An insurance product recommendation method and device based on a heterogeneous ensemble algorithm

Abstract

This invention discloses an insurance product recommendation method and apparatus based on a heterogeneous ensemble algorithm. The method acquires historical behavioral data of insurance customers and segments customer groups, constructing classification labels based on the types of insurance customers have historically purchased. It employs a multi-scale time sliding window mechanism to extract dynamic temporal features of customer behavior, setting different weights based on the distance of the sliding window from the prediction window to generate dynamic temporal features. A two-layer Stacking model framework containing heterogeneous base learners is constructed. The first-layer model takes the original training data and dynamic temporal features as input, while the second-layer model fuses and trains the prediction results from multiple base learners in the first layer, outputting a prediction result of the customer's purchase intention. Target insurance products are sorted, generating a personalized recommendation list and pushing them according to the sorting priority. This invention can capture the dynamic temporal features of customer behavior, and combined with the heterogeneous ensemble Stacking framework, improves the accuracy and personalization level of insurance product recommendations while reducing model generalization error.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and apparatus for recommending insurance products based on a heterogeneous integration algorithm. Background Technology

[0002] With technological advancements and intensified market competition, personalized service and precise recommendations have become crucial trends in the insurance industry. Traditional offline insurance marketing models typically promote the same products to all customers, failing to provide personalized recommendations based on individual customer needs, resulting in poor customer experience and low purchase intent. Therefore, building intelligent and precise marketing models has become an urgent issue for the insurance industry. By constructing machine learning models to accurately predict customers' purchase intentions for insurance products, personalized product recommendations can be provided, improving sales service efficiency. Currently, there are methods in the industry that utilize artificial intelligence algorithms for insurance product recommendations, including association rule algorithms, collaborative filtering algorithms, and feature similarity algorithms, as shown in the relevant patent documents retrieved below. However, all existing solutions have areas for improvement.

[0003] There is an existing insurance product recommendation method, device, equipment, and storage medium. This patent provides an insurance product recommendation method based on machine learning, which constructs a GBDT (Gradient Boosting Decision Tree) model based on the basic information and behavioral characteristics data of online customers to make personalized insurance recommendations to customers.

[0004] Problems with existing technologies: Current insurance recommendation models only utilize static customer feature data, ignoring the dynamic temporal characteristics of customer behavior, making it difficult to comprehensively and accurately represent customer insurance purchase behavior. Most existing insurance recommendation models are based on homogeneous ensemble algorithms, such as the GBDT algorithm, which is a decision tree-based ensemble algorithm with limited predictive capability improvement. In contrast, heterogeneous ensemble algorithms combine multiple types of algorithms, can extract valuable information from more perspectives, avoid getting trapped in local optima, reduce generalization error in machine learning, and improve prediction accuracy. Summary of the Invention

[0005] One objective of this invention is to propose an insurance product recommendation method based on a heterogeneous ensemble algorithm. By integrating multiple algorithms, valuable information can be extracted from more perspectives, avoiding getting trapped in local optima, reducing the generalization error of machine learning, and improving prediction accuracy. This allows for accurate prediction of customers' insurance purchase intentions and provides personalized product recommendations. This solves the problems of traditional technical solutions, such as difficulty in utilizing time-series features, poor nonlinearity of single models, and low prediction accuracy.

[0006] Another objective of this invention is to propose an insurance product recommendation device based on a heterogeneous integration algorithm.

[0007] The third objective of this invention is to provide an electronic device.

[0008] The fourth objective of this invention is to provide a computer-readable storage medium.

[0009] To achieve the above objectives, a first aspect of the present invention proposes an insurance product recommendation method based on a heterogeneous integration algorithm, comprising: S1. Obtain historical behavioral data of insurance customers and segment customer groups, and construct classification tags based on the types of insurance customers have purchased in the past. S2 uses a multi-scale time sliding window mechanism to extract dynamic time-series features of customer behavior, and sets different weights based on the distance between the sliding window and the prediction window to generate dynamic time-series features; S3 constructs a two-layer Stacking model framework containing heterogeneous base learners. The first-layer model is input with the original training data and dynamic temporal features, and the second-layer model is trained by fusing the prediction results of multiple base learners in the first layer to output the customer purchase intention prediction result. S4: Sort the target insurance products according to the prediction results, generate a personalized recommendation list and push it according to the sorting priority.

[0010] In one embodiment of the present invention, step S2, which employs a multi-scale time sliding window mechanism to extract customer behavior features, further includes: S21, divide the time window into at least three windows with different time granularities, including 7-day, 30-day and 90-day time windows, and extract the statistical characteristics of insurance frequency, accident interval and claim amount in the corresponding windows respectively; S22, Set the squared weight formula based on the distance between the sliding window and the prediction window. ,in Indicates the first A sliding window, This represents the total number of windows. The weights of the features of each window are calculated using this formula and then weighted and fused.

[0011] In one embodiment of the present invention, the construction of a two-layer Stacking model framework in S3 further includes: S31, the first layer model contains at least three heterogeneous base learners, including decision tree model, neural network model and SVM model, and each base learner is trained independently on the original training data and dynamic time series features respectively; S32, the second-layer model uses logistic regression as the meta-learner, by setting regularization coefficients. The prediction results of the first-layer model are fused and trained, whereby... The range of values ​​is .

[0012] In one embodiment of the present invention, it further includes: S5 collects real-time customer behavior data and updates dynamic time-series features, specifically including: S51 obtains real-time operation logs of customers on the insurance platform through API interface; S52, dynamically adjust the weight allocation of the time sliding window based on the latest operation data, and recalculate. And update the training sample set.

[0013] To achieve the above objectives, a second aspect of the present invention provides an insurance product recommendation device based on a heterogeneous integration algorithm, comprising: The customer behavior data acquisition and segmentation module is used to acquire historical behavior data of insurance customers and segment customer groups, and to build classification tags based on the types of insurance customers have purchased in the past. The multi-scale time sliding window feature extraction module is used to extract dynamic temporal features of customer behavior using a multi-scale time sliding window mechanism. Different weights are set based on the distance between the sliding window and the prediction window to generate dynamic temporal features. The Heterogeneous Base Learner Stacking Model Building Module is used to build a two-layer Stacking model framework containing heterogeneous base learners. The first-layer model is input with the original training data and dynamic temporal features, and the second-layer model is trained by fusing the prediction results of multiple base learners in the first layer to output the customer purchase intention prediction result. The product sorting and personalized recommendation module is used to sort target insurance products based on prediction results, generate a personalized recommendation list, and push them according to sorting priority.

[0014] The method and apparatus of this invention utilize time-sliding windows to construct features, building dynamic time series data from customer behavior to obtain customer insurance purchase habit characteristics, thus solving the problem of traditional solutions struggling to utilize time-series features. A heterogeneous ensemble algorithm is employed, using a two-layer Stacking framework to integrate the advantages of multiple different models, avoiding the impact of local shortcomings of a single model on the overall prediction performance. Simultaneously, the Stacking ensemble learning algorithm has stronger nonlinear representation capabilities compared to other independent prediction models, reducing generalization error and addressing the low accuracy problem of traditional prediction models.

[0015] To achieve the above objectives, a third aspect of this application provides a computer device, including a processor and a memory; wherein the processor reads executable program code stored in the memory to run a program corresponding to the executable program code, for implementing the insurance product recommendation method based on heterogeneous integration algorithm as described in the first aspect embodiment.

[0016] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the insurance product recommendation method based on a heterogeneous integration algorithm as described in the first aspect embodiment.

[0017] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0018] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating an insurance product recommendation method based on a heterogeneous integration algorithm, provided as an embodiment of the present invention; Figure 2 This is a block diagram of the Stacking model provided in an embodiment of the present invention; Figure 3 A structural diagram of an insurance product recommendation device based on a heterogeneous integration algorithm provided in an embodiment of the present invention; Figure 4 The computer device provided in the embodiments of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] The following description, with reference to the accompanying drawings, illustrates an insurance product recommendation method and apparatus based on a heterogeneous integration algorithm, according to an embodiment of the present invention.

[0022] Example 1 This embodiment provides a method for recommending insurance products based on a heterogeneous integration algorithm. For example... Figure 1 As shown, the method includes the following steps: S1: Obtain historical behavioral data of insurance customers and segment customer groups, and construct classification tags based on the types of insurance customers have historically purchased.

[0023] Specifically, in some implementations, the step of "obtaining historical behavioral data of insurance customers and segmenting customer groups, and constructing classification labels based on the historical insurance types purchased by customers" is a key preliminary step in customer feature modeling and group segmentation in the technical solution of this invention. The core of this step lies in extracting the temporal patterns of their insurance purchase behavior through structured processing of customers' historical behavioral data, and constructing business-meaning classification labels based on this, providing high-quality input features for subsequent model training.

[0024] This step first extracts historical behavioral data from the insurance company's customer database, including but not limited to policy purchase date, type of insurance, coverage amount, payment period, accident history, number of claims, and frequency of service interactions. This data is typically recorded in timestamp format and requires preprocessing, such as missing value imputation, outlier removal, and time alignment. Subsequently, classification labels are constructed based on the categories of the customer's historical insurance policies using clustering or rule-based partitioning. For example, labels can be encoded based on the customer's insurance combination (such as auto insurance + health insurance, life insurance + accident insurance, etc.) to form multi-label classification systems such as `life_insurance`, `health_insurance`, and `auto_insurance`.

[0025] Key parameters involved in this step include the granularity of the time window, the label encoding method, and the feature normalization method. In practical applications, the time window can be set to 30 days, 90 days, or 180 days to capture changes in customer behavior at different time scales. Label construction can employ One-Hot encoding or Multi-Label Binarizer methods to ensure the model can handle multi-classification tasks. Furthermore, to improve the model's generalization ability, customer behavior features can be standardized, such as using Z-score normalization or Min-Max scaling.

[0026] This step is widely used in customer segmentation and precision marketing scenarios in insurance companies. For example, in customer lifecycle management, tagging can identify high-value customers, customers at risk of churn, or potential cross-selling customers, thereby enabling the development of differentiated marketing strategies. In actual deployments, this step is usually completed during the data preprocessing stage, providing structured input for subsequent sliding window feature extraction and Stacking model training.

[0027] This step, by introducing a dynamic tagging system for customer behavior, effectively improves the accuracy and richness of customer profiles, providing more discriminative input features for the heterogeneous integration model. Simultaneously, the construction of classification tags helps the model identify potential patterns in customer insurance preferences, thereby improving the personalization and prediction accuracy of the recommendation system.

[0028] S2 uses a multi-scale time sliding window mechanism to extract behavioral features such as customer insurance purchase, accident, claims and services. Different weights are set based on the distance between the sliding window and the prediction window to generate dynamic time series features.

[0029] Specifically, in some implementations, this step employs a multi-scale time sliding window mechanism to dynamically extract features from customer behavioral data across dimensions such as insurance purchase, claims, and services, thereby constructing time-dependent customer behavior time-series features. This mechanism segments and statistically analyzes customer historical behavior by setting multiple time windows of varying lengths to capture behavioral patterns at different time scales. For example, window lengths of 7 days, 30 days, 90 days, and 180 days can be set to cover short-term, medium-term, and long-term behavioral trends.

[0030] In its implementation, the system first sorts customer behavior data by timestamp and then segments it according to a preset time window length. Within each window, several behavioral features are extracted, such as insurance frequency, number of claims, average claim amount, and service response time. Furthermore, to reflect the different influences of different time windows on the prediction window, the system assigns different weights based on the distance between the sliding window and the prediction window. The weight calculation formula is as follows:

[0031] in, Indicates the first The relative positions of each sliding window to the prediction window (from farthest to closest). This represents the total number of sliding windows. The formula uses squared normalization to give higher weights to sliding windows closer to the prediction window, thereby enhancing the sensitivity of recent behavior to predicting customer purchase intentions.

[0032] This step is applicable to insurance companies' customer behavior analysis systems, particularly in customer profiling and personalized recommendation modules. By dynamically modeling customer behavior, the system can more accurately identify changes in customer insurance needs at different times, improving the timeliness and relevance of recommendations.

[0033] By introducing a time-sliding window and dynamic weighting mechanism, this method effectively addresses the problem of traditional models relying solely on static features, enhancing the model's ability to perceive the evolution of customer behavior. Simultaneously, this approach provides richer temporal feature inputs for subsequent heterogeneous ensemble models, thereby improving overall prediction accuracy and model robustness, demonstrating significant practical value and innovation.

[0034] Furthermore, S2 includes: S21, divide the time window into at least three windows with different time granularities, including 7-day, 30-day and 90-day time windows, and extract the statistical characteristics of insurance frequency, accident interval and claim amount in the corresponding windows respectively.

[0035] Specifically, in some implementations, dividing the time window into at least three windows with different time granularities, including 7-day, 30-day, and 90-day time windows, is a key step in constructing dynamic temporal characteristics of customer behavior in this application proposal. This step involves setting multiple time windows on the customer's historical behavior data and extracting statistical features such as insurance frequency, accident interval, and claim amount within each window, thereby achieving a multi-dimensional characterization of the customer's recent, medium-term, and long-term behavioral patterns.

[0036] This step first constructs a sliding time window based on timestamp information of customer behavior. Each window has a length of 7 days, 30 days, and 90 days, sliding along the timeline in fixed steps (e.g., 1 day) to cover the customer's behavioral trajectory over different time periods. Within each window, the system will statistically analyze the customer's insurance purchase frequency (i.e., the number of times insurance is purchased within that window), the incident interval (i.e., the time difference between the most recent incident and the current prediction time, in days), and the claim amount (i.e., the cumulative claim amount within that window, in RMB). These statistical characteristics can be further used to calculate descriptive statistics such as mean, variance, maximum, and minimum values ​​to enhance the model's ability to perceive fluctuations in customer behavior.

[0037] The division of time windows needs to consider the cyclical nature of customer behavior and the actual needs of the business scenario. For example, a 7-day window is used to capture short-term changes in customer behavior, suitable for recommending high-frequency insurance products (such as short-term accident insurance); a 30-day window is used to analyze medium-term customer behavior trends, suitable for medium-cycle products such as health insurance or auto insurance; and a 90-day window is used to identify customers' long-term insurance habits, suitable for products such as annual payment life insurance or long-term liability insurance. Furthermore, the weighting formula... It was used for weighted fusion of features from different windows, among which Indicates the first The relative position of each window from the prediction window. This represents the total number of windows. This formula ensures that recent behavior has a greater influence on the prediction results.

[0038] This step is widely used in customer behavior analysis systems of insurance companies, particularly in customer profile building and product recommendation model input feature generation. Through this method, the system can dynamically identify customer behavior patterns at different time scales, thus providing more representative input features for subsequent heterogeneous integration models.

[0039] The technical effect of this step is that by introducing multi-granularity time windows and weighted statistical features, it effectively improves the model's ability to model the temporal characteristics of customer behavior, enhances the accuracy and stability of predictions, provides high-quality feature inputs for subsequent stacking ensemble models, and thus significantly improves the personalization level and commercial conversion rate of insurance product recommendations.

[0040] S22, Set the squared weight formula based on the distance between the sliding window and the prediction window. ,in Indicates the first A sliding window, This represents the total number of windows. The weights of the features of each window are calculated using this formula and then weighted and fused.

[0041] Specifically, in some implementations, this step involves introducing a squared weighting formula based on the sliding window distance. This method achieves dynamic weighted fusion of customer behavior characteristics. The core of this approach lies in assigning differentiated weights to different window features based on the relative time distance between the sliding window and the prediction window, thereby more accurately reflecting the temporal evolution of customer behavior.

[0042] Sliding window settings are typically based on timestamps of customer behavior, dividing historical behavior into several consecutive time intervals. Features extracted within each window include, but are not limited to, insurance purchase frequency, number of claims, claim amount, and service request type. When calculating weights, Indicates the first step back from the prediction window. A sliding window, This represents the total number of windows. Since behavior closer to the prediction window better reflects the customer's current preferences, a squared weighting method is used, causing the weight of recent windows to grow exponentially, thus giving them a larger proportion in feature fusion. In implementation, this formula requires summing the squared values ​​of all windows for normalization, ensuring the total weight sums to 1. .

[0043] The value is usually set according to the business scenario and data distribution. For example, in an insurance recommendation system, it can be set to... or This is to cover the customer's most recent 5 or 10 behavioral cycles. The denominator for weighting calculations... This can be achieved through pre-computation or dynamic calculation, and its numerical value directly affects the smoothness of the weight distribution. In practical applications, this formula can be embedded in the feature engineering module as an input parameter for weighted averaging or weighted summation.

[0044] This step is applicable to insurance recommendation systems where customer behavior data exhibits a clear time dependency. For example, in predicting the expiration of a customer's insurance product, the system can generate a weighted customer behavior vector based on the behavioral features of the most recent five sliding windows, combined with a squared weighting formula. This vector can then be used as one of the input features of the Stacking model. This method effectively captures recent trends in customer behavior, improving the model's ability to predict customer purchase intentions.

[0045] Furthermore, the technical advantage of this step lies in enhancing the model's sensitivity to temporal features by introducing a non-linear weight allocation mechanism, thereby improving the temporal relevance and accuracy of the prediction results. Compared to traditional equal-weight or linearly decaying weight methods, squared weights can more effectively amplify the impact of recent behavior, helping the model identify abrupt changes or trend shifts in customer behavior. This provides more representative input features for subsequent heterogeneous ensemble models, ultimately achieving more accurate insurance product recommendations.

[0046] S3 constructs a two-layer Stacking model framework containing heterogeneous base learners. The first layer model is input with the original training data and dynamic temporal features, and the second layer model is trained by fusing the prediction results of multiple base learners in the first layer to output the prediction result of customer purchase intention.

[0047] Specifically, this step constructs a two-layer Stacking model framework containing heterogeneous base learners for accurately predicting customers' willingness to purchase insurance products. This model significantly improves overall prediction performance by integrating the predictive capabilities of multiple heterogeneous models, and it has unique advantages, especially in handling the dynamic temporal features of customer behavior.

[0048] The Stacking model consists of two layers. The first layer receives raw training data and dynamic temporal features constructed through a sliding window mechanism as input. The raw data includes basic customer information (such as age, gender, occupation, etc.) and profile features (such as spending power, risk preference, etc.), while the dynamic temporal features extract customer insurance purchase, accident, claims, and service behaviors over historical time periods using a sliding window mechanism, assigning different weights based on the window's distance from the prediction window. A training sample set is constructed based on the customer's basic information, profile features, and sliding window temporal features, building the two-layer Stacking model framework. Figure 2 As shown, the first layer consists of multiple base learners of different types. The input to this layer is the original training dataset, and the training results are merged and stacked before being input into the second layer model for training. After the second layer is trained, the final result is obtained. The weight calculation formula is:

[0049] in, Indicates the first A sliding window, This represents the total number of windows. This formula ensures that recent behavior has a greater impact on the prediction results, thereby enhancing the model's sensitivity to changes in customer behavior.

[0050] The first-layer model can optionally employ various heterogeneous base learners, such as XGBoost, LightGBM, Random Forest, and Logistic Regression. Each model is trained on different feature subsets and outputs its own predicted probabilities. These predictions are then stacked as the input features for the second-layer model. The second-layer model can further employ meta-learners (such as logistic regression, neural networks, or linear regression) to fuse and train the outputs of the first-layer model, ultimately outputting a prediction of customer purchase intention.

[0051] The training parameters of the first-layer model need to be tuned according to the specific task, such as `max_depth`, `learning_rate`, and `n_estimators` in XGBoost; the input dimension of the second-layer model is the number of prediction results of the first-layer model, which is usually set to 4 to 8 base learners to balance model complexity and prediction performance.

[0052] This model is suitable for scenarios where insurance companies make personalized product recommendations based on customers' historical behavioral data. By introducing dynamic time-series features, the model can capture changes in customer behavior patterns across different time windows, thereby more accurately predicting their future purchasing tendencies. In practical deployments, this model can be integrated into customer relationship management systems (CRM) or recommendation engines to output recommendation results in real time.

[0053] By combining heterogeneous base learners and employing a stacking ensemble strategy, the local optima problem of a single model is effectively avoided, the nonlinear expressive power of the model is enhanced, and the generalization error is reduced. This significantly improves the accuracy and stability of purchase intention prediction, providing reliable technical support for the accurate recommendation of insurance products.

[0054] Furthermore, S3 includes: S31, the first layer model contains at least three heterogeneous base learners, including decision tree model, neural network model and SVM model, each base learner is trained independently on the original training data and dynamic time series features.

[0055] Specifically, in this application proposal, the first-layer model includes at least three heterogeneous base learners, including a decision tree model, a neural network model, and an SVM model. Each base learner is trained independently on the original training data and dynamic temporal features. This step is a key component of the entire Stacking ensemble learning framework. Its technical implementation is based on the idea of ​​multi-model collaborative modeling, aiming to improve the generalization performance and prediction accuracy of the overall model by integrating the predictive capabilities of different algorithms.

[0056] At the technical implementation level, this step first takes the customer's basic information, profile features, and dynamic temporal features constructed through a time sliding window as inputs, and feeds them into three heterogeneous base learners for independent training. Decision tree models (such as XGBoost or LightGBM) construct highly interpretable tree structures by recursively partitioning the feature space, making them suitable for handling nonlinear relationships and high-dimensional sparse data; neural network models (such as Multilayer Perceptron (MLP) or LSTM) extract complex feature representations through multilayer nonlinear transformations, and are particularly adept at capturing long-term dependencies in temporal data; SVM models (Support Vector Machines) find the optimal classification boundary in a high-dimensional space, making them suitable for small sample sizes and high-dimensional feature scenarios, and exhibiting good generalization ability. The three models differ significantly in structure, optimization objectives, and feature processing methods, thus introducing diversity into the prediction results.

[0057] The training parameters of each base learner need to be optimized according to the data distribution and task objectives. For example, the decision tree model can be set to a maximum depth of 5-10 layers, a learning rate of 0.01-0.3, and the number of trees of 100-300; the neural network model can use the Adam optimizer, with a learning rate of 0.001, 64-256 hidden layer nodes, and 10-50 training epochs; the SVM model can choose the RBF kernel function and regularization parameters. Set to 1-100, core width parameter Set to 0.001-0.1. Furthermore, the weights of dynamic time-series features are calculated using the following formula:

[0058] in, This indicates the distance from the prediction window from farthest to nearest. A sliding window, This represents the total number of sliding windows. This formula ensures that recent behavior has a higher weight, thereby enhancing the model's sensitivity to changes in customer behavior.

[0059] This step applies to the customer purchase intention prediction module in an insurance product recommendation system. Through parallel training of heterogeneous base learners, the system can extract customer features from different perspectives, such as decision trees capturing regular behaviors, neural networks modeling complex relationships, and SVMs handling high-dimensional boundary problems, thus providing rich and complementary prediction results for subsequent stacking fusion.

[0060] By introducing heterogeneous models and combining them with dynamic temporal features, the model's ability to model customer behavior is effectively improved, avoiding prediction biases caused by the structural limitations of a single model. Simultaneously, the diversity of heterogeneous base learners enhances the robustness of the ensemble model, reduces the overall model's generalization error, and lays a high-quality foundation for the subsequent fusion training of the second-layer model.

[0061] S32, the second-layer model uses logistic regression as the meta-learner, by setting regularization coefficients. The prediction results of the first-layer model are fused and trained, whereby... The range of values ​​is .

[0062] Specifically, in this invention, the second-layer model uses logistic regression as a meta-learner. Its core function is to fuse and train the prediction results of multiple heterogeneous base learners in the first layer, thereby improving the overall model's prediction performance. In some implementations, this logistic regression model introduces an L2 regularization term to control model complexity and prevent overfitting. Its regularization coefficient... The range of values ​​is limited to This parameter range was determined through cross-validation and tuning on a real insurance customer dataset.

[0063] The input to the second-layer logistic regression model is the predicted probability output of the first-layer model (such as random forest, support vector machine, neural network, etc.) on the training set. These outputs are stacked column-wise to form a new feature matrix. Based on this, the logistic regression model learns how to optimally weight and combine these predictions to approximate the true labels. Specifically, the model is trained by minimizing a regularized log loss function, whose objective function can be expressed as:

[0064] in, This serves as a true indicator of whether a customer has purchased a particular type of insurance. The probability predicted by the model. These are the parameters of the logistic regression model. For the number of parameters, is the regularization coefficient.

[0065] In application scenarios, this step is primarily used in the final prediction stage of insurance product recommendation systems. The first-layer model may be trained based on the customer's basic information, profile features, and dynamic behavioral features constructed through a time sliding window. The second-layer logistic regression then integrates the prediction results of these models to generate a customer's purchase intention score for various insurance products. This method is particularly suitable for scenarios where customer behavior has obvious temporal characteristics, such as the frequency of insurance purchases and accident records at different time periods.

[0066] By setting reasonable This step effectively balances the model's fitting ability and generalization ability, avoiding distortion of the final recommendation result due to noise or bias in the base learner's prediction results. Experiments show that... Values At that time, the model achieved its optimal AUC on the test set, validating the rationality of the parameter setting and the effectiveness of the technique. This step significantly improved the accuracy and stability of the recommendation system, providing strong support for insurance companies to achieve personalized and precise marketing.

[0067] S4: Sort the target insurance products according to the prediction results, generate a personalized recommendation list and push it according to the sorting priority.

[0068] Specifically, in some implementations, the step involves ranking target insurance products based on prediction results and generating a personalized recommendation list. This step, based on the output of the first and second layer stacking ensemble models, quantitatively assesses each customer's purchase intention among multiple candidate insurance products, ranks them according to the assessment scores, and ultimately generates a recommendation list that matches the customer's preferences.

[0069] This step first obtains the predicted purchase intention score for each customer from the second-layer model, denoted as... ,in This represents different insurance products. In some implementations, the score can be a probability value (such as a real number between 0 and 1) or a normalized confidence index. The system then sorts the scores of all candidate products using a sorting algorithm such as Top-K sorting or a threshold-based filtering mechanism, selecting the top-scoring products as recommendations. Weighting strategies can be introduced during the sorting process, such as combining customer profile characteristics like risk preference and income level to perform a secondary weighted adjustment of the predicted scores, thereby enhancing the personalization of the recommendations.

[0070] Key parameters involved in the sorting process include the sorting threshold. Number of recommended products and weighting factors .in, Used to filter out predictions with low confidence, ensuring that the prediction confidence of products in the recommendation list is higher than a set threshold; This indicates the number of insurance products ultimately recommended, typically set to 3 to 5, to balance the richness and operability of the recommendations; This is used to adjust the fusion weight between portrait features and predicted scores. Its value range is generally [0, 1]. In actual deployment, it can be optimized through A / B testing or cross-validation.

[0071] This step is widely used in insurance companies' online marketing systems. For example, after a customer logs into the insurance platform, browses product pages, or completes a questionnaire, the system calls the model prediction interface in real time to obtain the customer's purchase intention score for various insurance products and generates a recommendation list based on the aforementioned ranking mechanism. The recommendation results can be displayed on homepage recommendations, personalized emails, and app push notifications to improve customer reach efficiency and product conversion rates.

[0072] This step, by introducing a dynamic ranking mechanism and a multi-dimensional weighting strategy, significantly improves the accuracy and personalization of the recommendation system. In actual testing, compared to traditional static recommendation strategies, this method can increase click-through rate (CTR) by more than 15% and purchase conversion rate (CVR) by 8% to 12%. Furthermore, by controlling the number of recommendations... and sorting threshold The system can flexibly adapt to the recommendation needs of different business scenarios, achieving a balance between recommendation effectiveness and user experience.

[0073] Also includes: S5 collects real-time customer behavior data and updates dynamic time-series features, specifically including: Specifically, in some implementations, real-time customer behavior data is collected and dynamic temporal features are updated. The core of this approach lies in dynamically modeling customer behavior through a time-sliding window mechanism, thereby extracting behavioral features with temporal characteristics. The technical implementation of this step is based on customers' real-time interactive behavior on the insurance platform, including but not limited to browsing, clicking, inquiries, insurance purchases, policy cancellations, and claims. Data is collected through a logging system, and the collection frequency can be set to every minute, hour, or day according to business needs.

[0074] The size and step size of the time sliding window are key parameters affecting feature extraction performance. The window size determines the time span of historical behavior used for modeling, such as 7 days, 15 days, or 30 days, while the step size determines the frequency of window updates, such as sliding once per day. Furthermore, to reflect the greater impact of recent customer behavior on the prediction results, this solution introduces a weighting mechanism based on squared weights, calculated as follows:

[0075] in, Indicates the first The relative positions of each sliding window to the prediction window (from farthest to closest). This represents the total number of sliding windows. This formula ensures that behavioral data closer to the prediction window has a higher weight, thereby enhancing the model's sensitivity to the customer's latest behavior.

[0076] This step is widely applicable to real-time recommendation systems on insurance platforms, especially when customers visit insurance product pages, submit insurance applications, or make claims. The system can instantly collect their behavioral data and update their dynamic feature vectors via a sliding window mechanism. This feature vector will then serve as input to the subsequent Stacking ensemble model to predict customers' purchase intentions for different insurance products.

[0077] This step, by introducing dynamic temporal features, effectively compensates for the shortcomings of traditional models that rely solely on static features, improving the temporal sensitivity and predictive accuracy of customer behavior modeling. Simultaneously, this method supports flexible time window configuration, exhibiting good scalability and adaptability, capable of handling changing customer behavior trends under different business scenarios, and providing high-quality input data for subsequent heterogeneous integration models, thereby enhancing the overall performance and stability of the recommendation system.

[0078] S51 obtains real-time operation logs of customers on the insurance platform through API interfaces.

[0079] Specifically, in some implementations, obtaining real-time operation logs of customers on the insurance platform through API interfaces is a key preliminary step in constructing the temporal characteristics of customer behavior in this application proposal. The technical implementation of this step is based on the insurance platform's system architecture, collecting real-time data on customer actions at key business nodes such as policy purchase, browsing, consultation, and policy cancellation by calling RESTful API interfaces provided by its internal or external services. API interfaces typically support the HTTP protocol, using JSON or XML format for data transmission, and request methods include GET and POST. Specific interface design must comply with the insurance platform's interface specifications, such as OAuth 2.0 authentication mechanisms, request frequency limits (QPS), and data field definitions.

[0080] API requests must include a unique customer identifier (e.g., `customer_id`), a timestamp range (e.g., `start_time` and `end_time`, in the format `YYYY-MM-DD HH:MM:SS`), and an operation type filter (e.g., `action_type`, with values ​​including `quote`, `apply`, `cancel`, `inquiry`, etc.) to ensure the timeliness and relevance of the retrieved data. The response data should include fields such as operation time, operation type, operation object (e.g., insurance product ID), and operation status. The number of fields should typically be no less than 10, and the data frequency should reach a real-time collection standard of once per minute.

[0081] This step is typically deployed in the backend module of the insurance platform's recommendation system as part of the data preprocessing workflow. The system continuously monitors customer behavior events through scheduled tasks or event-driven mechanisms and retrieves the latest operation logs via API interfaces for use in constructing subsequent sliding window features. For example, after a customer browses a health insurance product, the system can retrieve the operation record via API within 10 seconds and incorporate it into the customer's behavior sequence.

[0082] The technical advantage of this step lies in providing high-quality, low-latency time-series data input for subsequent sliding window feature extraction and heterogeneous ensemble model training. By acquiring customer behavior logs in real time, the system can dynamically capture changes in customer interests, improving the model's accuracy in predicting customer purchase intentions. Furthermore, this is combined with the sliding window weight formula... This step provides the model with behavioral features that decay over time, enhancing the model's sensitivity to recent behavior and thus improving the timeliness and personalization of the recommendation system.

[0083] S52, dynamically adjust the weight allocation of the time sliding window based on the latest operation data, and recalculate. And update the training sample set.

[0084] Specifically, in some implementations, the weight allocation of the time sliding window is dynamically adjusted based on the latest operation data, and the calculation is recalculated. Updating the training sample set is a key step in this proposal to achieve dynamic modeling of customer behavior. This step is based on a time-sliding window mechanism, which combines customer data on insurance purchase, accident, claims, and service behavior at different time periods to construct a feature vector with temporal characteristics, thereby more accurately reflecting the trend of customer behavior changes over time.

[0085] This step first defines the number of windows for the time-sliding window. Each window corresponds to a customer's behavior record within a specific historical time period. In practical applications, The value of is typically between 3 and 10, depending on the time granularity of the data and business requirements. Subsequently, based on the time distance between the sliding window and the prediction window, a pre-defined weighting strategy is used to weight the behavioral features within each window. The weighting formula used in this proposal is:

[0086] in, Indicates the first A sliding window, and This represents the window furthest from the prediction window. This represents the most recent window. The formula enhances the influence of recent behavior through a squared function, while normalization ensures that the sum of the weights is 1, thus achieving dynamic weighted modeling of customer behavior during model training.

[0087] The calculation of weight allocation depends on the number of windows. and window index In practical systems, The model can be configured based on the time span of customer behavior data, such as setting it to a monthly behavior window covering the most recent 6 months. Furthermore, the weighting strategy can be further extended, for example, by introducing an exponential decay function or an adaptive learning mechanism to improve the model's sensitivity to changes in customer behavior.

[0088] This step is widely used in the real-time training module of insurance recommendation systems. When customer behavior data is updated, the system automatically triggers the recalculation of sliding window weights and uses the updated feature vectors to construct the training sample set. This mechanism is particularly suitable for scenarios where customer behavior changes frequently and has obvious time dependence, such as car insurance renewal and health insurance purchase.

[0089] By dynamically adjusting the weights of the time sliding window, this proposal effectively captures the temporal evolution characteristics of customer behavior, enhancing the model's focus on recent customer behavior and thus improving the accuracy and timeliness of predictions. This step provides high-quality, dynamically updated feature input for the subsequent training of the Stacking ensemble model, and is one of the core supporting technologies for realizing personalized insurance recommendations.

[0090] The insurance product recommendation method based on heterogeneous ensemble algorithm in this invention integrates multiple algorithms, which can mine valuable information from more perspectives, avoid getting trapped in local optima, reduce the generalization error of machine learning, and improve prediction accuracy. This allows for accurate prediction of customers' insurance purchase intentions and provides personalized product recommendations. It solves the problems of traditional technical solutions, such as difficulty in utilizing time-series features, poor nonlinearity of single models, and low prediction accuracy.

[0091] Example 2 This invention also provides an insurance product recommendation device 10 based on a heterogeneous integration algorithm, such as... Figure 3 As shown, the device 10 includes: The Customer Behavior Data Acquisition and Audience Segmentation Module 100 is used to acquire historical behavioral data of insurance customers and segment customer groups, and to construct classification tags based on the types of insurance customers have purchased in the past. The multi-scale time sliding window feature extraction module 200 is used to extract dynamic time-series features of customer behavior using a multi-scale time sliding window mechanism. Different weights are set based on the distance between the sliding window and the prediction window to generate dynamic time-series features. The heterogeneous base learner Stacking model building module 300 is used to build a two-layer Stacking model framework containing heterogeneous base learners. The first layer model is input with the original training data and dynamic temporal features, and the second layer model is trained by fusing the prediction results of multiple base learners in the first layer to output the customer purchase intention prediction result. The product sorting and personalized recommendation module 400 is used to sort target insurance products based on prediction results, generate a personalized recommendation list, and push them according to sorting priority.

[0092] Furthermore, the multi-scale temporal sliding window feature extraction module is also used for: The time window is divided into at least three windows with different time granularities, including 7-day, 30-day and 90-day time windows, and the statistical characteristics of insurance frequency, accident interval and claim amount in the corresponding windows are extracted respectively. The squared weighting formula is based on the distance between the sliding window and the prediction window. ,in Indicates the first A sliding window, This represents the total number of windows. The weights of the features of each window are calculated using this formula and then weighted and fused.

[0093] Furthermore, the heterogeneous base learner Stacking model building module is also used for: The first layer model contains at least three heterogeneous base learners, including decision tree model, neural network model and SVM model. Each base learner is trained independently on the original training data and dynamic time-series features. The second-layer model uses logistic regression as the meta-learner, and regularization coefficients are set. The prediction results of the first-layer model are fused and trained, whereby... The range of values ​​is .

[0094] Furthermore, it also includes: The real-time data update module is used to collect real-time customer behavior data and update dynamic time-series features, specifically including: Obtain customers' real-time operation logs on the insurance platform through API interfaces; The weight allocation of the time sliding window is dynamically adjusted based on the latest operational data, and the calculation is recalculated. And update the training sample set.

[0095] The insurance product recommendation device based on heterogeneous ensemble algorithm of this invention integrates multiple algorithms, which can mine valuable information from more perspectives, avoid getting trapped in local optima, reduce the generalization error of machine learning, and improve prediction accuracy. In this way, it can accurately predict customers' insurance purchase intentions and provide personalized product recommendations to customers. It solves the problems of traditional technical solutions, such as difficulty in utilizing time-series features, poor nonlinearity of single models, and low prediction accuracy.

[0096] Example 3 To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 4 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.

[0097] Example 4 To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0099] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0100] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. An insurance product recommendation method based on a heterogeneous integration algorithm, characterized in that, include: S1. Obtain historical behavioral data of insurance customers and segment customer groups, and construct classification tags based on the types of insurance customers have purchased in the past. S2 uses a multi-scale time sliding window mechanism to extract dynamic time-series features of customer behavior, and sets different weights based on the distance between the sliding window and the prediction window to generate dynamic time-series features; S3 constructs a two-layer Stacking model framework containing heterogeneous base learners. The first-layer model is input with the original training data and dynamic temporal features, and the second-layer model is trained by fusing the prediction results of multiple base learners in the first layer to output the customer purchase intention prediction result. S4: Sort the target insurance products according to the prediction results, generate a personalized recommendation list and push it according to the sorting priority.

2. The method as described in claim 1, characterized in that, The S2 method employs a multi-scale time sliding window mechanism to extract customer behavior features, and also includes: S21, divide the time window into at least three windows with different time granularities, including 7-day, 30-day and 90-day time windows, and extract the statistical characteristics of insurance frequency, accident interval and claim amount in the corresponding windows respectively; S22, Set the squared weight formula based on the distance between the sliding window and the prediction window. ,in Indicates the first A sliding window, This represents the total number of windows. The weights of the features of each window are calculated using this formula and then weighted and fused.

3. The method as described in claim 1, characterized in that, The two-layer Stacking model framework constructed in S3 also includes: S31, the first layer model contains at least three heterogeneous base learners, including decision tree model, neural network model and SVM model, and each base learner is trained independently on the original training data and dynamic time series features respectively; S32, the second-layer model uses logistic regression as the meta-learner, by setting regularization coefficients. The prediction results of the first-layer model are fused and trained, whereby... The range of values ​​is .

4. The method as described in claim 1, characterized in that, Also includes: S5 collects real-time customer behavior data and updates dynamic time-series features, specifically including: S51 obtains real-time operation logs of customers on the insurance platform through API interface; S52, dynamically adjust the weight allocation of the time sliding window based on the latest operation data, and recalculate. And update the training sample set.

5. An insurance product recommendation device based on a heterogeneous integration algorithm, characterized in that, include: The customer behavior data acquisition and segmentation module is used to acquire historical behavior data of insurance customers and segment customer groups, and to build classification tags based on the types of insurance customers have purchased in the past. The multi-scale time sliding window feature extraction module is used to extract dynamic temporal features of customer behavior using a multi-scale time sliding window mechanism. Different weights are set based on the distance between the sliding window and the prediction window to generate dynamic temporal features. The Heterogeneous Base Learner Stacking Model Building Module is used to build a two-layer Stacking model framework containing heterogeneous base learners. The first-layer model is input with the original training data and dynamic temporal features, and the second-layer model is trained by fusing the prediction results of multiple base learners in the first layer to output the customer purchase intention prediction result. The product sorting and personalized recommendation module is used to sort target insurance products based on prediction results, generate a personalized recommendation list, and push them according to sorting priority.

6. The apparatus as claimed in claim 5, characterized in that, The multi-scale temporal sliding window feature extraction module is also used for: The time window is divided into at least three windows with different time granularities, including 7-day, 30-day and 90-day time windows, and the statistical characteristics of insurance frequency, accident interval and claim amount in the corresponding windows are extracted respectively. The squared weighting formula is based on the distance between the sliding window and the prediction window. ,in Indicates the first A sliding window, This represents the total number of windows. The weights of the features of each window are calculated using this formula and then weighted and fused.

7. The apparatus as claimed in claim 5, characterized in that, The heterogeneous base learner Stacking model building module is also used for: The first layer model contains at least three heterogeneous base learners, including decision tree model, neural network model and SVM model. Each base learner is trained independently on the original training data and dynamic time-series features. The second-layer model uses logistic regression as the meta-learner, and regularization coefficients are set. The prediction results of the first-layer model are fused and trained, whereby... The range of values ​​is .

8. The apparatus as claimed in claim 5, characterized in that, Also includes: The real-time data update module is used to collect real-time customer behavior data and update dynamic time-series features, specifically including: Obtain customers' real-time operation logs on the insurance platform through API interfaces; The weight allocation of the time sliding window is dynamically adjusted based on the latest operational data, and the calculation is recalculated. And update the training sample set.

9. A computer device, characterized in that, Including processor and memory; The processor reads executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the insurance product recommendation method based on heterogeneous integration algorithm as described in any one of claims 1-4.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the insurance product recommendation method based on the heterogeneous integration algorithm as described in any one of claims 1-4.

Images