient of the local linear model, obtained through least squares fitting, and the influence strength needs to be judged in conjunction with feature standardization. In practical applications, both the original contribution value (absolute influence size) and the standardized contribution value (the proportion of the sum of the absolute values ​​of all feature contributions, reflecting relative importance) are output simultaneously.

[0046] Finally, key features are selected based on preset thresholds: the thresholds are determined using statistical methods (set to 1.5 or 2 times the standard deviation) or business experience methods (set to an absolute contribution value ≥ 0.15). Features with an absolute contribution value exceeding the threshold are identified as key features, and a key impact factor report is generated. The report is categorized and arranged into positive and negative key features (positive features are sorted in descending order of contribution, and negative features are sorted in descending order of absolute contribution value), and the intervention priority of each feature is marked. It includes basic feature information (name, source dimension, original value), contribution information (original value, standardized value, direction of impact), and business interpretation (e.g., for positive features, contribution value 0.18, high number of logins in the past 7 days significantly increases the probability of renewal, it is recommended to maintain customer APP activity).

[0047] S105: Based on the aforementioned key features, generate a personalized retention strategy for the target customer. Specifically, strategy generation relies on a pre-defined feature-strategy mapping rule library. This rule library is developed by insurance company business experts in conjunction with data analysts, based on historical intervention effect data and business experience to ensure the effectiveness and operability of the rules. The basic rule format is IF key feature condition THEN retention intervention action. The key feature condition can be a single condition (e.g., a decrease of more than 70% in APP login frequency in the past month) or a combination of conditions (e.g., a recent emotional score ≤ -0.6 AND ≥ 2 customer service inquiries in the past 30 days). The retention intervention action clearly defines the executing entity (e.g., APP operation team, customer service center), the execution channel (e.g., APP push, human telephone), the execution timing (e.g., 21 days before the policy expires), and the execution content (e.g., a 50 yuan premium coupon, valid for 7 days).

[0048] The rule base sets rule priorities and conflict resolution mechanisms: priorities are ranked according to intervention cost and effectiveness (e.g., manual follow-up visits have higher priority than app push notifications); when multiple rules match simultaneously, the rule with the highest priority is selected, or the most suitable rule is selected by combining information supplemented by dynamic customer profiles (e.g., frequently used interaction channels) (e.g., if a customer frequently uses SMS, the app push notification will be changed to an SMS push notification). During the matching process, key features (especially negative features) are transformed into feature condition statements and accurately matched with the conditions in the rule base.

[0049] If only one rule is matched, the intervention action according to that rule is directly adopted; if multiple non-conflicting rules are matched, all intervention actions are combined to form a combined strategy (e.g., pushing a 50 yuan coupon to the app + sending a health education SMS once a week). Subsequently, the intervention actions are fine-tuned based on the customer's dynamic profile: the execution channel is adapted to the customer's commonly used interaction methods, the execution timing is adapted to the customer's behavioral habits (e.g., if the customer logs into the app on the weekend, the push time is adjusted), and the execution content details are adapted to the customer's core concerns (e.g., if the customer frequently mentions the efficiency of claims processing, a description of the priority review service for claims processing is added).

[0050] The final personalized retention strategy suggestion document includes strategy objectives (such as increasing customer app activity and reducing churn risk), a list of intervention actions (detailing the implementing entity, channel, time, and content of each action), execution requirements (such as tracking whether SMS messages have been read within 24 hours of being sent), expected results (estimated renewal probability based on historical data, such as 8%-12%), and risk warnings (such as triggering a second SMS reminder if the customer does not respond to app push notifications within 3 days), ensuring that the strategy can be directly implemented.

[0051] In one optional implementation, the multimodal data includes static attribute data, time-series behavioral log data, unstructured text data, and macro-environmental data.

[0052] Specifically, the four types of data complement each other in terms of data source, type, and function, together forming the complete data foundation for dynamic customer profiles. Static attribute data is structured data, sourced from the insurance company's core business system, recording customer demographic attributes (age, gender, occupation, region) and basic policy attributes (insurance type, insurance period, premium amount, sum insured), providing fixed core information about customers and policies.

[0053] The time-series behavioral log data is semi-structured data, sourced from user interaction channels such as apps, CRM systems, and official websites. It records customer actions in chronological order (e.g., app login on May 1, 2024; viewing car insurance policies on May 3, 2024) to capture real-time customer behavior tendencies and pattern changes. The unstructured text data is in natural language form, sourced from call center recordings transcribed into text, online customer service chat logs, and work order descriptions. It requires the extraction of sentiment and thematic information using natural language processing technology. Sentiment information is extracted based on a BERT fine-tuning model, using discrete classification + continuous value quantization in the range of [-1,1]. Thematic information is mined based on an LDA model to uncover preset business themes such as claims efficiency and premium discounts, reflecting customer sentiment tendencies and core concerns.

[0054] Macroeconomic data is external structured data sourced from government statistical departments, industry research institutions, or third-party data service providers. It includes macroeconomic indicators (GDP growth rate, per capita disposable income of residents), industry dynamic data (information on the launch of new insurance products by competitors, industry average renewal rate), and seasonal and regional data (auto insurance claim rate during the rainy season, auto insurance penetration rate in first-tier cities), supporting quantitative analysis of the impact of the external environment on customers' willingness to renew their insurance policies.

[0055] Referring to Figure 2, which shows a flowchart of a dynamic customer profile construction method provided in Embodiment 1 of this application, the step of constructing and updating the dynamic customer profile of the target customer based on the multimodal data includes steps S201-S203: S201: The static attribute data, the time-series behavior log data, the unstructured text data, and the macro-environment data are subjected to multimodal data fusion processing to construct the dynamic customer profile. The dynamic customer profile includes customer attribute dimensions, behavioral trend dimensions, sentiment tendency dimensions, and environmental association dimensions. Specifically, the multimodal data fusion processing follows a standardized process of association-cleaning-integration to ensure data quality and integration effectiveness. In the data association stage, the customer's unique identifier (such as customer ID) is used as the core key to accurately associate static attribute data (associating customer ID with policy ID), time-series behavior log data (associating customer ID with device ID / account ID), and unstructured text data (associating customer ID with interaction record ID) with the same customer. Simultaneously, auxiliary keys such as timestamps and geographic codes are used to match macro-environment data with customer data; for example, the penetration rate of competing auto insurance products in a certain city in Q2 2024 is associated with customers in that city.

[0056] The data cleaning process involves targeted processing based on the characteristics of different data types: for static attribute data, it corrects missing occupational classifications and incorrect regional codes; for time-series behavioral log data, it removes duplicate login logs and abnormal high-frequency operation records (such as viewing the policy 100 times in a short period of time); for unstructured text data, it filters out invalid text (such as noise characters in audio-to-text conversion) and standardizes expression (such as grouping slow claims processing and poor claims processing speed into the same topic).

[0057] The data integration process divides and integrates the cleaned data according to dimensions: the customer attribute dimension integrates static attribute data and core policy information to form a basic customer profile; the behavior trend dimension integrates statistical and trend information from time-series behavior log data to form a customer behavior trajectory; the sentiment tendency dimension integrates sentiment and thematic information extracted from text data to form a customer psychological state; and the environmental correlation dimension integrates macro-environment matching information to form the customer's external context, ultimately constructing a multi-dimensional, dynamically updatable customer profile.

[0058] S202: The dynamic customer profile is continuously evolved using an event-triggered mechanism and a periodic incremental update mechanism, wherein: the event-triggered mechanism immediately updates the corresponding dimension of the dynamic customer profile when a specific business event occurs; the periodic incremental update mechanism performs sliding window calculations on the time-series behavior log data and the sentiment sequence data extracted from unstructured text data according to a fixed time window, so as to update the trend indicators in the dynamic customer profile.

[0059] Specifically, the event triggering mechanism and the periodic incremental update mechanism work together to adapt to the update needs of sudden critical changes and gradual trend changes, respectively. Specific business events in the event triggering mechanism must meet conditions that significantly impact customers' willingness to renew their policies. These include customers initiating claims, claims settlement, customer service complaints, policy information changes (such as sum assured adjustments), and participation in / abandonment of marketing activities. When such events occur, the system captures event information in real time through the event notification interface of the business system, locates the event-related dimensions in the customer profile within 5 minutes, and completes the update. For example, claims settlement events update the customer attribute dimension—historical claims records—and the behavioral trend dimension—the frequency of claims-related operations; customer service complaint events update the sentiment tendency dimension—recent sentiment score—and the sentiment tendency dimension—complaint topic records.

[0060] The sliding window parameters of the periodic incremental update mechanism are initialized based on prior knowledge of the business cycle (e.g., a 30-day window + a 1-day step), and can be adaptively adjusted according to customer event density or policy lifecycle stage. For time-series behavioral log data, behavioral statistics (e.g., number of APP logins in the past 7 days) and trend indicators (e.g., the month-on-month change rate of logins in the past 7 days and the past 14 days) are calculated within the window. For sentiment sequence data (the raw time-series data of sentiment curves), sentiment statistics (e.g., the average sentiment score in the past 7 days) and topic frequency indicators (e.g., the number of times premiums were mentioned in the past 7 days) are calculated within the window. After the calculation is completed, only the trend dimension fields related to these indicators in the profile are updated, rather than a full update, which reduces system resource consumption while ensuring the timeliness of the profile.

[0061] In an optional implementation, static attribute features and policy features are extracted from the customer attribute dimension of the dynamic customer profile.

[0062] Specifically, the extraction of static attribute features and policy features follows the dual principles of stability and business relevance to ensure that the features accurately reflect the core static information of customers and policies. Static attribute features are extracted from customer demographic attributes: age is divided into segmented features by 18-25 years old, 26-35 years old, 36-45 years old, 46-55 years old, and 56 years old and above; occupation is divided into categorized features by corporate employees, civil servants, freelancers, students, and retirees; and region is divided into level features by first-tier cities, new first-tier cities, second-tier cities, third- and fourth-tier cities, and other regions. At the same time, features such as whether family members have purchased the company's products and whether the customer is a VIP customer are derived to further enrich the static dimension representation of customers.

[0063] Policy features are extracted from the basic policy information under the customer's name, focusing on attributes strongly related to renewal intention: insurance type is divided into categories such as car insurance, health insurance, life insurance, and property insurance; insurance period is divided into ranges such as 1 year, 2-5 years, and more than 5 years; premium amount is divided into ranges such as below 5,000 yuan, 5,000-10,000 yuan, 10,000-20,000 yuan, and more than 20,000 yuan; and payment method is divided into types such as annual payment, monthly payment, and quarterly payment. At the same time, features such as policy effective duration and whether value-added services are attached (such as free roadside assistance and green channel for medical treatment) are derived, providing comprehensive policy-related inputs for the static feature channel.

[0064] From the behavioral trend dimension of the dynamic customer profile, statistical features, trend features, and sequence pattern features of the dynamic behavior sequence are extracted.

[0065] Specifically, the three types of features depict customer behavior in a layered manner from three dimensions: quantitative statistics, trend changes, and pattern regularity, ensuring a comprehensive mining of the value of dynamic behavior. Statistical features are a quantitative summary of customer behavior within a fixed time window, typically set at 7 days, 15 days, or 30 days. These include the number of times the app has logged in in the past 7 days, the number of times policy details have been viewed in the past 15 days, the number of times health tasks have been completed in the past 30 days, and the number of times marketing activities have been clicked in the past 7 days, directly reflecting the customer's activity level within a specific period.

[0066] The trend feature is a comparative analysis of statistical features within different time windows, used to capture the direction and magnitude of changes in customer behavior. Specifically, it includes the month-on-month change rate of the number of APP logins in the past 7 days compared to the past 14 days, and the year-on-year change rate of the number of policy views in the past 15 days compared to the same period last month. The calculation method is (current window statistical value - historical window statistical value) / historical window statistical value × 100%. A positive feature value indicates an increase in behavioral activity, while a negative value indicates a decrease.

[0067] Sequence pattern features are the patterns mined from sequences of customer behavioral events. Frequent sequence pattern mining algorithms such as PrefixSpan are used to identify recurring combinations of customer behaviors. Specifically, this includes whether the sequence of logging into the APP → viewing the policy → viewing the claim progress → exiting is frequent, and whether the sequence of receiving marketing text messages → clicking links → not placing an order is frequent. These features are used to reflect customer behavior habits and potential needs (e.g., frequently checking the claim progress may indicate concern about claim efficiency).

[0068] From the emotional tendency dimension of the dynamic customer profile, extract emotional score features and interest topic features.

[0069] Specifically, both types of features are extracted from the results of natural language processing of unstructured text data to achieve a quantitative representation of the customer's emotional state and core concerns. The emotional score feature first uses a BERT emotional classification model fine-tuned for the insurance field to score the emotional state of each customer's text interaction (such as customer service chat, work order description), with a score range of [-1, 1] (-1 is extremely negative, 1 is extremely positive); then, based on the scoring results, multi-dimensional statistical indicators are calculated to form the emotional score feature, which specifically includes the average emotional score of the last 10 interactions, the lowest emotional score of the last 10 interactions, the number of negative emotions (score ≤ -0.6) in the last 30 days, and the trend of emotional score changes in the last 7 days (rising / falling / stable), directly reflecting the customer's emotional state and emotional change trend.

[0070] By leveraging LDA to mine customer text interactions, we pre-define business-related topics such as claims efficiency, premium discounts, service quality, product coverage, and app user experience. Then, we analyze the frequency and percentage of customers mentioning each topic within a fixed time window to identify key themes. These themes include the number of times claims efficiency was mentioned in the past 30 days, the frequency and percentage of premium discounts mentioned in the past 15 days (number of premium discount mentions / total number of topic mentions), and newly added topics in the past 7 days. This helps us understand customers' current core concerns and changing needs.

[0071] Environmental association features are extracted from the environmental association dimension of the dynamic customer profile.

[0072] Specifically, environmental correlation features are the product of precise matching between external environmental data and customer information. The core objective is to quantify the impact of the external environment on the renewal intentions of specific customers. For macroeconomic data, matching and conversion are performed according to the customer dimension: the national consumer confidence index is weighted and adjusted according to the economic level of the customer's region to obtain the regional consumer confidence index; the national average renewal rate of the insurance industry is broken down according to the type of insurance purchased by the customer to obtain the average renewal rate of the insurance industry to which the customer belongs; and the year-on-year growth rate of per capita disposable income is correlated with the customer's premium amount range to obtain the matching value of income growth rate for customers within the same premium range.

[0073] Based on industry competition dynamics data, we conduct precise matching at the customer level: according to the customer's region and the type of insurance, we extract characteristics such as the average premium of the same type of insurance offered by competitors in the customer's region, the discounts for new customers of the same type of insurance offered by competitors in the customer's region, and the renewal rate of the same type of insurance offered by competitors in the customer's region; according to the customer's policy expiration date, we extract characteristics such as the number of new insurance products launched in the industry in the month the policy expires, and the frequency of competitor marketing activities in the 30 days before the policy expires.

[0074] For seasonal and regional data, customer-dimensional adaptation is performed: for auto insurance customers, features such as the current quarterly auto insurance claim rate (reflecting seasonal risk) and the duration of the rainy / snowy season in the customer's region (affecting auto insurance demand) are extracted; for health insurance customers, features such as the current influenza incidence rate in the customer's region (affecting health insurance awareness) and the tightness of medical resources in the customer's region (affecting the expected claims experience) are extracted to ensure that environmental correlation features can accurately reflect the impact of the external environment on specific customers.

[0075] In an optional implementation, see Figure 3 As shown, Figure 3 The flowchart illustrates a method for determining the policy renewal probability provided in Embodiment 1 of this application, wherein the step of inputting the customer characteristics into a dual-channel renewal prediction model to obtain the policy renewal probability includes steps S301 to S303: S301: Input the static attribute features, the policy features, and the environmental association features into the static feature channel of the dual-channel renewal prediction model, and process them through a fully connected network to obtain a static feature vector.

[0076] Specifically, the core function of the static feature channel is to achieve non-linear fusion and key information extraction of static class features. First, the input static features are preprocessed: categorical features (such as occupational classification and insurance type) are encoded using one-hot encoding or embedding encoding. Embedding encoding can map high-dimensional categorical features into low-dimensional dense vectors, reducing computational complexity and preserving category association information; numerical features (such as age segmentation encoding values ​​and industry average renewal rates) are processed using Z-Score to eliminate the interference of features of different magnitudes on model training and ensure the comparability between features.

[0077] The preprocessed features are input into a fully connected network, which typically contains 2-3 hidden layers. Each hidden layer uses either ReLU or Leaky ReLU activation functions—ReLU effectively alleviates the vanishing gradient problem, while Leaky ReLU preserves negative gradient information, improving the model's representational ability. The first hidden layer maps the preprocessed features from their original dimension (e.g., 50-dimensional) to a higher dimension (e.g., 128-dimensional), capturing the basic relationships between features. The second hidden layer maps the 128-dimensional features to 64-dimensional features, further refining key information. Finally, the output layer of the fully connected network (without an activation function) outputs a 32-dimensional or 64-dimensional static feature vector. This vector integrates core information from the customer's static attributes, policy information, and the external environment, laying the foundation for subsequent fusion with dynamic temporal features.

[0078] S302: Input the statistical features, trend features, sequence pattern features, sentiment score features, and topic of concern features into the dynamic time-series channel of the dual-channel insurance renewal prediction model, model the time-series dependencies through a gated recurrent unit or a long short-term memory network, and use an attention mechanism to assign weights to features at different time steps to obtain a dynamic time-series feature vector.

[0079] Specifically, the core function of the dynamic time-series channel is to capture the temporal dependence of customer dynamic behavior and emotions, and highlight the characteristic value of key time steps. First, the five types of input dynamic features are arranged according to time steps to form a time-series feature sequence—using days as time steps, a time-series sequence of nearly 30 days is constructed. Each time step includes features such as the number of logins on that day (statistical feature), the month-on-month change in the number of logins (trend feature), behavioral sequence pattern tags (sequence pattern feature), the average sentiment score on that day (sentiment score feature), and the frequency of topic mentions on that day (theme of interest feature).

[0080] The temporal feature sequence is then input into GRU or LSTM. Both networks can effectively handle the long-distance dependency problem in temporal data: GRU adaptively forgets irrelevant information and updates useful information through the synergy of update and reset gates; LSTM achieves accurate memorization of long-term information through the cooperation of cell state and three gate structures, avoiding the gradient vanishing problem of traditional recurrent neural networks. After network processing, the hidden state vector (e.g., 64-dimensional) of each time step is output, reflecting the encoded information of the dynamic features of that time step.

[0081] To further enhance the representation capability of dynamic features, an attention mechanism is introduced to assign weights to the hidden state vectors at each time step. First, the similarity score is calculated between the hidden state vector at each time step and the query vector (usually the hidden state vector of the last time step in a GRU or LSTM system). Then, the similarity score is normalized using the Softmax function to obtain the attention weight for each time step—a higher weight indicates that the dynamic features at that time step are more critical for renewal prediction; for example, the weight of the hidden state vector 7 days before policy expiration is usually higher than that 30 days before expiration. Finally, the hidden state vector at each time step is multiplied by its corresponding attention weight, and all weighted vectors are summed to obtain a dynamic temporal feature vector (with the same dimensions as the static feature vector, such as 32 or 64 dimensions). This vector highlights the dynamic feature information of key time steps, improving the representation capability of dynamic dimensions.

[0082] S303: The static feature vector and the dynamic time-series feature vector are fused together, and the policy renewal probability is calculated and output through a fully connected layer and the Softmax function.

[0083] Specifically, the feature fusion and probability calculation stage aims to integrate the core information of the two types of features to output accurate renewal probability prediction results. Feature fusion adopts a vector concatenation method, merging static feature vectors and dynamic time-series feature vectors—for example, a 32-dimensional static feature vector and a 32-dimensional dynamic time-series feature vector are concatenated to form a 64-dimensional fused feature vector. This method can retain the original information of the two types of features to the greatest extent and avoid information loss.

[0084] The fused feature vector is input into a fully connected layer, which contains one hidden layer and one output layer. The hidden layer uses the ReLU activation function to map the 64-dimensional fused feature vector to 16-dimensional or 32-dimensional vectors, further exploring the interaction between static and dynamic features (such as the positive impact of the combination of young customers and high login frequency in the past 7 days on policy renewal). The output layer contains two neurons, corresponding to the renewal and non-renewal labels, and is used to output the raw scores of the two labels. Finally, the raw scores of the output layer are input into a Softmax function, which transforms the raw scores into probability values ​​in the range of 0-1, and the sum of the probabilities of the two classes is 1. The probability value corresponding to the renewal label is the final output policy renewal probability, which directly quantifies the customer's likelihood of renewing the policy.

[0085] In an optional implementation, referring to Figure 4, which shows a flowchart of a key feature determination method provided in Embodiment 1 of this application, wherein interpretability analysis is performed on the policy renewal probability to determine the key features affecting the renewal probability, including steps S401~S403: S401: Decision attribution is performed on the policy renewal probability, and SHAP or LIME interpretability analysis technology is applied to perform reverse analysis on the model decision process that generates the probability.

[0086] Specifically, the two interpretability techniques achieve decision attribution from different perspectives, both effectively breaking the black box of the model. When applying SHAP technology, the core is based on the Shapley value principle—Shapley values ​​are originally used in game theory to allocate cooperative payoffs. In model interpretation, each input feature is regarded as a game participant, and the renewal probability is regarded as a cooperative payoff. By calculating the average contribution of each feature across all possible feature subsets, the final contribution of that feature to the probability is determined. During the reverse analysis process, SHAP generates a SHAP value map, clearly showing the SHAP value (i.e., contribution) of each feature, the original value of the feature, and the direction of the feature's influence on the probability (positive SHAP value increases the probability, negative SHAP value decreases the probability). At the same time, it can also present the interaction between features (such as the additional contribution of the interaction between the age segment 26-35 years old and the high number of logins in the past 7 days to the probability).

[0087] For abstract features of dynamic time-series channels (such as behavioral sequence vectors), SHAP analysis first uses attention weights to locate key time windows affecting prediction results. Then, it uses the original business features within that window (such as daily login counts and daily sentiment scores) as the analysis objects, outputting their SHAP values ​​to trace from abstract vectors to specific business features. When applying LIME technology, the core principle is based on local approximation. Because the dual-channel model is a complex nonlinear model, direct interpretation is difficult. LIME generates a large number of similar samples in the local neighborhood of the current customer feature samples (such as fine-tuning the login counts of the past 7 days, sentiment scores, etc.), and then uses a simple linear model to fit the model prediction results of these similar samples. The coefficients of the linear model approximate the contribution of each feature in the original model. During the reverse analysis, LIME generates a local interpretation map, showing the coefficients (i.e., approximate contribution), feature values, and the fitting effect of the linear model for each feature, helping business personnel understand how each feature affects the prediction results in the current customer's feature scenario. Similarly, for abstract time-series features, LIME uses the original business features within the key time window as the fine-tuning objects to ensure that the interpretation results correspond to specific business scenarios.

[0088] S402: Based on the reverse analysis, calculate the contribution of each input feature of the dual-channel renewal prediction model to the policy renewal probability prediction result.

[0089] Specifically, contribution calculation needs to be combined with reverse analysis results to achieve precise quantification of the impact on each feature, providing a basis for subsequent key feature selection. For SHAP technology, contribution is the SHAP value of a feature. The calculation process needs to traverse all possible feature subsets: for example, for feature A, the difference between the model prediction value of all feature subsets containing A and the prediction value of the corresponding subsets not containing A needs to be calculated, and then all differences are averaged to obtain the SHAP value (contribution) of A. The range of SHAP value is related to the model output, usually between -1 and 1. The larger the value, the stronger the positive contribution to the renewal probability, and the smaller the value, the stronger the negative contribution. For dynamic time-series features, SHAP value calculation will focus on the original business features within the key time window, such as the SHAP value of the number of logins in the past 7 days, rather than the SHAP value of the abstract behavioral sequence vector, to ensure that the contribution has business interpretability.

[0090] For LIME technology, contribution is the feature coefficient of the local linear model. The calculation process involves fitting a linear relationship between the prediction results of locally similar samples and the feature values ​​using the least squares method to obtain the coefficient (contribution) of each feature. The range of coefficient values ​​is not fixed; their influence strength must be judged in conjunction with the feature's standardization. For example, a coefficient of 0.2 for the number of logins in the past 7 days indicates that within the local neighborhood, for every 1 standardization unit increase in this feature, the renewal probability increases by an average of 0.2. In practical applications, both the original contribution value and the standardized contribution value for each feature are output simultaneously: the original contribution value reflects the absolute magnitude of the feature's influence, while the standardized contribution value (the sum of the absolute values ​​of all feature contributions, then calculating the proportion of each feature's contribution) reflects the relative influence weight of the feature, allowing business personnel to comprehensively understand the feature's role from both absolute and relative perspectives.

[0091] S403: Based on a preset contribution threshold, input features whose absolute contribution value exceeds the threshold are identified as key features, and a key influence factor report containing each key feature and its corresponding contribution value is generated.

[0092] Specifically, the key feature selection and report generation must balance technical rationality and business practicality to ensure that the selection results directly support business decisions. First, a contribution threshold is determined, using either a statistical method or a business experience method. The statistical method calculates the standard deviation of the contribution of all sample features, setting the threshold to 1.5 or 2 times the standard deviation to ensure that key features have a significant impact on the prediction results and eliminate the interference of random fluctuations. The business experience method is set by business experts based on historical intervention effects; for example, setting the threshold to an absolute contribution value ≥ 0.15 to ensure that the selected key features correspond to specific business intervention actions and have practical operational value. After determining the threshold, the contribution of all input features is iterated, and features with absolute values ​​exceeding the threshold are marked as key features.

[0093] A key impact factor report is then generated, which must include three core modules: basic feature information, contribution information, and business interpretation. The basic feature information includes the feature name (e.g., number of APP logins in the past 7 days), the feature source dimension (e.g., behavioral trend dimension), and the original feature value (e.g., 5 times). The contribution information includes the original contribution value (e.g., 0.18), the standardized contribution value (e.g., 15%), and the direction of influence (e.g., positive). The business interpretation section is automatically generated by the system based on preset rules. For example, if the feature is a positive key feature with an original contribution of 0.18, it indicates that a higher number of logins in the past 7 days has a significant effect on improving the probability of policy renewal, and it is recommended to maintain the customer's APP usage activity. The report will also be categorized and arranged into positive and negative key features. Positive key features will be sorted from largest to smallest by their contribution (highlighting factors that are beneficial to renewal), while negative key features will be sorted from largest to smallest by their absolute contribution (highlighting churn risk factors that require key intervention). At the same time, the intervention priority of each key feature will be marked (e.g., a negative key feature with a recent sentiment score of -0.8 has a higher priority than a 10% decrease in login frequency in the past 30 days), providing a clear priority basis for the generation of subsequent retention strategies.

[0094] In an optional implementation, see Figure 5 As shown, Figure 5 The flowchart illustrates a personalized retention strategy generation method provided in Embodiment 1 of this application, wherein the step of generating a personalized retention strategy for the target customer based on the key features includes steps S501-S502: S501: Match the key features with a preset feature strategy mapping rule library, wherein the feature strategy mapping rule library predefines the mapping relationship between key feature combinations and retention intervention actions.

[0095] Specifically, the feature-strategy mapping rule base is the core bridge connecting data features and business actions. Its construction and matching logic directly determine the effectiveness of the strategy. The rule base is jointly developed by insurance company business experts and data analysts. The development process must fully integrate historical intervention effect data and business experience to ensure the scientific nature and operability of the rules. The basic format of the rules is IF key feature condition THEN retention intervention action. The key feature condition can include a single feature condition (such as a decrease of more than 70% in the APP login frequency in the past month) or a combination of feature conditions (such as a recent emotional score ≤ -0.6 AND the number of customer service inquiries ≥ 2 in the past 30 days). The retention intervention action must clearly define the implementing entity, implementation channel, implementation content, and implementation timing. For example: IF key feature condition = decrease of more than 70% in the APP login frequency in the past month THEN intervention action = the APP operation team pushes a special wake-up message (including a 50 yuan premium coupon, valid for 7 days) through the APP 21 days before the policy expires.

[0096] The rule base also includes rule priority and conflict resolution mechanisms: rule priority is set based on the cost and effectiveness of intervention actions. For example, the rule priority for manual follow-up visits is higher than that for app push notifications (because manual costs are higher, high-risk customers need to be matched first). When multiple rules are matched simultaneously, the conflict resolution mechanism will select the rule with the highest priority, or combine information supplemented by the customer's dynamic profile (such as the customer's frequently used interaction channels) to select the most suitable rule—for example, if a customer is matched with both app push coupon and SMS push coupon rules, and the profile shows that the customer frequently uses app channels, then the app push rule will be selected. During the matching process, the system will first convert the key features (especially negative key features) determined in step S403 into feature condition statements, and then perform precise matching with the key feature conditions in the rule base. For example, the key feature "recent sentiment score -0.7" will be converted into the condition "recent sentiment score ≤ -0.6", and then the rules in the rule base containing this condition will be matched.

[0097] S502: Based on the matching results, generate personalized retention strategy suggestions that include at least one specific intervention action.

[0098] Specifically, strategy recommendations are generated based on matching results, and then fine-tuned using personalized customer information to ensure the accuracy and feasibility of the strategy. First, the final intervention action is selected from the matching results: if only one rule is matched, the intervention action for that rule is directly adopted; if multiple rules are matched without conflict, all intervention actions are combined to form a combined strategy. For example, if both the rules for pushing coupons via the app and sending health insurance knowledge articles are matched, the strategy would be to push a 50 yuan coupon via the app plus a weekly health education SMS message.

[0099] Then, based on the supplementary information from the customer dynamic profile, the intervention actions are fine-tuned. The fine-tuning covers the execution channels, timing, and details of the execution content: the execution channels need to be matched with the customer's commonly used interaction channels. For example, if the customer often uses SMS instead of the APP, then the APP push will be changed to SMS push; the timing needs to be combined with the policy expiration date and the customer's behavioral habits. For example, if the customer usually logs into the APP on weekends, then the push time will be adjusted to weekends; the details of the execution content need to be matched with the customer's concerns. For example, if the customer has frequently mentioned the efficiency of claims processing in the past 30 days, then the explanation of the fast-track claims process will be added to the coupon push content (such as receiving the coupon to enjoy priority claims review service).

[0100] Finally, a personalized retention strategy suggestion document is generated, which includes four core parts: strategy objectives, intervention action list, execution requirements, and expected results. The strategy objectives clearly state the core purpose of the strategy (e.g., increasing customer app activity and reducing churn risk); the intervention action list details the implementing entity, channel, time, and content of each action; the execution requirements clearly state the execution standards for each action (e.g., SMS messages must be tracked for read status within 24 hours of being sent); and the expected results estimate the renewal probability that the strategy can improve based on historical data (e.g., an expected increase of 8%-12%), while also indicating the risk warnings of the strategy (e.g., if the customer does not respond to app push notifications within 3 days, a second SMS reminder must be triggered) to ensure that business personnel can clearly understand and implement the strategy.

[0101] In an optional implementation, see Figure 6 As shown, Figure 6 A flowchart of a closed-loop optimization method provided in Embodiment 1 of this application is shown, wherein the method further includes steps S601-S603: S601: After implementing the personalized retention strategy, collect customer feedback data and renewal result data.

[0102] Specifically, data collection needs to cover three dimensions: strategy execution process, customer response behavior, and final results, ensuring that the data comprehensively and accurately reflects the strategy's effectiveness and provides a reliable basis for subsequent optimization. Customer feedback data mainly records customer response behavior and attitudes to retention strategies, including: strategy execution records (such as whether APP push notifications were successfully delivered and whether manual follow-up calls were answered), collected in real time through business system interfaces (such as APP push data obtained from the APP operation platform and manual follow-up data obtained from the CRM system); customer response behavior (such as whether they clicked on APP push links, whether they claimed coupons, and whether they participated in the follow-up communication); and customer attitude data (such as emotional feedback records during follow-ups, in-APP satisfaction ratings, and proactively provided opinions).

[0103] Renewal results data is the core basis for judging the final effectiveness of the strategy. Specifically, it includes: whether the customer renewed their policy before its expiry date (yes / no tag), the renewal completion time (e.g., 15 days before policy expiry), changes to policy information during renewal (e.g., whether the sum assured was adjusted, whether new services were added), and the reasons for non-renewal (e.g., customer's proactive feedback of choosing a competitor, temporary lack of need, or system inference of no intervention action). This type of data is collected within 7 days of policy expiry. The reasons for non-renewal need to be supplemented and confirmed through customer service follow-ups or data analysis to ensure data accuracy. In addition, strategy execution cost data (e.g., the time cost of manual follow-ups, the cost of coupons) is also collected for subsequent calculation of the strategy's return on investment, providing a cost dimension reference for rule base optimization.

[0104] S602: The customer feedback data and the renewal result data are used as new multimodal data and fed back into the dynamic customer profile update process.

[0105] Specifically, data feedback is key to the continuous evolution of dynamic profiles, ensuring that newly generated data can be promptly integrated into the profile, improving its timeliness and accuracy. First, customer feedback data is categorized and its features extracted, transforming it into structured data usable for profile updates: for example, actions such as clicking app push links or claiming coupons are converted into time-series behavioral log data in the behavioral trend dimension (e.g., the number of app push clicks and coupon claims in the past 7 days); emotional feedback records during follow-ups are converted into emotional score features in the emotional tendency dimension (e.g., an emotional score of 0.8 for the most recent follow-up); and opinions proactively provided by customers (e.g., a desire for more favorable premiums) are converted into focus-related features in the emotional tendency dimension (e.g., mentioning premium discounts once in the past 7 days).

[0106] Then, the renewal results data are structured and integrated into the corresponding dimensions of the customer profile: for example, the result of completing renewal is transformed into historical renewal records in the customer attribute dimension (e.g., three consecutive years of renewal); the result and reason for non-renewal are transformed into churn tags in the customer attribute dimension (e.g., churn in Q2 2024) and churn reason themes in the sentiment dimension (e.g., mentioning competitor discounts once). Finally, updates are performed according to the dynamic profile update mechanism (event-triggered update or periodic incremental update): for example, claiming coupons is an immediate behavior, and the behavior trend dimension is updated immediately through the event-triggered mechanism; the reason for non-renewal is a phased result, and the sentiment dimension is updated through the periodic incremental update mechanism (e.g., weekly), ensuring that new data can be reflected in the customer profile in a timely manner, providing support for other subsequent business scenarios (e.g., potential customer reactivation).

[0107] S603: Using the renewal result data as labels, the dual-channel renewal prediction model is retrained to achieve closed-loop optimization.

[0108] Specifically, model retraining is a core step in improving prediction accuracy. By constructing a complete training data pipeline, it ensures that the model can continuously learn the patterns of business changes, while simultaneously optimizing the feature strategy mapping rule base to achieve collaborative evolution of the model and strategy. First, a retraining dataset is constructed, using customers and policies as units. New data (customer feedback data, renewal result data) is integrated with the original multimodal data to form a complete customer feature sample. The renewal result is used as a label (renewal is marked as 1, non-renewal as 0). Each sample contains a feature set (static features + dynamic features, with dynamic features including sentiment scores, behavioral statistics, etc., calculated based on new feedback data) and a label. The sample size of the retraining dataset typically needs to reach more than 20% of the original training set to ensure representativeness.

[0109] Then, the model retraining process is executed, using incremental training instead of full retraining to reduce system resource consumption: First, the historical training parameters of the dual-channel model are loaded (such as the weights of the fully connected network, the hidden layer parameters of GRU / LSTM, and the weights of the attention mechanism); then, the retraining dataset is divided into a training set and a validation set in a 7:3 ratio, and the model is incrementally trained using the training set (the learning rate is set to 1 / 10 of the initial learning rate to avoid parameter oscillation); the model performance metrics (such as AUC value, accuracy, and recall) are monitored using the validation set, and training is stopped when the metrics do not improve for three consecutive epochs to avoid overfitting.

[0110] Finally, model evaluation and deployment are conducted. The performance of the newly trained model and the original model is compared on a test set (historical data independent of the retraining dataset). If the AUC value of the new model improves by ≥2% (or other preset indicators are met), the new model is deployed and deployed, replacing the original model. If the indicators are not met, the reasons are analyzed (such as insufficient sample size or missing features), the retraining plan is adjusted, and the process is repeated. At the same time, information such as model parameters, performance indicators, and sample distribution during the retraining process is recorded in the model iteration log to provide a traceability basis for subsequent model optimization. In addition, the feature-policy mapping rule base is optimized based on the retraining results: for example, if the data shows that pushing a 100 yuan coupon has a significant effect on improving the renewal rate of customers whose login frequency has decreased by 50%-70% in the past month (improving by 15%), the priority of this rule is increased; if manual follow-up is ineffective for customers with a sentiment score ≤-0.6 in the last time (improving by only 3%), the intervention action is adjusted (such as changing to a combination of senior customer service manager + exclusive claims commitment) to achieve synchronous optimization of strategy and model.

[0111] See Figure 7 As shown, Figure 7 This paper illustrates the architecture of a personalized customer retention strategy technology process provided in Embodiment 1 of this application, which is divided into three core levels: Data Acquisition and Processing Layer: This layer contains four types of data ("Static Data: Core Business Data", "Dynamic Behavioral Data: APP / CRM / Website, etc.", "Text Interaction Data: Work Order System / Call Center", and "External Environment Data: Macro / Industry Data"). These data enter the "Multimodal Data Fusion (including feature-level stitching and model-level deep alignment)" stage in the form of "raw data streams". After "generation and updating (including event triggering and adaptive sliding window periodic updates)", a "Dynamic Profile Library (continuously updated)" is formed. This layer corresponds to the "Dynamic Profile Building and Update Engine" on the left.

[0112] The core layer of intelligent prediction and decision-making: Based on the "dynamic profile features (including statistical / trend features generated by predefined templates and sequence pattern features automatically learned by the model)", "feature engineering" is carried out to obtain "feature vectors", which are then input into the "renewal prediction model (dual-channel structure, the two channels are processed independently and then spliced ​​and fused)" to output "predicted probabilities". Then, through "interpretability analysis (e.g., SHAP, which supports the traceability from abstract features to the original business dimensions)", "key influencing factors" are determined. After that, the "strategy recommendation engine" generates "push strategies". This layer corresponds to the "intelligent prediction and decision-making core" on the left.

[0113] Business Application and Feedback Layer: The "push strategy" is transmitted to the "prediction and strategy dashboard", which then executes "precise intervention", and finally obtains "effect feedback data (including emotional feedback, renewal results, etc.)". This layer corresponds to the "business application and feedback layer" on the left. At the same time, the "effect feedback data" flows back to the relevant links through "closed-loop feedback for model optimization" to realize process iteration.

[0114] Example 2 See Figure 8 As shown, Figure 8 This illustration shows a structural schematic diagram of a customer personalized retention strategy determination device according to Embodiment 2 of this application, wherein the device includes: The multimodal data acquisition module 801 is used to collect multimodal data of the target customer; The dynamic customer profile building module 802 is used to build and update dynamic customer profiles of target customers based on the multimodal data. The policy renewal probability determination module 803 is used to extract customer features from the dynamic customer profile and input the customer features into the dual-channel renewal prediction model to obtain the policy renewal probability. The key feature determination module 804 is used to perform interpretability analysis on the policy renewal probability and determine the key features that affect the renewal probability. The retention strategy generation module 805 is used to generate a personalized retention strategy for the target customer based on the key features.

[0115] In an optional implementation, the multimodal data includes static attribute data, time-series behavioral log data, unstructured text data, and macro-environmental data; The process of constructing and updating a dynamic customer profile for the target customer based on the multimodal data includes: The static attribute data, the time-series behavior log data, the unstructured text data, and the macro-environment data are subjected to multimodal data fusion processing to construct the dynamic customer profile, wherein the dynamic customer profile includes customer attribute dimensions, behavioral trend dimensions, emotional tendency dimensions, and environmental association dimensions. The dynamic customer profile is continuously evolved using an event-triggered mechanism and a periodic incremental update mechanism, wherein: The event triggering mechanism immediately updates the corresponding dimension of the dynamic customer profile when a specific business event occurs; The periodic incremental update mechanism performs sliding window calculations on the time-series behavior log data and the sentiment sequence data extracted from unstructured text data according to a fixed time window, so as to update the trend indicators in the dynamic customer profile.

[0116] In an optional implementation, customer features are extracted from the dynamic customer profile, including: Static attribute features and policy features are extracted from the customer attribute dimensions of the dynamic customer profile. From the behavioral trend dimension of the dynamic customer profile, extract the statistical features, trend features, and sequence pattern features of the dynamic behavior sequence; From the emotional tendency dimension of the dynamic customer profile, extract emotional score features and interest topic features; Environmental association features are extracted from the environmental association dimension of the dynamic customer profile.

[0117] In an optional implementation, the step of inputting the customer characteristics into the dual-channel renewal prediction model to obtain the policy renewal probability includes: The static attribute features, the policy features, and the environmental association features are input into the static feature channel of the dual-channel renewal prediction model and processed through a fully connected network to obtain a static feature vector. The statistical features, trend features, sequence pattern features, sentiment score features, and topic of concern features are input into the dynamic time-series channel of the dual-channel insurance renewal prediction model. The time-series dependencies are modeled through gated recurrent units or long short-term memory networks, and an attention mechanism is used to assign weights to features at different time steps to obtain a dynamic time-series feature vector. The static feature vector and the dynamic time-series feature vector are fused together, and the policy renewal probability is output by calculating through a fully connected layer and the Softmax function.

[0118] In an optional implementation, an interpretability analysis is performed on the policy renewal probability to identify key characteristics affecting the renewal probability, including: The decision attribution of the policy renewal probability is performed, and SHAP or LIME interpretability analysis technology is applied to perform reverse analysis of the model decision process that generates the probability. Based on the reverse analysis, the contribution of each input feature of the dual-channel renewal prediction model to the policy renewal probability prediction result is calculated. Based on a preset contribution threshold, input features whose absolute contribution value exceeds the threshold are identified as key features, and a key influence factor report containing each key feature and its corresponding contribution value is generated.

[0119] In an optional implementation, generating a personalized retention strategy for the target customer based on the key features includes: The key features are matched with a preset feature strategy mapping rule library, wherein the feature strategy mapping rule library predefines the mapping relationship between key feature combinations and retention intervention actions; Based on the matching results, generate personalized retention strategy suggestions that include at least one specific intervention action.

[0120] In an optional implementation, the device further includes a closed-loop optimization module for: After implementing the personalized retention strategy, collect customer feedback data and renewal result data; The customer feedback data and the renewal result data are used as new multimodal data and fed back into the dynamic customer profile update process; Using the renewal result data as labels, the dual-channel renewal prediction model is retrained to achieve closed-loop optimization.

[0121] Example 3 Based on the same application concept, see [link / reference] Figure 9 As shown, Figure 9 This illustration shows a structural schematic diagram of a computer device provided in Embodiment 3 of this application, wherein, as shown... Figure 9 As shown, the computer device 900 provided in Embodiment 3 of this application includes: The computer device 900 includes a processor 901, a memory 902, and a bus 903. The memory 902 stores machine-readable instructions that can be executed by the processor 901. When the computer device 900 is running, the processor 901 communicates with the memory 902 through the bus 903. When the machine-readable instructions are executed by the processor 901, the steps of the customer personalized retention strategy determination method shown in Embodiment 1 above are performed.

[0122] Example 4 Based on the same concept, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the customer personalized retention strategy determination method described in any of the above embodiments.

[0123] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0124] The computer program product for determining personalized customer retention strategies provided in this application includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the preceding method embodiments. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0125] The customer personalized retention strategy determination device provided in this application embodiment can be specific hardware on a device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this application embodiment are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0126] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0127] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0128] In addition, the functional units in the embodiments provided in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0129] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0130] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0131] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for determining personalized customer retention strategies, characterized in that, include: Collect multimodal data from target customers; Based on the multimodal data, construct and update dynamic customer profiles of target customers; Customer features are extracted from the dynamic customer profile, and the customer features are input into the dual-channel renewal prediction model to obtain the policy renewal probability; An interpretability analysis was performed on the policy renewal probability to identify the key features affecting the renewal probability. Based on the aforementioned key features, a personalized retention strategy for the target customer is generated.

2. The method according to claim 1, characterized in that, The multimodal data includes static attribute data, time-series behavior log data, unstructured text data, and macro-environmental data; The process of constructing and updating a dynamic customer profile for the target customer based on the multimodal data includes: The static attribute data, the time-series behavior log data, the unstructured text data, and the macro-environment data are subjected to multimodal data fusion processing to construct the dynamic customer profile, wherein the dynamic customer profile includes customer attribute dimensions, behavioral trend dimensions, emotional tendency dimensions, and environmental association dimensions. The dynamic customer profile is continuously evolved using an event-triggered mechanism and a periodic incremental update mechanism, wherein: The event triggering mechanism immediately updates the corresponding dimension of the dynamic customer profile when a specific business event occurs; The periodic incremental update mechanism performs sliding window calculations on the time-series behavior log data and the sentiment sequence data extracted from unstructured text data according to a fixed time window, so as to update the trend indicators in the dynamic customer profile.

3. The method according to claim 2, characterized in that, Extracting customer features from the dynamic customer profile includes: Static attribute features and policy features are extracted from the customer attribute dimensions of the dynamic customer profile. From the behavioral trend dimension of the dynamic customer profile, extract the statistical features, trend features, and sequence pattern features of the dynamic behavior sequence; From the emotional tendency dimension of the dynamic customer profile, extract emotional score features and interest topic features; Environmental association features are extracted from the environmental association dimension of the dynamic customer profile.

4. The method according to claim 3, characterized in that, The step of inputting the customer characteristics into the dual-channel renewal prediction model to obtain the policy renewal probability includes: The static attribute features, the policy features, and the environmental association features are input into the static feature channel of the dual-channel renewal prediction model and processed through a fully connected network to obtain a static feature vector. The statistical features, trend features, sequence pattern features, sentiment score features, and topic of concern features are input into the dynamic time-series channel of the dual-channel insurance renewal prediction model. The time-series dependencies are modeled through gated recurrent units or long short-term memory networks, and an attention mechanism is used to assign weights to features at different time steps to obtain a dynamic time-series feature vector. The static feature vector and the dynamic time-series feature vector are fused together, and the policy renewal probability is output by calculating through a fully connected layer and the Softmax function.

5. The method according to claim 4, characterized in that, An interpretability analysis was performed on the policy renewal probability to identify key features affecting the renewal probability, including: The decision attribution of the policy renewal probability is performed, and SHAP or LIME interpretability analysis technology is applied to perform reverse analysis of the model decision process that generates the probability. Based on the reverse analysis, the contribution of each input feature of the dual-channel renewal prediction model to the policy renewal probability prediction result is calculated. Based on a preset contribution threshold, input features whose absolute contribution value exceeds the threshold are identified as key features, and a key influence factor report containing each key feature and its corresponding contribution value is generated.

6. The method according to claim 5, characterized in that, The process of generating personalized retention strategies for target customers based on the aforementioned key features includes: The key features are matched with a preset feature strategy mapping rule library, wherein the feature strategy mapping rule library predefines the mapping relationship between key feature combinations and retention intervention actions; Based on the matching results, generate personalized retention strategy suggestions that include at least one specific intervention action.

7. The method according to claim 1, characterized in that, The method further includes: After implementing the personalized retention strategy, collect customer feedback data and renewal result data; The customer feedback data and the renewal result data are used as new multimodal data and fed back into the dynamic customer profile update process; Using the renewal result data as labels, the dual-channel renewal prediction model is retrained to achieve closed-loop optimization.

8. A device for determining personalized customer retention strategies, characterized in that, The device includes: The multimodal data acquisition module is used to collect multimodal data from target customers. The dynamic customer profile building module is used to build and update dynamic customer profiles of target customers based on the multimodal data. The policy renewal probability determination module is used to extract customer features from the dynamic customer profile and input the customer features into the dual-channel renewal prediction model to obtain the policy renewal probability. The key feature determination module is used to perform interpretability analysis on the policy renewal probability and determine the key features that affect the renewal probability. The retention strategy generation module is used to generate personalized retention strategies for the target customers based on the key features.

9. A computer device, characterized in that, include: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the customer personalized retention strategy determination method as described in any one of claims 1 to 7 are performed.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the customer personalized retention strategy determination method as described in any one of claims 1 to 7.

Images