Insurance loyalty prediction method and system based on multi-dimensional customer behavior data

Abstract

The invention provides an insurance loyalty prediction method and system based on multi-dimensional customer behavior data, and the method comprises the steps: extracting key behavior characteristics and behavior correlation characteristics based on the multi-dimensional customer behavior data, and constructing a multi-dimensional input vector; constructing an insurance loyalty prediction model by taking the multi-dimensional input vector as input and the insurance loyalty score as output; wherein the insurance loyalty prediction model comprises a base model and a meta model; inputting a multi-dimensional input vector corresponding to the multi-dimensional behavior data of the to-be-predicted customer into the base model and the meta-model, and obtaining an insurance loyalty predicted value of the to-be-predicted customer according to the output and the weight of the base model and the meta-model; and outputting the corresponding execution strategy according to the insurance loyalty prediction value of the to-be-predicted customer, thereby effectively improving the accuracy and reliability of insurance customer loyalty prediction.

Description

Technical Field

[0001] This invention relates to the field of insurance forecasting, and in particular to a method and system for predicting insurance loyalty based on multi-dimensional customer behavior data. Background Technology

[0002] With increasingly fierce competition in the insurance industry, customer loyalty has become a key factor for the long-term stable development of insurance companies. However, traditional customer assessment methods usually rely on static data and manual analysis, making it difficult to reflect the dynamic changes in customers in real time.

[0003] While existing technologies exist for predicting and evaluating customer loyalty based on user behavior data, most rely on a single type of user behavior data or a single predictive evaluation model. They lack comprehensive and accurate methods for evaluating and predicting customer loyalty across different dimensions and using different evaluation models. In particular, there are significant shortcomings in the analysis and integration of multi-dimensional data, resulting in low accuracy and reliability in insurance customer loyalty prediction.

[0004] To address this problem, the present invention provides a method and system for predicting insurance loyalty based on multi-dimensional customer behavior data, thereby solving the aforementioned issues. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention innovatively proposes an insurance loyalty prediction method and system based on multi-dimensional customer behavior data. This effectively solves the problem of low accuracy and reliability in insurance customer loyalty prediction caused by existing technologies, and effectively improves the accuracy and reliability of insurance customer loyalty prediction.

[0006] The first aspect of this invention provides a method for predicting insurance loyalty based on multi-dimensional customer behavior data, comprising: Collect multi-dimensional customer behavior data, extract key behavioral features and behavioral correlation features based on the multi-dimensional customer behavior data, and construct multi-dimensional input vectors; An insurance loyalty prediction model is constructed using a multi-dimensional input vector as input and an insurance loyalty score as output; wherein, the insurance loyalty prediction model includes a base model and a meta-model; The multi-dimensional input vectors corresponding to the multi-dimensional behavioral data of the customer to be predicted are input into the base model and the meta-model respectively. Based on the output and weight of the base model and the meta-model, the predicted value of the insurance loyalty of the customer to be predicted is obtained. Based on the predicted insurance loyalty value of the customer to be predicted, output the corresponding execution strategy.

[0007] Optionally, the customer multi-dimensional behavioral data includes insurance purchase records, claims records, renewal and cancellation behaviors, and marketing interaction behaviors, including invitation records, sales conversion records, and visit records.

[0008] Optionally, based on multi-dimensional customer behavior data, key behavioral features and behavioral correlation features are extracted to construct a multi-dimensional input vector, specifically including: Based on multi-dimensional customer behavior data, key behavioral characteristics are identified; these key behavioral characteristics include insurance purchase frequency index, insurance purchase-surrender cycle index, claims response index, renewal stability index, surrender frequency index, marketing response index, and visit response index. Based on behavioral correlation characteristics, the key behavioral characteristics or their weights are adjusted; among them, the behavioral correlation characteristics include the correlation between claims response time and marketing response index, the correlation between renewal reminder reach and monthly renewal rate, and the correlation between competitor consultation after policy cancellation and re-insurance probability. The modified key behavioral features are vectorized to construct a multi-dimensional input vector.

[0009] Furthermore, based on multi-dimensional customer behavioral data, the key behavioral characteristics are identified as follows: The insurance frequency index is the score corresponding to the range of insurance frequency. The specific calculation method for the policy application-surrender cycle index is as follows: T = min(100, D / 365×100); where T is the insurance-surrender cycle index, and D is the absolute value of the difference in days between the insurance application time and the surrender time. The specific calculation method for the claims response index is as follows: C=C w，时效 ×C Claims processing time score +C w，满意 × Claims satisfaction; where C is the claims response index, C w，时效 C is the weighting coefficient for the claims processing timeliness score. w，满意 This is a weighting coefficient for claims satisfaction. The specific calculation method for the renewal stability index is as follows: S = Number of consecutive renewal years / Total number of insurance years × 100; where S is the renewal stability index; The policy surrender frequency index is the score corresponding to the range of policy surrender frequencies; The marketing response index is calculated as follows: I = (Number of signed contracts at the meeting / Number of invited meetings) × 100; where I is the marketing response index; the specific calculation method for the visit response index is as follows: B = Number of valid responses / Number of visits × 100; where B is the visit response index.

[0010] Furthermore, based on behavioral association characteristics, the key behavioral features or their weights are adjusted as follows: When the customer claim response time exceeds the first preset time threshold, the corresponding visit response index is reduced by the first preset percentage, the weight of the claim response index is increased, and the weight of the marketing response index is reduced. If a customer's renewal completion rate exceeds the second preset percentage within 30 days of receiving a renewal reminder, the weight of the renewal stability index will be increased. If a customer consults about competing products after canceling their policy and is more than 300% likely to renew their policy within the first preset timeframe, the weighting of the marketing response index and the policy renewal frequency index will be increased.

[0011] Optionally, the base model includes a linear regression model, a random forest model, and an SVM model, and the meta-model is an LSTM time series model; the objective functions of the base model and the meta-model are: L = w1 × F + w2 × C + w3 × S w4×R+w5×I+w6×B+w7×T; Wherein, L is the score of the objective function, F is the insurance frequency index, w1 is the weight of the insurance frequency index, C is the claims response index, w2 is the weight of the claims response index, S is the renewal stability index, w3 is the weight of the renewal stability index, R is the surrender frequency index, w4 is the weight of the surrender frequency index, I is the marketing response index, w5 is the weight of the marketing response index, B is the visit response index, w6 is the weight of the visit response index, T is the insurance-surrender cycle index, and w7 is the weight of the insurance-surrender cycle index; wherein, the weights w1 of the insurance frequency index, w2 of the claims response index, w3 of the renewal stability index, w4 of the surrender frequency index, w5 of the marketing response index, w6 of the visit response index, and w7 of the insurance-surrender cycle index are all weight coefficients obtained through preprocessing and normalized; the preprocessing includes training with historical data and dynamic adjustment according to the type of insurance and the customer life cycle.

[0012] Furthermore, based on the outputs and weights of the base model and meta-model, the predicted insurance loyalty value for the customer to be predicted is obtained as follows: L 总 =W1×L1+W2×L2+W3×L3+W4×L4; Among them, L 总Let L1 be the score of the objective function output by the linear regression model, W1 be the weight output by the linear regression model, L2 be the score of the objective function output by the random forest model, W2 be the weight output by the random forest model, L3 be the score of the objective function output by the SVM model, W3 be the weight output by the SVM model, L4 be the score of the objective function output by the LSTM time series model, and W4 be the weight output by the LSTM time series model. Among these, the weights W1, W2, and W3 output by the linear regression model, random forest model, and SVM model are the same, and the weights W1, W2, and W3 output by the linear regression model, random forest model, and SVM model are all less than the weights W4 output by the LSTM time series model.

[0013] Optionally, based on the predicted insurance loyalty value of the customer to be predicted, the corresponding execution strategy is output, specifically including: When the predicted insurance loyalty value of the customer to be predicted is greater than the first preset value, the execution strategy of conducting a dedicated service follow-up once a month and giving priority to recommending high-end products will be output. When the predicted insurance loyalty value of the customer to be predicted is no greater than the first preset value but greater than the second preset value, the execution strategy for quarterly follow-up visits and targeted discounts will be output. When the predicted insurance loyalty value of the customer to be predicted is less than the second preset value, the execution strategy of triggering two promotional pushes and a salesperson's door-to-door visit within half a month will be output.

[0014] Furthermore, based on the predicted insurance loyalty values ​​of the customers to be predicted, the corresponding execution strategies output also include: For customers whose insurance loyalty prediction value is lower than the second preset value, the decision tree analysis is used to analyze the combination of behaviors and output the highest-ranked incentives in sequence, and a matching solution is matched for each incentive.

[0015] A second aspect of the present invention provides an insurance loyalty prediction system based on multi-dimensional customer behavior data, comprising: The data acquisition module collects multi-dimensional customer behavior data, extracts key behavioral features and behavioral correlation features based on the multi-dimensional customer behavior data, and constructs multi-dimensional input vectors. A construction module is used to build an insurance loyalty prediction model, taking a multi-dimensional input vector as input and an insurance loyalty score as output; wherein, the insurance loyalty prediction model includes a base model and a meta-model; The input module takes the multi-dimensional input vectors corresponding to the multi-dimensional behavioral data of the customer to be predicted and inputs them into the base model and the meta-model respectively. Based on the outputs and weights of the base model and the meta-model, the predicted value of the insurance loyalty of the customer to be predicted is obtained. The output module outputs the corresponding execution strategy based on the predicted insurance loyalty value of the customer to be predicted.

[0016] The technical solution adopted in this invention has the following technical effects: 1. The technical solution of this invention extracts key behavioral features and behavioral correlation features based on multi-dimensional customer behavioral data to construct a multi-dimensional input vector; using the multi-dimensional input vector as input and insurance loyalty score as output, an insurance loyalty prediction model is constructed; wherein, the insurance loyalty prediction model includes a base model and a meta-model; the multi-dimensional input vector corresponding to the multi-dimensional behavioral data of the customer to be predicted is input into the base model and the meta-model respectively, and the predicted insurance loyalty value of the customer to be predicted is obtained according to the output and weight of the base model and the meta-model; based on the predicted insurance loyalty value of the customer to be predicted, the corresponding execution strategy is output, effectively solving the problem of low accuracy and reliability of insurance customer loyalty prediction caused by existing technology, and effectively improving the accuracy and reliability of insurance customer loyalty prediction.

[0017] 2. The technical solution of this invention determines key behavioral features based on multi-dimensional customer behavior data; it corrects the key behavioral features or their weights based on behavioral correlation features; and it vectorizes the corrected key behavioral features to construct a multi-dimensional input vector. This not only considers behavioral features of different dimensions but also the correlation features between behaviors, further improving the accuracy and reliability of insurance customer loyalty prediction.

[0018] 3. In the objective function calculation of the loyalty prediction model in the technical solution of this invention, the weights of multiple feature vectors are weight coefficients obtained through preprocessing and normalized. The preprocessing includes training with historical data and dynamic adjustment based on insurance type and customer life cycle. Dynamic adjustment based on historical data training and insurance type and customer life cycle ensures the smooth progress of insurance customer loyalty prediction and improves the adaptability of insurance customer loyalty prediction.

[0019] 4. In the technical solution of this invention, different execution strategies are output according to the different predicted values ​​of the insurance loyalty of the customers to be predicted. Moreover, for customers whose predicted insurance loyalty values ​​are less than a second preset value, the behavior combination is analyzed through decision tree analysis, and the highest-ranked incentives are output in sequence. A matching solution is matched for each incentive. This makes it easier to optimize customer management, marketing strategies and risk control based on the predicted customer loyalty.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the method of Embodiment 1 in the present invention; Figure 2 This is a schematic diagram of the architecture of the method in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the system structure in Embodiment 2 of the present invention. Detailed Implementation

[0023] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0024] Example 1 like Figures 1-2 As shown, this invention provides a method for predicting insurance loyalty based on multi-dimensional customer behavior data, including: S1. Collect multi-dimensional customer behavior data, extract key behavioral features and behavioral correlation features based on the multi-dimensional customer behavior data, and construct multi-dimensional input vectors; S2, using a multi-dimensional input vector as input and an insurance loyalty score as output, constructs an insurance loyalty prediction model; wherein, the insurance loyalty prediction model includes a base model and a meta-model; S3. Input the multi-dimensional input vectors corresponding to the multi-dimensional behavioral data of the customer to be predicted into the base model and the meta-model respectively. Based on the outputs and weights of the base model and the meta-model, obtain the predicted value of the insurance loyalty of the customer to be predicted. S4 outputs the corresponding execution strategy based on the predicted insurance loyalty value of the customer to be predicted.

[0025] In step S1, i.e. Figure 2The middle data collection layer collects multi-dimensional customer behavior data, including insurance application records, claims records, renewal and cancellation behaviors, and marketing interaction behaviors. These marketing interaction behaviors include invitation and attendance records, sales conversion records, and visit records. Multi-dimensional user behavior data is collected through compliant data sources within the insurance company, including insurance application records, claims records, renewal and cancellation behaviors, and marketing interactions (invitation and attendance records, sales conversion records, and visit records). All data collection complies with Article 13 of the Personal Information Protection Law, processing only customer data necessary for fulfilling insurance contracts and after-sales services. Data transmission uses the TLS 1.3 encryption protocol, storage uses AES-256 encryption, and access control uses the RBAC (Role-Based Access Control) model.

[0026] In step S1 (i.e.) Figure 2 The middle data processing layer extracts key behavioral features and behavioral correlation features based on multi-dimensional customer behavioral data, and constructs multi-dimensional input vectors, specifically including: S11. Based on multi-dimensional customer behavior data, determine key behavioral characteristics; among which, key behavioral characteristics include insurance purchase frequency index, insurance purchase-surrender cycle index, claims response index, renewal stability index, surrender frequency index, marketing response index, and visit response index (each key behavioral characteristic ranges from 0 to 100). Specifically, the insurance frequency index is the score corresponding to the range of insurance frequency; The correspondence between the insurance frequency (times / year) and the corresponding score can be set in advance: 0 times / year → 0 points; 1 time / year → 40 points; 2 times / year → 70 points; ≥3 times / year → 100 points; or other corresponding ranges can be used, which are not limited in this embodiment.

[0027] The specific calculation method for the policy application-surrender cycle index is as follows: T = min(100, D / 365×100); where T is the insurance-surrender cycle index, and D is the absolute value of the difference in days between the insurance application time and the surrender time. Specifically, when there is no policy cancellation, T=95; when there is a short-term policy cancellation (D≤30 days), T=0.

[0028] The specific calculation method for the claims response index is as follows: C=C w，时效 ×C Claims processing time score +C w，满意 × Claims satisfaction; where C is the claims response index, C w，时效 C is the weighting coefficient for the claims processing timeliness score. w，满意 This is a weighting coefficient for claims satisfaction. Specifically, it includes two sub-indicators: ① Claims processing time score: unit = days; ≤3 days → 90–100; 3–7 days → 60–80; >7 days → 0–50; ② Claims satisfaction score: unit = 0–5 points (or other scores); conversion: satisfaction score = claims satisfaction / 5 × 100; Comprehensive calculation: C = 0.6 × claims processing time score + 0.4 × claims satisfaction score.

[0029] The specific calculation method for the renewal stability index is as follows: S = Number of consecutive renewal years / Total number of insurance years × 100; where S is the renewal stability index; The surrender frequency index is the score corresponding to the range of surrender frequencies. The correspondence between the surrender frequency (times / year) range and the corresponding score can be set in advance: 0 times / year → 0 points; 1 time / year → 60 points; ≥2 times / year → 100 points; or other corresponding ranges can be used, which are not limited in this embodiment.

[0030] The marketing response index is calculated as follows: I = (Number of signed contracts at the event / Number of invited attendees) × 100; where I is the marketing response index. Specifically, if there is no invitation record, the default score is 50 (neutral value).

[0031] The specific method for calculating the visit response index is as follows: B = Number of valid responses / Number of visits × 100; where B is the visit response index.

[0032] S12, Based on the behavioral correlation characteristics, adjust the key behavioral characteristics or the weights of key behavioral characteristics; among them, the behavioral correlation characteristics include the correlation characteristics between claims response time and marketing response index, the correlation characteristics between renewal reminder reach and monthly renewal rate, and the correlation characteristics between competitor consultation after policy cancellation and re-insurance probability. Specifically, when the customer claim response time exceeds the first preset time threshold, the corresponding visit response index is reduced by the first preset percentage, the weight of the claim response index is increased, and the weight of the marketing response index is reduced. Claims response time → Marketing response index: For customers whose claims take longer than 7 days, their visit response index decreases by an average of 10% (first preset percentage), and the weights of the claims response index C and the marketing response index I are adjusted or corrected. This invention constructs a causal relationship model of "claims experience → marketing interaction behavior" based on historical data analysis. When a user's claims response time exceeds 7 days, their subsequent visit response index decreases by an average of 10%. Based on this, the weights are dynamically adjusted: the weight w2 of the claims response index C (which can be the initial weight or a weight trained on historical data, i.e., the weight before normalization) is increased by 0.03–0.05, while the weight w5 of the marketing response index I (which can be the initial weight or a weight trained on historical data, i.e., the weight before normalization) is weakened (reduced) to improve the interpretability and accuracy of loyalty scoring.

[0033] If a customer's renewal completion rate exceeds the second preset percentage within 30 days of receiving a renewal reminder, the weight of the renewal stability index will be increased. Specifically, the renewal reminder reaches the 30-day renewal rate (if the renewal completion rate within 30 days after the reminder is >80%, the weight w3 of the renewal stability index S is increased by 0.05). Historical data shows that the renewal completion rate of users within 30 days after receiving the renewal reminder (SMS or phone call) exceeds 80% (the second preset percentage). Based on this, the behavioral causal chain of "renewal reminder reaching → renewal completion" is identified. When this chain is established, the weight w3 of the renewal stability index S (which can be the initial weight or the weight trained with historical data, i.e., the weight before normalization) is automatically increased by 0.03–0.05 to enhance the model's sensitivity to renewal behavior.

[0034] If a customer consults about competing products after canceling their policy and is more than 300% likely to renew their policy within the first preset timeframe, the weighting of the marketing response index and the policy renewal frequency index will be increased.

[0035] Specifically, the probability of a customer consulting with competitors after canceling their policy is higher than 60%, automatically classifying them as a high-recovery-potential customer. Behavioral sequence analysis reveals that if a user consults competitors after canceling their policy, the probability of them re-purchasing insurance within the next 3-6 months (a first preset timeframe) exceeds 60%. Based on this causal relationship, this invention automatically classifies such users as "high-recovery-potential customers" and increases the weight w5 of their marketing response index I and the weight w1 of their insurance frequency index F (this increase can be based on the initial weights or weights trained with historical data, i.e., weights before normalization) to enhance the accuracy of loyalty prediction.

[0036] S13, vectorize the corrected key behavioral features to construct a multi-dimensional input vector.

[0037] In step S2, the base models include linear regression, random forest, and SVM models, and the meta-model is an LSTM time series model. The objective functions of the base model and the meta-model are: L = w1 × F + w2 × C + w3 × S w4×R+w5×I+w6×B+w7×T; Wherein, L is the score of the objective function, F is the insurance frequency index, w1 is the weight of the insurance frequency index, C is the claims response index, w2 is the weight of the claims response index, S is the renewal stability index, w3 is the weight of the renewal stability index, R is the surrender frequency index, w4 is the weight of the surrender frequency index, I is the marketing response index, w5 is the weight of the marketing response index, B is the visit response index, w6 is the weight of the visit response index, T is the insurance-surrender cycle index, and w7 is the weight of the insurance-surrender cycle index. The weights w1 (insurance frequency index), w2 (claims response index), w3 (renewal stability index), w4 (surrender frequency index), w5 (marketing response index), w6 (visit response index), and w7 (insurance-surrender cycle index) are all normalized weight coefficients obtained through preprocessing. Preprocessing includes training with historical data and dynamic adjustments based on insurance type and customer lifecycle.

[0038] Specifically, the weights w1 for the insurance purchase frequency index, w2 for the claims response index, w3 for the renewal stability index, w4 for the surrender frequency index, w5 for the marketing response index, w6 for the visit response index, and w7 for the insurance purchase-surrender cycle index can be obtained through training on historical data from over 100,000 users over the past three years. This training employs a gradient descent method iteratively for over 1000 iterations, with a convergence threshold set to 1e-5. The initial weight range is as follows: w1 (insurance frequency index weight): 0.25–0.30; w2 (Claims Response Index Weighting): 0.20–0.30; w3 (Renewal Stability Index Weight): 0.25–0.35; w4 (surrender frequency index weight): 0.10–0.20 (negative weight); W5 (Marketing Response Index Weighting): 0.15–0.25; w6 (Visit Response Index Weighting): 0.10–0.20; w7 (Insurance application and surrender cycle index weight): 0.10–0.15; Weight normalization: ;in, Let be the normalized weight coefficient of the i-th term. is the weight coefficient of the i-th term before normalization.

[0039] After training on historical data, but before normalization, the weight coefficients can be dynamically adjusted: a scenario-based dynamic weight adjustment mechanism is designed to adapt to different insurance types and customer lifecycles. When adjusted according to insurance type: Health insurance: W2 (claims response index weighting) increased to 0.30; Life insurance: W3 (renewal stability index weighting) increased to 0.40; Property Insurance: W1 (insurance frequency index weight) is increased to 0.30.

[0040] Adjusted by customer lifecycle: New customers (≤1 year): W5 (marketing response index weight) changes from 0.15 to 0.25; Existing customers (≥5 years): w3 (renewal stability index weight) changed from 0.30 to 0.35.

[0041] Regional and group optimization: Combining regional characteristics and user groups, a clustering algorithm is used to adjust the weights after training on historical data. After the second adjustment, the weight coefficients are normalized to achieve personalized loyalty modeling of "one weight per person".

[0042] In step S3 (i.e.) Figure 2 The loyalty calculation engine, corresponding to steps S2-S3, obtains the predicted insurance loyalty value of the customer to be predicted based on the outputs and weights of the base model and meta-model. L 总 =W1×L1+W2×L2+W3×L3+W4×L4; Among them, L 总 Let L1 be the score of the objective function output by the linear regression model, W1 be the weight output by the linear regression model, L2 be the score of the objective function output by the random forest model, W2 be the weight output by the random forest model, L3 be the score of the objective function output by the SVM model, W3 be the weight output by the SVM model, L4 be the score of the objective function output by the LSTM time series model, and W4 be the weight output by the LSTM time series model. Among these, the weights W1, W2, and W3 output by the linear regression model, random forest model, and SVM model are the same, and the weights W1, W2, and W3 output by the linear regression model, random forest model, and SVM model are all less than the weights W4 output by the LSTM time series model.

[0043] Specifically, the base models are: weight W1 (weight 0.2) from the linear regression model, weight W2 (weight 0.2) from the random forest model, and weight W3 (weight 0.2) from the SVM model; the meta-model is: weight W4 (weight 0.4) from the LSTM time series model, which is used to capture the trend of loyalty changes; the final score is output through weighted voting.

[0044] The dynamic optimization strategy is as follows: Monitoring frequency: Data is scanned once per hour. If the user's key indicators fluctuate by ≥10%, the score will be automatically updated. Model retraining conditions: Triggered every quarter or when ≥50,000 new data entries are added. A / B testing criteria (for each model): 10,000 users are selected as the group, and the testing period is 1 month. The model with an accuracy improvement of ≥5% is the optimal base model or meta-model.

[0045] In step S4 (i.e.) Figure 2 The front-end presentation layer and back-end management module output corresponding execution strategies based on the predicted insurance loyalty values ​​of the customers to be predicted, specifically including: S41, when the predicted value of the insurance loyalty of the customer to be predicted is greater than the first preset value, output the execution strategy of monthly exclusive service follow-up and priority recommendation of high-end products; High loyalty customers (L) 总 ≥80 (first preset value is 80): monthly dedicated service follow-up + priority recommendation of high-end products.

[0046] S42, when the predicted value of the insurance loyalty of the customer to be predicted is not greater than the first preset value but greater than the second preset value, output the execution strategy of quarterly return visit and targeted discount; Medium-loyalty customers (50≤L<80, second preset value is 50): Quarterly follow-up visits + targeted discounts.

[0047] S43, when the predicted value of the customer's insurance loyalty is less than the second preset value, output the execution strategy of triggering two promotional pushes and a salesperson's door-to-door visit within half a month.

[0048] Low loyalty customers (L<50): Trigger two promotional pushes within 15 days + a salesperson visit.

[0049] Preferably, step S4 further includes: S44: For customers whose insurance loyalty prediction value is less than the second preset value, the decision tree analyzes the combination of behaviors and outputs the highest-ranked incentives in sequence, and matches a solution for each incentive.

[0050] For customers with a loyalty level below 50 (L<50), the "Choking Trigger Identification and Intelligent Solution" process is triggered. Specifically, churn trigger identification involves analyzing behavioral combinations using a decision tree, such as "≥2 policy cancellations + claims delays >7 days + lack of customer visits," to output the top 3 triggers (e.g., low claims efficiency, poor product fit). Intelligent solution matching follows. Low claims processing efficiency → Activate the "Fast Claims Processing Channel" (response within 24 hours); Poor product compatibility → Push personalized insurance combination plans; Lack of marketing follow-up → Assign a dedicated account manager for a second visit.

[0051] We comply with data protection regulations during data collection and processing, and employ methods such as data encryption, access control, and de-identification to ensure customer privacy and security.

[0052] This invention, through multi-dimensional data analysis and machine learning algorithms, can comprehensively calculate user loyalty and dynamically adjust prediction results. Based on user loyalty scores, insurance companies can formulate more targeted marketing strategies and customer management measures, improving customer retention and reducing churn. Through in-depth analysis of user behavior, the system can provide real-time feedback on user needs, offering personalized insurance product recommendations and services, thereby improving customer satisfaction. It can effectively monitor low-loyalty users and take timely intervention measures, helping insurance companies reduce fraud and management risks, achieving a closed-loop management system of "early warning – attribution – resolution." It can process massive amounts of user data and maintain efficient loyalty calculation capabilities during the business expansion of insurance companies, exhibiting good scalability.

[0053] This invention's technical solution extracts key behavioral features and behavioral correlation features from multi-dimensional customer behavioral data to construct a multi-dimensional input vector. Using this multi-dimensional input vector as input and an insurance loyalty score as output, an insurance loyalty prediction model is constructed. This model includes a base model and a meta-model. The multi-dimensional input vector corresponding to the multi-dimensional behavioral data of the customer to be predicted is input into both the base model and the meta-model. Based on the outputs and weights of the base model and meta-model, the predicted insurance loyalty value of the customer is obtained. Based on the predicted insurance loyalty value, a corresponding execution strategy is output. This effectively solves the problem of low accuracy and reliability in insurance customer loyalty prediction caused by existing technologies, and effectively improves the accuracy and reliability of insurance customer loyalty prediction.

[0054] The technical solution of this invention determines key behavioral features based on multi-dimensional customer behavior data; it corrects the key behavioral features or their weights based on behavioral correlation features; and it vectorizes the corrected key behavioral features to construct a multi-dimensional input vector. This not only considers behavioral features of different dimensions but also the correlation features between behaviors, further improving the accuracy and reliability of insurance customer loyalty prediction.

[0055] In the objective function calculation of the loyalty prediction model in this invention, the weights of multiple feature vectors are weight coefficients obtained through preprocessing and normalized. The preprocessing includes training with historical data and dynamic adjustment based on insurance type and customer lifecycle. Dynamic adjustment based on historical data training and insurance type and customer lifecycle ensures the smooth progress of insurance customer loyalty prediction and improves the adaptability of insurance customer loyalty prediction.

[0056] In this invention, different execution strategies are output according to the different predicted insurance loyalty values ​​of the customers to be predicted. Moreover, for customers whose predicted insurance loyalty values ​​are less than a second preset value, the behavior combinations are analyzed through decision tree analysis, and the highest-ranked incentives are output in sequence. A matching solution is provided for each incentive. This makes it easier to optimize customer management, marketing strategies and risk control based on the predicted customer loyalty.

[0057] Example 2 like Figure 3 As shown, the technical solution of the present invention also provides an insurance loyalty prediction system based on multi-dimensional customer behavior data, including: The data acquisition module 101 collects multi-dimensional customer behavior data, extracts key behavioral features and behavioral correlation features based on the multi-dimensional customer behavior data, and constructs a multi-dimensional input vector. Module 102 constructs an insurance loyalty prediction model by taking a multi-dimensional input vector as input and an insurance loyalty score as output; wherein, the insurance loyalty prediction model includes a base model and a meta-model; The input module 103 inputs the multi-dimensional input vectors corresponding to the multi-dimensional behavioral data of the customer to be predicted into the base model and the meta-model respectively. Based on the outputs and weights of the base model and the meta-model, the predicted value of the insurance loyalty of the customer to be predicted is obtained. Output module 104 outputs the corresponding execution strategy based on the predicted insurance loyalty value of the customer to be predicted.

[0058] The implementation process of the acquisition module 101, construction module 102, input module 103, and output module 104 in this embodiment 2 corresponds to the method steps in embodiment 1, and will not be repeated here.

[0059] This invention's technical solution extracts key behavioral features and behavioral correlation features from multi-dimensional customer behavioral data to construct a multi-dimensional input vector. Using this multi-dimensional input vector as input and an insurance loyalty score as output, an insurance loyalty prediction model is constructed. This model includes a base model and a meta-model. The multi-dimensional input vector corresponding to the multi-dimensional behavioral data of the customer to be predicted is input into both the base model and the meta-model. Based on the outputs and weights of the base model and meta-model, the predicted insurance loyalty value of the customer is obtained. Based on the predicted insurance loyalty value, a corresponding execution strategy is output. This effectively solves the problem of low accuracy and reliability in insurance customer loyalty prediction caused by existing technologies, and effectively improves the accuracy and reliability of insurance customer loyalty prediction.

[0060] The technical solution of this invention determines key behavioral features based on multi-dimensional customer behavior data; it corrects the key behavioral features or their weights based on behavioral correlation features; and it vectorizes the corrected key behavioral features to construct a multi-dimensional input vector. This not only considers behavioral features of different dimensions but also the correlation features between behaviors, further improving the accuracy and reliability of insurance customer loyalty prediction.

[0061] In the objective function calculation of the loyalty prediction model in this invention, the weights of multiple feature vectors are weight coefficients obtained through preprocessing and normalized. The preprocessing includes training with historical data and dynamic adjustment based on insurance type and customer lifecycle. Dynamic adjustment based on historical data training and insurance type and customer lifecycle ensures the smooth progress of insurance customer loyalty prediction and improves the adaptability of insurance customer loyalty prediction.

[0062] In this invention, different execution strategies are output according to the different predicted insurance loyalty values ​​of the customers to be predicted. Moreover, for customers whose predicted insurance loyalty values ​​are less than a second preset value, the behavior combinations are analyzed through decision tree analysis, and the highest-ranked incentives are output in sequence. A matching solution is provided for each incentive. This makes it easier to optimize customer management, marketing strategies and risk control based on the predicted customer loyalty.

[0063] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting insurance loyalty based on multi-dimensional customer behavior data, characterized in that, include: Collect multi-dimensional customer behavior data, extract key behavioral features and behavioral correlation features based on the multi-dimensional customer behavior data, and construct multi-dimensional input vectors; An insurance loyalty prediction model is constructed using a multi-dimensional input vector as input and an insurance loyalty score as output; wherein, the insurance loyalty prediction model includes a base model and a meta-model; The multi-dimensional input vectors corresponding to the multi-dimensional behavioral data of the customer to be predicted are input into the base model and the meta-model respectively. Based on the output and weight of the base model and the meta-model, the predicted value of the insurance loyalty of the customer to be predicted is obtained. Based on the predicted insurance loyalty value of the customer to be predicted, output the corresponding execution strategy.

2. The insurance loyalty prediction method based on multi-dimensional customer behavior data according to claim 1, characterized in that, The customer's multi-dimensional behavioral data includes insurance purchase records, claims records, renewal and cancellation behaviors, and marketing interaction behaviors, including invitation records, sales conversion records, and visit records.

3. The insurance loyalty prediction method based on multi-dimensional customer behavior data according to claim 1, characterized in that, Based on multi-dimensional customer behavior data, key behavioral features and behavioral correlation features are extracted to construct a multi-dimensional input vector, specifically including: Based on multi-dimensional customer behavior data, key behavioral characteristics are identified; these key behavioral characteristics include insurance purchase frequency index, insurance purchase-surrender cycle index, claims response index, renewal stability index, surrender frequency index, marketing response index, and visit response index. Based on behavioral correlation characteristics, the key behavioral characteristics or their weights are adjusted; among them, the behavioral correlation characteristics include the correlation between claims response time and marketing response index, the correlation between renewal reminder reach and monthly renewal rate, and the correlation between competitor consultation after policy cancellation and re-insurance probability. The modified key behavioral features are vectorized to construct a multi-dimensional input vector.

4. The insurance loyalty prediction method based on multi-dimensional customer behavior data according to claim 3, characterized in that, Based on multi-dimensional customer behavior data, the key behavioral characteristics are identified as follows: The insurance frequency index is the score corresponding to the range of insurance frequency. The specific calculation method for the policy application-surrender cycle index is as follows: T = min(100, D / 365×100); where T is the insurance-surrender cycle index, and D is the absolute value of the difference in days between the insurance application time and the surrender time. The specific calculation method for the claims response index is as follows: C=C w，时效 ×C Claims processing time score +C w，满意 × Claims satisfaction; where C is the claims response index, C w，时效 C is the weighting coefficient for the claims processing timeliness score. w，满意 This is a weighting coefficient for claims satisfaction. The specific calculation method for the renewal stability index is as follows: S = Number of consecutive renewal years / Total number of insurance years × 100; where S is the renewal stability index; The policy surrender frequency index is the score corresponding to the range of policy surrender frequencies; The marketing response index is calculated as follows: I = (Number of signed contracts at the meeting / Number of invited meetings) × 100; where I is the marketing response index; the specific calculation method for the visit response index is as follows: B = Number of valid responses / Number of visits × 100; where B is the visit response index.

5. The insurance loyalty prediction method based on multi-dimensional customer behavior data according to claim 4, characterized in that, Based on behavioral correlation characteristics, the key behavioral features or their weights are adjusted as follows: When the customer claim response time exceeds the first preset time threshold, the corresponding visit response index is reduced by the first preset percentage, the weight of the claim response index is increased, and the weight of the marketing response index is reduced. If a customer's renewal completion rate exceeds the second preset percentage within 30 days of receiving a renewal reminder, the weight of the renewal stability index will be increased. If a customer consults about competing products after canceling their policy and is more than 300% likely to renew their policy within the first preset timeframe, the weighting of the marketing response index and the policy renewal frequency index will be increased.

6. The insurance loyalty prediction method based on multi-dimensional customer behavior data according to claim 3, characterized in that, The base models include linear regression, random forest, and SVM models, and the meta-model is an LSTM time series model; the objective functions of the base models and the meta-model are: L=w1×F+w2×C+w3×S w4×R+w5×I+w6×B+w7×T; Wherein, L is the score of the objective function, F is the insurance frequency index, w1 is the weight of the insurance frequency index, C is the claims response index, w2 is the weight of the claims response index, S is the renewal stability index, w3 is the weight of the renewal stability index, R is the surrender frequency index, w4 is the weight of the surrender frequency index, I is the marketing response index, w5 is the weight of the marketing response index, B is the visit response index, w6 is the weight of the visit response index, T is the insurance-surrender cycle index, and w7 is the weight of the insurance-surrender cycle index; wherein, the weights w1 of the insurance frequency index, w2 of the claims response index, w3 of the renewal stability index, w4 of the surrender frequency index, w5 of the marketing response index, w6 of the visit response index, and w7 of the insurance-surrender cycle index are all weight coefficients obtained through preprocessing and normalized; the preprocessing includes training with historical data and dynamic adjustment according to the type of insurance and the customer life cycle.

7. The insurance loyalty prediction method based on multi-dimensional customer behavior data according to claim 6, characterized in that, Based on the outputs and weights of the base model and meta-model, the predicted insurance loyalty value for the customer to be predicted is as follows: L 总 =W1×L1+W2×L2+W3×L3+W4×L4; Among them, L 总 Let L1 be the score of the objective function output by the linear regression model, W1 be the weight output by the linear regression model, L2 be the score of the objective function output by the random forest model, W2 be the weight output by the random forest model, L3 be the score of the objective function output by the SVM model, W3 be the weight output by the SVM model, L4 be the score of the objective function output by the LSTM time series model, and W4 be the weight output by the LSTM time series model. Among these, the weights W1, W2, and W3 output by the linear regression model, random forest model, and SVM model are the same, and the weights W1, W2, and W3 output by the linear regression model, random forest model, and SVM model are all less than the weights W4 output by the LSTM time series model.

8. The insurance loyalty prediction method based on multi-dimensional customer behavior data according to claim 1, characterized in that, Based on the predicted insurance loyalty value of the customers to be predicted, the corresponding execution strategy is output, including: When the predicted insurance loyalty value of the customer to be predicted is greater than the first preset value, the execution strategy of conducting a dedicated service follow-up once a month and giving priority to recommending high-end products will be output. When the predicted insurance loyalty value of the customer to be predicted is no greater than the first preset value but greater than the second preset value, the execution strategy for quarterly follow-up visits and targeted discounts will be output. When the predicted insurance loyalty value of the customer to be predicted is less than the second preset value, the execution strategy of triggering two promotional pushes and a salesperson's door-to-door visit within half a month will be output.

9. The insurance loyalty prediction method based on multi-dimensional customer behavior data according to claim 8, characterized in that, Based on the predicted insurance loyalty values ​​of the customers to be predicted, the corresponding execution strategy output also includes: For customers whose insurance loyalty prediction value is lower than the second preset value, the decision tree analysis is used to analyze the combination of behaviors and output the highest-ranked incentives in sequence, and a matching solution is matched for each incentive.

10. An insurance loyalty prediction system based on multi-dimensional customer behavior data, characterized in that: include: The data acquisition module collects multi-dimensional customer behavior data, extracts key behavioral features and behavioral correlation features based on the multi-dimensional customer behavior data, and constructs multi-dimensional input vectors. A construction module is used to build an insurance loyalty prediction model, taking a multi-dimensional input vector as input and an insurance loyalty score as output; wherein, the insurance loyalty prediction model includes a base model and a meta-model; The input module takes the multi-dimensional input vectors corresponding to the multi-dimensional behavioral data of the customer to be predicted and inputs them into the base model and the meta-model respectively. Based on the outputs and weights of the base model and the meta-model, the predicted value of the insurance loyalty of the customer to be predicted is obtained. The output module outputs the corresponding execution strategy based on the predicted insurance loyalty value of the customer to be predicted.

Images