An insurance risk assessment method based on sports health data and federated learning and related products

Abstract

Embodiments of the present application provide an insurance risk assessment method based on sports health data and federated learning and related products. The method simultaneously collects static risk data and dynamic behavior data of a target object, assesses the risk degree from the dimensions of inherent risk and behavior risk, and avoids misjudgment caused by ignoring dynamic factors. Before data fusion, the quality of the two types of data is evaluated respectively to obtain their respective quality scores; then, weight analysis is performed based on the insurance product type and the quality scores to generate adaptive weights. This dual driving mechanism can not only make the weights match the business scenarios, but also automatically reduce the weight of a certain type of data when the quality of the data is low, thereby reducing the negative impact of low-quality data. The risk assessment model is trained based on a preset federated learning framework and has certain generalization ability. The model performs weighted calculation on the two types of data according to the adaptive weights to output a risk score, avoids the deviation caused by artificially fixed weights, and improves the evaluation accuracy and real risk mapping ability.

Description

Technical Field

[0001] This application relates to the field of insurance risk assessment technology, and in particular to an insurance risk assessment method and related products based on sports and health data and federated learning. Background Technology

[0002] Current risk assessment schemes based on medical insurance for insured individuals rely heavily on traditional methods, such as insurance company internal claims records, health disclosures, or limited medical examination data. These methods suffer from limited data dimensionality and are susceptible to subjective bias. While some schemes incorporate medical data like hospital electronic medical records and reports from medical examination institutions, the problem of data silos prevents information sharing between different institutions, hindering the formation of large-scale, high-quality training samples. Consequently, the models are unable to effectively capture risk signals for accurate medical insurance risk assessment. Summary of the Invention

[0003] To address the aforementioned issues and improve the accuracy of insurance risk assessment, this application provides an insurance risk assessment method and related products based on sports and health data and federated learning.

[0004] The embodiments of this application disclose the following technical solutions: In a first aspect, embodiments of this application provide an insurance risk assessment method based on sports and health data and federated learning, the method comprising: Acquire static risk data and dynamic behavioral data for the target object; the static risk data is used to characterize the static characteristics and historical state information of the target object; the dynamic behavioral data is used to characterize the periodically changing monitoring parameters of the target object; the dynamic behavioral data includes exercise and health data; The static risk data and the dynamic behavior data are evaluated for data quality to obtain static risk data quality scores and dynamic behavior data quality scores. Based on the insurance type of the target insurance product, a weight analysis is performed according to the static risk data quality score and the dynamic behavior data quality score to obtain the static risk data weight and the dynamic behavior data weight. Based on a pre-built risk assessment model, risk assessment is performed according to the weights of the static risk data, the weights of the dynamic behavior data, the static risk data, and the dynamic behavior data to obtain a risk assessment score for the target object; the risk assessment model is trained based on a preset federated learning architecture.

[0005] In one possible implementation, the pre-defined federated learning architecture includes multiple local data nodes and a single central data node, and each of the local data nodes is configured with the same baseline risk assessment model. The local data node is used to train the benchmark risk assessment model locally using the GBDT algorithm and local data, and to send the updated parameters of the trained model to the central data node; the GBDT algorithm is a gradient boosting decision tree algorithm. The central data node is used to perform iterative training of the model based on the model update parameters from each of the local data nodes to obtain the risk assessment model.

[0006] In one possible implementation, the step of iteratively training the model based on the model update parameters from each of the local data nodes to obtain the risk assessment model includes: Based on a preset parameter aggregation algorithm, the model update parameters obtained from training each local data node are aggregated to obtain global model update parameters; Based on a preset model convergence determination algorithm, the global model update parameters are converged; the preset model convergence determination algorithm includes convergence conditions. If the global model update parameters are determined to meet the convergence condition, the risk assessment model is generated based on the global model update parameters. If it is determined that the global model update parameters do not meet the convergence condition, the global model update parameters are sent to each of the local data nodes; Obtain the model update parameters obtained by each local data node in the next round of iterative training based on the global model update parameters; The risk assessment model is obtained by iteratively training the model based on the model update parameters obtained from the next round of training of each local data node.

[0007] In one possible implementation, the step of aggregating the model update parameters obtained from training each of the local data nodes based on a preset parameter aggregation algorithm to obtain global model update parameters includes: The sum of the products of the aggregate weights associated with each local data node and the corresponding model update parameters is determined as the global model update parameters; the aggregate weights are determined based on the proportion of data volume, the proportion of data quality score, and the proportion of training accuracy of each local data node. Wherein, the data volume ratio is the proportion of the training data volume of the local data node in the total training data of all the local data nodes; the data quality score ratio is the proportion of the data quality score of the local data node in the total data quality score of all the local data nodes; and the training accuracy ratio is the proportion of the local model training accuracy of the local data node in the total local model training accuracy of all the local data nodes.

[0008] In one possible implementation, the convergence determination of the global model update parameters based on a preset model convergence determination algorithm includes: Obtain the current iteration training number corresponding to the global model update parameters; Based on the mean squared error loss function, determine the loss function value of the model output corresponding to the global model update parameters; If the loss function value is not greater than the loss function threshold or the current number of training iterations is greater than the maximum number of iterations, it is determined that the global model update parameters satisfy the convergence condition. If the loss function value is greater than the loss function threshold and the current iteration training number is less than or equal to the maximum iteration number, it is determined that the global model update parameters do not meet the convergence condition.

[0009] In one possible implementation, after generating the risk assessment model, the method further includes: Based on the risk assessment model, risk prediction is performed according to a preset verification dataset associated with the target object to obtain multiple sets of predicted risk results; the preset verification dataset includes multiple sets of original risk data and risk type identifiers corresponding to each set of original risk data; Based on the multiple sets of predicted risk results and the associated risk type identifiers, the risk assessment model is evaluated to obtain model performance indicators. If the model performance index is less than the performance index threshold, a data update instruction is sent to each of the local data nodes to control each of the local data nodes to perform the local model update based on the updated local data.

[0010] In one possible implementation, the risk assessment based on a pre-built risk assessment model, according to the static risk data weights, the dynamic behavioral data weights, the static risk data, and the dynamic behavioral data, to obtain a risk assessment score for the target object, includes: Based on the risk assessment model, risk assessments are performed according to the static risk data and the dynamic behavioral data to obtain static risk scores and dynamic risk scores. The risk assessment score is determined by summing the product of the static risk data weight and the static risk score, and the product of the dynamic risk score and the dynamic behavioral data weight.

[0011] In one possible implementation, before sending the trained model update parameters to the central data node, the method further includes: Differential privacy encryption is applied to the model update parameters obtained from the training to obtain the encrypted model update parameters.

[0012] Secondly, embodiments of this application provide an insurance risk assessment system based on sports and health data and federated learning, the system comprising: The data acquisition module is used to acquire static risk data and dynamic behavioral data for the target object; the static risk data is used to characterize the static characteristics and historical state information of the target object; the dynamic behavioral data is used to characterize the periodically changing monitoring parameters of the target object; the dynamic behavioral data includes exercise and health data; The quality assessment module is used to assess the data quality of the static risk data and the dynamic behavior data to obtain a static risk data quality score and a dynamic behavior data quality score. The weighting analysis module is used to perform weighting analysis based on the insurance type of the target insurance product, according to the static risk data quality score and the dynamic behavior data quality score, to obtain the static risk data weight and the dynamic behavior data weight. The risk assessment module is used to perform risk assessment based on a pre-built risk assessment model, according to the weights of the static risk data, the weights of the dynamic behavior data, the static risk data, and the dynamic behavior data, to obtain a risk assessment score for the target object; the risk assessment model is trained based on a preset federated learning architecture.

[0013] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the possible insurance risk assessment methods in the first aspect.

[0014] Compared to existing technologies, this application offers the following advantages: This application provides an insurance risk assessment method and related products based on sports and health data and federated learning. In this method, static risk data and dynamic behavioral data of the target object are collected simultaneously. The risk level of the target object is assessed from both inherent and behavioral risks, thereby reducing risk misjudgments caused by ignoring dynamic factors. Furthermore, before data fusion, the two types of data are assessed separately to obtain static risk data quality scores and dynamic behavioral data quality scores. Weight analysis is then performed based on the type of the target insurance product and the quality scores to generate static risk data weights and dynamic behavioral data weights. This dual-driven mechanism adapts the weights to the business logic of the insurance scenario and automatically reduces the weight of a certain type of data when its quality is low, avoiding negative impacts on the assessment results from low-quality data. Finally, the risk assessment model trained based on a preset federated learning framework can break down data silos, achieve multi-source privacy-preserving joint modeling, and improve generalization ability. The risk assessment model performs weighted calculations based on the above adaptive weights and the two types of data to output a risk score. This approach ensures that the assessment results accurately reflect the true level of risk while avoiding bias from artificially fixed weights, thereby improving the accuracy of insurance risk assessment. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating an insurance risk assessment method based on sports health data and federated learning, provided for an embodiment of this application; Figure 2 A flowchart illustrating a model iterative training method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an insurance risk assessment system based on sports health data and federated learning, provided as an embodiment of this application. Detailed Implementation

[0017] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0018] As described earlier, current risk assessment schemes based on medical insurance for insured individuals rely heavily on traditional methods, such as insurance company internal claims records, health declarations, or limited medical examination data. These methods suffer from limited data dimensionality and are susceptible to subjective bias. While some schemes incorporate medical data such as hospital electronic medical records and reports from medical examination institutions, the problem of data silos prevents the sharing of medical information between different institutions, hindering the formation of large-scale, high-quality training samples. Consequently, the models are unable to effectively capture risk signals for accurate medical insurance risk assessment.

[0019] Based on this, this application provides an insurance risk assessment method and related products based on sports and health data and federated learning. In this method, static risk data and dynamic behavioral data of the target object are collected simultaneously. The risk level of the target object is assessed from both inherent and behavioral risks, thereby reducing risk misjudgments caused by ignoring dynamic factors. Furthermore, before data fusion, the two types of data are separately quality-assessed to obtain static risk data quality scores and dynamic behavioral data quality scores. Weight analysis is then performed based on the type of the target insurance product and the quality scores to generate static risk data weights and dynamic behavioral data weights. This dual-driven mechanism adapts the weights to the business logic of the insurance scenario and automatically reduces the weight of a certain type of data when its quality is low, avoiding negative impacts on the assessment results. Finally, the risk assessment model trained based on a preset federated learning framework can break down data silos, achieve multi-source privacy-preserving joint modeling, and improve generalization ability. The risk assessment model performs weighted calculations based on the above adaptive weights and the two types of data to output a risk score. This ensures accurate mapping of the assessment results to the true risk level while avoiding bias from manually fixed weights, thereby improving the accuracy of insurance risk assessment.

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

[0021] It should be noted that in the following description of this embodiment, medical insurance will be used as an example of an insurance product.

[0022] See Figure 1 The figure is a flowchart illustrating an insurance risk assessment method based on sports health data and federated learning provided in an embodiment of this application, specifically including the following steps: S101: Acquire static risk data and dynamic behavioral data for the target object; the static risk data is used to characterize the static characteristics and historical state information of the target object; the dynamic behavioral data is used to characterize the periodic changes in monitoring parameters of the target object; the dynamic behavioral data includes exercise and health data.

[0023] Static risk data refers to information related to the inherent attributes, historical status, or long-term stable characteristics of a target object. This data does not change significantly over a short timescale or is considered relatively fixed benchmark information within the assessment period. Specifically, static risk data is used to characterize the static characteristics and historical status information of a target object. Taking the risk assessment scenario of medical insurance as an example, static characteristics include, but are not limited to, the target object's age, gender, place of birth, genetic information, past medical history, surgical records, and chronic disease diagnosis results; historical status information covers health events, medical records, medication history, and abnormal indicators in physical examination reports that have occurred in the past period. In the context of medical insurance, static risk data can be understood as the insured's existing medical records and physical examination results; in the context of auto insurance, static risk data corresponds to vehicle model, vehicle age, driver's age, historical traffic violation records, and accident records. The characteristic of static risk data is high stability, but its disadvantage is low update frequency, which cannot reflect the impact of recent behavioral changes on the target object's risk status in a timely manner. Therefore, in the insurance risk assessment system of this embodiment, static risk data serves as the basis for basic risk assessment, providing a long-term risk benchmark for the target object in the risk assessment model.

[0024] Dynamic behavioral data refers to monitoring parameters used to characterize the periodic changes or real-time fluctuations of a target object. It reflects the target object's behavioral patterns, activity regularities, or trends in physiological indicators over time. Dynamic behavioral data is typically collected continuously or frequently through devices such as wearable devices, mobile applications, vehicle sensors, and IoT terminals, and is characterized by rapid updates. In medical insurance risk assessment scenarios, common dynamic behavioral data include daily average exercise volume, steps, activity duration, exercise intensity, heart rate and its variability, sleep duration and quality, sedentary time, respiratory rate, and body temperature changes. In auto insurance risk assessment scenarios, dynamic behavioral data sets can include the number of times of sudden acceleration and braking, mileage and time periods, and frequency of device use. In medical insurance scenarios, dynamic behavioral data can help assess the target object's daily health management habits; for example, a long-term lack of exercise or insufficient sleep significantly increases the risk of cardiovascular disease. In auto insurance scenarios, dynamic behavioral data can reflect a driver's driving style in real time; for example, frequent sudden braking indicates a higher probability of accidents.

[0025] In one possible implementation, after obtaining static risk data and dynamic behavioral data, data preprocessing procedures such as de-identification and standardization can be performed. De-identification removes personally identifiable information from the data, such as names, ID numbers, device IDs, and user nicknames, retaining only health or behavioral characteristics. A unique identifier is then added to the processed data to support traceability.

[0026] S102: Perform data quality assessment on the static risk data and the dynamic behavior data to obtain static risk data quality score and dynamic behavior data quality score.

[0027] In this embodiment, to ensure the accuracy of risk assessment, the process of risk assessment based on static risk data and dynamic behavioral data needs to determine the weight of each data in the risk assessment based on the data quality of each data.

[0028] Specifically, the quality assessment of static risk data mainly revolves around three dimensions: completeness, accuracy, and timeliness. Completeness assessment checks for missing key fields in the static risk data. For example, in a medical insurance scenario, it checks whether core fields such as the age, gender, past medical history, and medication records of the assessed individual are complete. If the missing percentage exceeds a preset threshold, the completeness score for that portion of the data will decrease accordingly. Accuracy assessment focuses on the authenticity and reliability of the data, such as whether diagnostic records come from legitimate medical institutions and whether physical examination reports are certified. In practical applications, methods such as inter-field logical consistency checks can be used for quantitative scoring. Timeliness assessment determines whether the static risk data is still valid. For example, physical examination data is generally valid for 6 to 12 months; data exceeding this period will have a lower timeliness score, while lifelong, unchanging data such as birth dates and genetic information will receive full marks for timeliness. The static risk data quality score is calculated using a weighted average, for example, 40% for completeness, 40% for accuracy, and 20% for timeliness, ultimately resulting in a percentage score. A higher quality score indicates that the batch of static risk data is more reliable and contributes more to subsequent weighted fusion or federated learning aggregation.

[0029] In addition to completeness, accuracy, and timeliness, quality assessment of dynamic behavioral data also requires attention to data continuity and sampling frequency. Completeness assessment primarily checks the coverage of dynamic behavioral data over time, such as whether daily average activity levels are recorded daily and whether there are long gaps in data collection. If the data missing rate exceeds 30% within a certain period, the quality score of the dynamic behavioral data at that point will be penalized. Accuracy assessment involves issues such as sensor noise and outlier detection. For example, a sudden jump in heart rate from 80 to 200 and then immediately returning to normal might be an artifact caused by a loose sensor, requiring the removal of such outliers through sliding window filtering or thresholding. Timeliness assessment in dynamic behavioral data reflects the real-time nature or freshness of the data. For example, when used for short-term risk prediction, data from the past week is more valuable than data from three months ago; therefore, a timeliness score can be calculated using a time decay weighting method. Dynamic behavioral data quality scoring typically uses a weighted summation method, such as 30% for completeness, 30% for accuracy, 20% for timeliness, and 20% for continuity. For nodes with excessively low quality scores, they can be temporarily excluded from federated training or their weight in parameter aggregation can be automatically reduced, thereby ensuring the robustness of the overall risk assessment model.

[0030] S103: Based on the insurance type of the target insurance product, perform weight analysis according to the static risk data quality score and the dynamic behavior data quality score to obtain the static risk data weight and the dynamic behavior data weight.

[0031] The first step in weighted analysis is to determine the baseline weights for static risk data and dynamic behavioral data based on the type of the target insurance product. Different types of insurance products have different core focuses in risk assessment, therefore the pre-set importance of the two types of data also differs accordingly. For example, in the risk assessment scenario of medical insurance, static characteristics such as the insured's past medical history and family history have a strong predictive power for the probability of disease occurrence; therefore, the baseline weight of static risk data is usually high, generally set at 0.7 to 0.9. Dynamic behavioral data, as an auxiliary indicator, has a lower baseline weight. Conversely, in the risk assessment scenario of auto insurance, real-time monitoring data of driving behavior can directly reflect the insured's current risk level; the baseline weight of dynamic behavioral data can be increased to 0.5 or even 0.7, while the weight of static risk data is correspondingly reduced. This baseline weight reflects prior business knowledge, ensuring a reasonable bias in the risk assessment model under different product scenarios and avoiding assessment distortion caused by a singular weight allocation.

[0032] After obtaining the baseline weights, the baseline weights are adaptively adjusted by combining the static risk data quality score and the dynamic behavioral data quality score, thereby avoiding the negative impact of low-quality data on the evaluation results. Specifically, when the quality score of either type of data is low (for example, the missing rate of dynamic behavioral data exceeds 40%), its actual contribution decreases. If the baseline weights are still maintained in this case, significant noise will be introduced. This solution adopts a quality-weighted adjustment strategy: the static baseline weight is multiplied by the static risk data quality score, and the dynamic baseline weight is multiplied by the dynamic behavioral data quality score. Then, the two are normalized so that the sum of the two adjusted weights is normalized again. The core idea is that, while preserving the business preferences determined by the type of insurance product, the weights are redistributed according to the actual reliability of the two types of data. The side with higher data quality will have its final weight appropriately increased relative to the baseline weight; the side with lower data quality will have its final weight decreased accordingly. For example, in a medical insurance risk assessment scenario, the static baseline weight should ideally be high. However, if the actual static physical examination data obtained is a year out of date and has a low quality score, while the dynamic behavioral data has a stable sampling frequency and an extremely low missing rate, then the weight of the dynamic behavioral data will be appropriately increased after correction, while the weight of the static risk data will be slightly decreased. This mechanism achieves adaptive adjustment of data quality while maintaining the business orientation of insurance product type, thereby improving the accuracy of risk assessment.

[0033] S104: Based on a pre-built risk assessment model, risk assessment is performed according to the static risk data weights, the dynamic behavior data weights, the static risk data, and the dynamic behavior data to obtain a risk assessment score for the target object; the risk assessment model is trained based on a preset federated learning architecture.

[0034] Based on a pre-built risk assessment model, pre-processed static risk data and dynamic behavioral data are used as input features. Combined with the weights of static risk data and dynamic behavioral data determined in the weight analysis phase, a comprehensive risk assessment is performed. Specifically, the risk assessment model first extracts key indicators from the static risk data (such as age, past medical history, and abnormal physical examination items) and time-series features from the dynamic behavioral data (such as daily average exercise volume, heart rate variability, and sedentary duration). Through non-linear mapping or linear weighting within the model, static risk feature scores and dynamic behavioral feature scores are calculated. Subsequently, the static risk data weights and dynamic behavioral data weights are used as fusion coefficients to perform a weighted sum of the two feature scores: the static risk feature score is multiplied by the static risk data weight, and the dynamic behavioral feature score is multiplied by the dynamic behavioral data weight. The sum of these two values ​​yields a comprehensive risk assessment score. This score is typically a continuous numerical value, with a range that can be preset according to business requirements (e.g., 0 to 100 or 0 to 1). A higher value indicates a higher overall risk level for the target object.

[0035] Specifically, the process of determining the risk assessment score in step S104 is achieved through the following two steps: Step 1: Based on the risk assessment model, perform risk assessments according to the static risk data and the dynamic behavioral data respectively to obtain static risk scores and dynamic risk scores.

[0036] The input data is split into a static risk data branch and a dynamic behavior data branch, and each branch is calculated independently through its corresponding sub-network or feature mapping layer in the model. For static risk data, the model extracts the inherent attributes (such as age, gender, past medical history, abnormal physical examination indicators, etc.) and historical status information (such as hospitalization records of the past three years, medication continuity, etc.). Through a pre-trained feature weight matrix and nonlinear transformation, it outputs a continuous static risk score, which reflects the basic risk level of the target object based on long-term stable factors, and typically ranges from 0 to 100.

[0037] For dynamic behavioral data, the model focuses on its temporal changes, such as the fluctuation trend of daily average exercise volume, recent changes in heart rate variability, and weekly cumulative sedentary time. Using the temporal branch of a recurrent neural network or gradient boosting tree, a dynamic risk score is calculated. This score captures the impact of the target's recent behavioral patterns on current risk. The calculation of the two types of scores is independent and does not interfere with each other, and both are based on the same set of trained and converged model parameters, ensuring consistency in the scoring scale. In practical applications, such as a medical insurance assessment scenario, the model outputs a static risk score of 75 based on the target's age, blood pressure, and other static characteristics, while simultaneously outputting a dynamic risk score of 90 based on dynamic behaviors such as a sharp drop in steps and severe sleep deprivation over the past week. These two scores reflect the different contributions of long-term health baseline and short-term behavioral deterioration to risk, respectively.

[0038] Step 2: The summation of the product of the static risk data weight and the static risk score, and the product of the dynamic risk score and the dynamic behavioral data weight, is determined as the risk assessment score.

[0039] After obtaining the static and dynamic risk scores, the weights of the static risk data and dynamic behavioral data determined in the weight analysis phase are combined and a weighted summation is performed to obtain the final risk assessment score. The specific calculation rule is: the static risk data weight multiplied by the static risk score, plus the dynamic behavioral data weight multiplied by the dynamic risk score; the result is the comprehensive risk assessment score. Since the sum of the static risk data weight and the dynamic behavioral data weight is 1, the final score is a convex combination of the two scores, always remaining within the original score range, facilitating subsequent risk level classification and business decisions. Taking the aforementioned example values, if the static risk data weight is 0.65 and the dynamic behavioral data weight is 0.35 after weight analysis, then the final risk score = 0.65 × 75 + 0.35 × 90 = 48.75 + 31.5 = 80.25. This score is higher than the static score but lower than the dynamic score, indicating that short-term behavioral risks are appropriately included in the overall assessment, but do not fully cover the long-term baseline. The final score can be directly compared with the preset risk level thresholds (e.g., low risk ≤60, medium risk 60-80, high risk >80) to output the risk level.

[0040] In this embodiment, the risk assessment model is trained based on a preset federated learning architecture, which includes multiple local data nodes and a central data node. The local data nodes are deployed at different data holders (e.g., multiple hospitals, medical examination institutions, health management platforms, wearable device manufacturers, or sports application platforms), each holding its own local user data, and the data remains within its domain. The central data node is deployed at an insurance institution or a third-party platform and is responsible for coordinating the overall training process. To ensure consistency in the initial model state and effective subsequent aggregation, each local data node is configured with the exact same baseline risk assessment model, i.e., a model with the same algorithm structure and the same initial model parameters. The local data nodes and the central data node in the preset federated learning architecture will be described in detail below.

[0041] The local data node is used to train the benchmark risk assessment model locally using the GBDT algorithm and local data, and send the updated model parameters obtained from the training to the central data node; the GBDT algorithm is a gradient boosting decision tree algorithm.

[0042] Within the pre-defined federated learning architecture of this embodiment, each local data node undertakes the core task of training the baseline risk assessment model using local data. The specific algorithm employed is Gradient Boosting Decision Tree (GBDT). GBDT is an ensemble learning algorithm that iteratively trains multiple decision trees to gradually approximate the true risk labels: each new tree attempts to fit the prediction residuals of the previous tree, and the prediction results of all trees are summed as the final output. During local training, each local data node starts with the baseline risk assessment model issued by the central data node, using static risk data and dynamic behavioral data stored locally as training samples. Features and labels remain within the local data node. In each iteration, the local data node calculates the residuals for each sample based on the current model, then constructs a new decision tree to fit these residuals. The leaf nodes of the new tree output incremental values, which are scaled by the learning rate and then accumulated onto the model.

[0043] After training, local data nodes send the updated model parameters obtained in this round of training to the central data node. These updated parameters can be all local model parameters or the parameter changes relative to the previous round of global model training. In one possible implementation, before sending parameters from the local data nodes to the central data node, differential privacy encryption can be applied to the trained model to send the encrypted updated model parameters to the central data node. Specifically, after completing local model training, each local node adds random noise following a Laplace distribution to its obtained model parameters. The noise amplitude is controlled by a very small privacy budget coefficient, typically between 0.01 and 0.05. The magnitude of this noise matches the sensitivity of the model parameters, ensuring that even if an attacker obtains the encrypted parameters, they cannot reverse-engineer any individual user's original information with high confidence. This effectively defends against common privacy threats such as parameter reverse engineering attacks and member inference attacks. Finally, each local node only uploads the encrypted model parameters to the central node, thus securely participating in federated modeling without leaking the original data, thereby ensuring the information security of the model under the federated training architecture.

[0044] The central data node is used to perform iterative training of the model based on the model update parameters from each of the local data nodes to obtain the risk assessment model.

[0045] On the other hand, the central data node is used to aggregate the training results of each local data node and drive iterative model updates. Specifically, the central data node receives model update parameters uploaded by all participating local data nodes. These parameters are generated by each node after independent training based on its own local data. The central node uses these parameters as input and generates a new round of global model parameters according to a preset federated optimization strategy. This parameter serves as both the output of the current iteration and the starting point for the next round of training. If the updated global model parameters do not meet the preset convergence conditions, the central node distributes them to each local node to trigger the next round of local training. By repeatedly executing such an iterative update cycle, the global model parameters continuously evolve between the local optimization of each node and the global coordination of the central node, gradually improving the model's ability to predict risk. When the set convergence conditions are met, the central node stops iterating and solidifies the finally aggregated model parameters, forming a risk assessment model that can be used in practice. Throughout the entire iterative training process, the central data node does not need to access any raw data; it only needs to process the encrypted model update parameters to complete the collaborative training of the global model, thereby achieving federated modeling of multi-source heterogeneous data.

[0046] Next, we will introduce the process of performing model iterative training at the central data node, with reference to the accompanying drawings of a specific implementation example.

[0047] See Figure 2 The figure is a flowchart illustrating a model iterative training method provided in an embodiment of this application, specifically including the following steps: S201: Based on a preset parameter aggregation algorithm, the model update parameters obtained from training each local data node are aggregated to obtain global model update parameters.

[0048] In a federated learning architecture, the central data node needs to merge the model update parameters uploaded by each local data node into a unified global model update parameter. This embodiment uses a weighted average aggregation algorithm to achieve the fusion of the model update parameters. Specifically, the central node multiplies the aggregation weight corresponding to each local data node by the model update parameter uploaded by that node to obtain the weighted model parameters of each node. Then, the weighted model parameters of all nodes are summed to obtain the final global model update parameter. Here, the aggregation weight is a value between 0 and 1, and the sum of the aggregation weights of all nodes is 1.

[0049] Specifically, the formula for the preset parameter aggregation algorithm is as follows: ; In the formula, This represents the aggregation weight corresponding to the local data node. ; The percentage of data in the i-th node. The percentage of data quality score for the i-th node. This represents the percentage of local model training accuracy at the i-th node. This represents the number of local data nodes. Let be the encryption model parameters for the i-th local data node. Update parameters for the global model.

[0050] The aggregation weights are determined based on three metrics: data volume percentage, data quality score percentage, and training accuracy percentage. Data volume percentage refers to the proportion of training data from a particular local data node to the total training data from all local nodes. For example, if all nodes have 100,000 training data entries, and a particular node contributes 20,000, then that node's data volume percentage is 0.2. This metric reflects the richness of the samples held by the node; the larger the data volume, the stronger the statistical representativeness, and the higher the weight should be. Data quality score percentage is the proportion of the quality score, derived from the completeness, accuracy, and timeliness of each node's data, to the total quality score of all nodes. The data quality score uses a percentage system and is calculated comprehensively from multiple dimensions such as missing data rate, outlier ratio, and timestamp freshness. The higher the quality score of a node, the more reliable its data, and the higher its weight will be during aggregation. Training accuracy percentage refers to the proportion of the prediction accuracy of a particular node's local model on the training set to the total accuracy of all nodes' local models. Accuracy can be calculated for each node using its local validation set (e.g., the ratio of correctly predicted samples to the total number of samples). Higher accuracy indicates a stronger fit to the data at that node, and more reliable its parameter updates. Finally, the aggregation weight is obtained by weighted summation of the three proportions mentioned above: 40% for data volume, and 30% each for data quality score and training accuracy. This combined design balances quantity, quality, and effectiveness, avoiding bias caused by a single metric dominating the calculations.

[0051] The model update parameters of each local data node are aggregated based on a preset parameter aggregation algorithm, so that nodes with good local training accuracy have a greater influence on the global parameters, thereby guiding the global model to optimize in a more accurate direction.

[0052] S202: Based on a preset model convergence determination algorithm, perform convergence determination on the global model update parameters; the preset model convergence determination algorithm includes convergence conditions.

[0053] After each round of parameter aggregation to obtain the global model update parameters, the central data node needs to perform a convergence determination. This involves judging whether the conditions defined by the preset model convergence determination algorithm are met, thereby deciding whether to terminate the iterative training. Specifically, step S202, based on the preset model convergence determination algorithm, performs the convergence determination of the global model update parameters through the following four steps: Step 1: Obtain the current iteration training number corresponding to the global model update parameters; Step 2: Based on the mean squared error loss function, determine the loss function value of the model output corresponding to the global model update parameters.

[0054] The first step in convergence determination is to obtain the iteration training rounds corresponding to the current global model update parameters. During the iterative process of federated learning, after each aggregation round, the central data node records the current iteration number, typically starting from the first round. The iteration count increments by one for each aggregation and distribution. This iteration count is the basic data for determining whether the preset maximum iteration limit has been reached. The second step is to determine the loss function value corresponding to the current global model parameters based on the mean squared error loss function. The mean squared error loss function is one of the most commonly used loss functions in regression tasks. Its calculation process is as follows: The central data node uses the current global model parameters to perform forward prediction on each sample in the training dataset, obtaining the model's predicted risk score; then, it calculates the square of the difference between the true risk label and the predicted score for each sample; finally, it sums the squared differences of all samples and divides by the total number of samples to obtain the mean squared error, which is the loss function value. The loss function value intuitively reflects the degree of fit of the current global model to the training data. The smaller the loss value, the closer the model's prediction result is to the true label, and the higher the model's accuracy. In actual federated training, since the original data does not leave the local machine, the central data node cannot directly obtain the true labels and predicted values ​​of all samples. Therefore, the global loss is usually obtained by calculating the local loss separately at each local node and then reporting it up for weighted averaging. However, the logic for determining convergence is completely equivalent to direct calculation. The loss function threshold (e.g., 0.03) and the maximum number of iterations (e.g., 200 rounds) are both preset by the system before training begins. The former depends on the accuracy requirements of the business, while the latter is used to avoid the training process from going on indefinitely.

[0055] Specifically, the loss function value is calculated as follows: ; In the formula, This represents the global model parameters in the k-th iteration, where N is the total amount of training data. For true health risk labels, To score the risk predicted by the model.

[0056] Step 3: If the loss function value is not greater than the loss function threshold or the current iteration training number is greater than the maximum iteration number, determine that the global model update parameters satisfy the convergence condition.

[0057] Step three is used to determine whether the convergence condition is met based on the loss function value and the current number of training iterations. The specific rule is: if the current loss function value is not greater than the preset loss function threshold, or the current number of training iterations is greater than the maximum number of iterations, the global model update parameters are considered to meet the convergence condition. The convergence determination uses an OR logic, meaning that convergence is triggered if either of the two sub-conditions is met. The first sub-condition is that the loss function value decreases to an acceptable range, meaning the model has learned a sufficiently effective risk mapping relationship, and the marginal benefit of further training is negligible. Terminating the iteration at this point avoids overfitting and wasting computational resources. For example, with a preset threshold of 0.03, the condition is met when the calculated mean squared error is 0.028. The second sub-condition is that the number of iterations has exceeded the maximum number of iterations. This is to prevent the model from getting stuck in an infinite loop if the loss cannot be reduced below the threshold due to extreme data distribution or improper learning rate settings. For example, with a maximum number of iterations set to 200, the system will forcibly stop and determine convergence at the 201st iteration, even if the loss is still higher than the threshold. This dual-safety mechanism balances model accuracy and training efficiency, prioritizing accuracy while setting a hard upper limit as a safety net. In practical applications, if the training run limit is triggered due to persistent loss, an alert can be issued, indicating the need to adjust the model structure, hyperparameters, or supplement with high-quality data.

[0058] Step 4: If the loss function value is greater than the loss function threshold and the current iteration training number is less than or equal to the maximum iteration number, it is determined that the global model update parameters do not meet the convergence condition.

[0059] Step four defines the scenarios where convergence conditions are not met. The specific rule is: if the current loss function value is greater than the loss function threshold, and the current training iteration count is less than or equal to the maximum number of iterations, the global model parameter update is deemed not to have met the convergence condition. This is an AND operation, requiring both conditions to be true simultaneously: the model's prediction error still exceeds the acceptable range, and the maximum number of iterations has not yet been reached. In this case, the system will not terminate training but will proceed to the next round of federated iterations.

[0060] S203: If the global model update parameters are determined to meet the convergence condition, the risk assessment model is generated based on the global model update parameters.

[0061] Once the convergence algorithm confirms that the global model update parameters meet the preset convergence conditions, the model solidification phase begins. Specifically, the central data node locks the currently aggregated global model update parameters, converting them from a temporary parameter state during training into a formal risk assessment model usable in the production environment. The generated assessment model can be deployed on the central node to receive real-time risk assessment requests, or it can be lightweighted and distributed to edge devices to support offline assessment scenarios. It's important to note that generating the risk assessment model does not mean the parameters are permanently frozen. In practical applications, the model can be retrained and iterated using newly collected labeled data periodically or triggered. However, for the current training cycle, the solidified risk assessment model represents the optimal risk assessment capability up to that point. After generating the risk assessment model, services can be provided externally, receiving static risk data and dynamic behavioral data of the target object, and outputting risk scores and levels.

[0062] S204: If it is determined that the global model update parameters do not meet the convergence condition, the global model update parameters are sent to each of the local data nodes.

[0063] If the convergence determination result indicates that the convergence condition is not met, it means that the current global model has not yet reached sufficient prediction accuracy and there is still room for further optimization, thus entering the preparation stage for a new round of iterative training. The specific operation of step S204 is as follows: the central data node sends the currently aggregated global model update parameters to all participating local data nodes. After receiving the latest global model parameters, each local data node uses them to overwrite its current model state, serving as the starting point for the next round of local training. The reason for distributing the global parameters to all nodes is that the core logic of federated learning requires each node to conduct independent training from a unified benchmark, ensuring the comparability and aggregability of the subsequently uploaded model update parameters. If some nodes remain on the old model version, their updated parameters will not be on the same optimization trajectory as those of other nodes, and direct aggregation will lead to model degradation or even divergence. Furthermore, the distribution process also includes training metadata for this iteration, such as the current loss value and suggested learning rate adjustment direction, so that each node can adaptively adjust the gradient descent step size or tree growth strategy during local training. After the distribution is completed, the central data node enters a waiting state until it receives a new round of model update parameters.

[0064] S205: Obtain the model update parameters obtained by each local data node in the next round of iterative training based on the global model update parameters.

[0065] After the global model update parameters are successfully distributed to each local data node, each node independently starts the next round of iterative training. The local training process is exactly the same as the first round: each node uses the received global parameters as the initial model, utilizes its own stored local data, and performs multiple rounds of local iteration using the gradient boosting decision tree algorithm to fit the residuals in the data and gradually update the model parameters. Due to the different data distributions of each node, the model parameters after local training will undergo differentiated changes. Some nodes may focus more on optimizing age-related splitting features, while others may adjust the leaf weights related to movement frequency. After training is completed, each node performs differential privacy encryption on the locally generated model update parameters and then uploads them back to the central data node through an encrypted channel.

[0066] Step S205 describes the behavior of the central node in collecting these encrypted model update parameters. The central data node waits for parameters to return from all normally participating local nodes, while setting a timeout threshold. Nodes that do not respond within the timeout period are considered temporarily offline or abnormal, and are temporarily skipped while the already received parameters are processed. The obtained parameter set is a direct product of the new round of training after the convergence condition is not met. In this way, the model can use the new information in the local data of each node to correct the current deficiencies of the global model, thereby making the global model more accurate in the next iteration.

[0067] S206: Based on the model update parameters obtained from the next round of iterative training of each local data node, perform model iterative training to obtain the risk assessment model.

[0068] After collecting the latest model update parameters from each local data node, the central data node re-executes parameter aggregation and convergence determination. Similar to the above process, the central node again uses a preset parameter aggregation algorithm to recalculate the aggregation weights for each node based on the amount, quality, and training accuracy of the newly reported data in this round. Then, it multiplies the encrypted model update parameters of each node by their corresponding weights and sums them to obtain the new round of global model update parameters, and performs the same convergence determination process. When the convergence condition is finally met, the central data node solidifies the global model update parameters obtained from the last aggregation into a formal risk assessment model. This model integrates the training results of all local nodes in the last iteration round, equivalent to the model performance obtained by training all local datasets centrally under ideal conditions.

[0069] In one possible implementation, after generating the risk assessment model, although a preset model convergence determination algorithm has determined that it meets the convergence condition, the convergence determination only indicates that the current training data has no significant effect on improving the performance and accuracy of the risk assessment model. The model may still have substandard performance or insufficient accuracy. Therefore, to address this issue, this embodiment pre-constructs a preset validation dataset to verify the model's accuracy, preventing situations where model training has converged but performance metrics still fail to meet requirements.

[0070] Specifically, the validation process for the risk assessment model is achieved through the following three steps: Step 1: Based on the risk assessment model, perform risk prediction according to the preset verification dataset associated with the target object to obtain multiple sets of predicted risk results; the preset verification dataset includes multiple sets of original risk data and risk type identifiers corresponding to each set of original risk data.

[0071] In this embodiment, the preset convergence determination algorithm typically terminates training based on the loss function value on the training data or the upper limit of the number of iterations. However, model convergence on the training data only means that its fitting error to the seen samples no longer decreases significantly; it does not fully guarantee that the model has good generalization ability in unseen real-world scenarios, nor does it directly reflect whether the model has met the performance indicators required by the business. To address this issue, this embodiment introduces an independent verification mechanism after model generation. Specifically, a preset verification dataset is pre-constructed. This dataset does not participate in any federated training process and is maintained uniformly by a third party or central node independent of each local node. The verification dataset contains multiple sets of raw risk data (e.g., historical user static and dynamic behavioral characteristics with real labels), and risk type identifiers corresponding to each set of data (e.g., high risk, medium risk, low risk, or specific incident labels). These data undergo the same de-identification and standardization preprocessing as the training data to ensure consistency in input format.

[0072] Step 2: Based on the multiple sets of predicted risk results and the associated risk type identifiers, evaluate the performance of the risk assessment model to obtain model performance indicators.

[0073] The first step in the validation process is to perform forward prediction on a pre-defined validation dataset using the generated risk assessment model. Each set of raw risk data from the validation dataset is input into the model, which then calculates and outputs the corresponding predicted risk result. Since the validation dataset already contains the true risk type identifier for each data point, the second step is to compare the model's prediction results with the true identifiers one by one, and to evaluate the model's performance by statistically analyzing various evaluation metrics. This embodiment supports the use of accuracy, precision, recall, and F1 score as model performance metrics. These metrics can reflect the model's classification or regression performance from different perspectives. For example, in medical insurance risk assessment, a high recall rate means that the model can effectively identify most high-risk individuals, avoiding missed diagnoses; while a high precision rate ensures that users marked as high-risk do indeed have a high probability of claiming, reducing underwriting costs caused by misjudgments. The calculated model performance metrics are one or more specific values, which the system compares with preset performance metric thresholds. These thresholds are usually set based on business requirements, such as requiring an accuracy of no less than 85% and a recall of no less than 80%. If all metrics reach or exceed the threshold, the model is considered ready for deployment; if any metric falls below the threshold, it means that although the current model has converged on the training data, its actual application performance is still unsatisfactory and requires further optimization.

[0074] Step 3: If the model performance index is less than the performance index threshold, send a data update instruction to each local data node to control each local data node to perform the local model update based on the updated local data.

[0075] When the model performance evaluation results show that one or more indicators are below a preset threshold, this embodiment provides an automatic response: In step three, a data update instruction is sent to each local data node. The purpose of this instruction is not simply to repeat the original training, but to require each node to re-execute the local model update using the updated local data. Here, the updated local data typically refers to two types of information: first, user data newly collected by each node that has not yet participated in training; and second, a data version that corrects possible labeling errors or formatting issues in the original data. By introducing fresh or higher-quality data, the model is expected to break through the performance bottleneck reached during the original training convergence. After receiving the data update instruction, each local data node will start from the latest global model parameters saved by the current central node, re-train locally using the updated local data, generate new model update parameters, and encrypt and upload them. The central node then executes the aggregation, convergence determination, and verification process again. This cycle can be repeated until the model's performance indicators on the verification dataset meet the business requirements. In this way, the model can be effectively prevented from prematurely stagnating in local optima or false convergence, ensuring that the final risk assessment model is both mathematically convergent and meets the expected business indicators, thereby truly meeting the practical application needs of high-frequency underwriting and accurate pricing in insurance.

[0076] This application provides an insurance risk assessment method based on sports and health data and federated learning. This method simultaneously collects static risk data and dynamic behavioral data of the target object, assessing the risk level of the target object from both inherent and behavioral risks, thereby reducing risk misjudgments caused by ignoring dynamic factors. Furthermore, before data fusion, the two types of data are separately quality-assessed to obtain static risk data quality scores and dynamic behavioral data quality scores. Weight analysis is then performed based on the type of the target insurance product and the quality scores to generate static risk data weights and dynamic behavioral data weights. This dual-driven mechanism ensures that the weights adapt to the business logic of the insurance scenario, and automatically reduces the weight of a certain type of data when its quality is low, avoiding negative impacts on the assessment results from low-quality data. Finally, the risk assessment model trained based on a preset federated learning framework can break down data silos, achieve multi-source privacy-preserving joint modeling, and improve generalization ability. The risk assessment model performs weighted calculations based on the above adaptive weights and the two types of data to output a risk score. This ensures that the assessment results accurately map to the true risk level while avoiding bias from manually fixed weights, thereby improving the accuracy of insurance risk assessment.

[0077] The following describes an insurance risk assessment system based on sports and health data and federated learning, provided by an embodiment of this application. The insurance risk assessment system based on sports and health data and federated learning described below can be referred to in correspondence with the insurance risk assessment method based on sports and health data and federated learning described above.

[0078] See Figure 3 The figure is a schematic diagram of the structure of an insurance risk assessment system based on sports health data and federated learning provided in an embodiment of this application, specifically including the following modules: The data acquisition module 100 is used to acquire static risk data and dynamic behavior data for the target object; the static risk data is used to characterize the static characteristics and historical state information of the target object; the dynamic behavior data is used to characterize the periodically changing monitoring parameters of the target object; The quality assessment module 200 is used to perform data quality assessment on the static risk data and the dynamic behavior data to obtain a static risk data quality score and a dynamic behavior data quality score. The weight analysis module 300 is used to perform weight analysis based on the insurance type of the target insurance product, according to the static risk data quality score and the dynamic behavior data quality score, to obtain the static risk data weight and the dynamic behavior data weight. The risk assessment module 400 is used to perform risk assessment based on a pre-built risk assessment model, according to the weights of the static risk data, the weights of the dynamic behavior data, the static risk data, and the dynamic behavior data, to obtain a risk assessment score for the target object; the risk assessment model is trained based on a preset federated learning architecture.

[0079] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, this application also provides a computer-readable storage medium storing computer instructions for causing the computer to execute the insurance risk assessment method based on sports health data and federated learning as described in any of the above embodiments.

[0080] The computer-readable media in this application embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0081] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the fault scenario screening method as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0082] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the system, method, and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The system, method, and medium embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0083] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An insurance risk assessment method based on sports and health data and federated learning, characterized in that, The method includes: Acquire static risk data and dynamic behavioral data for the target object; the static risk data is used to characterize the static characteristics and historical state information of the target object; the dynamic behavioral data is used to characterize the periodically changing monitoring parameters of the target object; the dynamic behavioral data includes exercise and health data; The static risk data and the dynamic behavior data are evaluated for data quality to obtain static risk data quality scores and dynamic behavior data quality scores. Based on the insurance type of the target insurance product, a weight analysis is performed according to the static risk data quality score and the dynamic behavior data quality score to obtain the static risk data weight and the dynamic behavior data weight. Based on a pre-built risk assessment model, risk assessment is performed according to the weights of the static risk data, the weights of the dynamic behavior data, the static risk data, and the dynamic behavior data to obtain a risk assessment score for the target object; the risk assessment model is trained based on a preset federated learning architecture.

2. The method according to claim 1, characterized in that, The preset federated learning architecture includes multiple local data nodes and a single central data node, and each of the local data nodes is configured with the same baseline risk assessment model. The local data node is used to train the benchmark risk assessment model locally using the GBDT algorithm and local data, and to send the updated parameters of the trained model to the central data node; the GBDT algorithm is a gradient boosting decision tree algorithm. The central data node is used to perform iterative training of the model based on the model update parameters from each of the local data nodes to obtain the risk assessment model.

3. The method according to claim 2, characterized in that, The step of iteratively training the model based on the model update parameters from each of the local data nodes to obtain the risk assessment model includes: Based on a preset parameter aggregation algorithm, the model update parameters obtained from training each local data node are aggregated to obtain global model update parameters; Based on a preset model convergence determination algorithm, the global model update parameters are converged; the preset model convergence determination algorithm includes convergence conditions. If the global model update parameters are determined to meet the convergence condition, the risk assessment model is generated based on the global model update parameters. If it is determined that the global model update parameters do not meet the convergence condition, the global model update parameters are sent to each of the local data nodes; Obtain the model update parameters obtained by each local data node in the next round of iterative training based on the global model update parameters; The risk assessment model is obtained by iteratively training the model based on the model update parameters obtained from the next round of training of each local data node.

4. The method according to claim 3, characterized in that, The preset parameter aggregation algorithm aggregates the model update parameters obtained from training each local data node to obtain global model update parameters, including: The sum of the products of the aggregate weights associated with each local data node and the corresponding model update parameters is determined as the global model update parameters; the aggregate weights are determined based on the proportion of data volume, the proportion of data quality score, and the proportion of training accuracy of each local data node. Wherein, the data volume ratio is the proportion of the training data volume of the local data node in the total training data of all the local data nodes; the data quality score ratio is the proportion of the data quality score of the local data node in the total data quality score of all the local data nodes; and the training accuracy ratio is the proportion of the local model training accuracy of the local data node in the total local model training accuracy of all the local data nodes.

5. The method according to claim 3, characterized in that, The convergence determination algorithm based on a preset model, which performs convergence determination on the global model update parameters, includes: Obtain the current iteration training number corresponding to the global model update parameters; Based on the mean squared error loss function, determine the loss function value of the model output corresponding to the global model update parameters; If the loss function value is not greater than the loss function threshold or the current number of training iterations is greater than the maximum number of iterations, it is determined that the global model update parameters satisfy the convergence condition. If the loss function value is greater than the loss function threshold and the current iteration training number is less than or equal to the maximum iteration number, it is determined that the global model update parameters do not meet the convergence condition.

6. The method according to claim 3, characterized in that, After generating the risk assessment model, the method further includes: Based on the risk assessment model, risk prediction is performed according to a preset verification dataset associated with the target object to obtain multiple sets of predicted risk results; the preset verification dataset includes multiple sets of original risk data and risk type identifiers corresponding to each set of original risk data; Based on the multiple sets of predicted risk results and the associated risk type identifiers, the risk assessment model is evaluated to obtain model performance indicators. If the model performance index is less than the performance index threshold, a data update instruction is sent to each of the local data nodes to control each of the local data nodes to perform the local model update based on the updated local data.

7. The method according to claim 1, characterized in that, The risk assessment model, based on a pre-built model, performs risk assessment according to the weights of static risk data, dynamic behavioral data, static risk data, and dynamic behavioral data to obtain a risk assessment score for the target object, including: Based on the risk assessment model, risk assessments are performed according to the static risk data and the dynamic behavioral data to obtain static risk scores and dynamic risk scores. The risk assessment score is determined by summing the product of the static risk data weight and the static risk score, and the product of the dynamic risk score and the dynamic behavioral data weight.

8. The method according to claim 2, characterized in that, Before sending the trained model update parameters to the central data node, the method further includes: Differential privacy encryption is applied to the model update parameters obtained from the training to obtain the encrypted model update parameters.

9. An insurance risk assessment system based on sports and health data and federated learning, characterized in that, The system includes: The data acquisition module is used to acquire static risk data and dynamic behavioral data for the target object; the static risk data is used to characterize the static characteristics and historical state information of the target object; the dynamic behavioral data is used to characterize the periodically changing monitoring parameters of the target object; the dynamic behavioral data includes exercise and health data; The quality assessment module is used to assess the data quality of the static risk data and the dynamic behavior data to obtain a static risk data quality score and a dynamic behavior data quality score. The weighting analysis module is used to perform weighting analysis based on the insurance type of the target insurance product, according to the static risk data quality score and the dynamic behavior data quality score, to obtain the static risk data weight and the dynamic behavior data weight. The risk assessment module is used to perform risk assessment based on a pre-built risk assessment model, according to the weights of the static risk data, the weights of the dynamic behavior data, the static risk data, and the dynamic behavior data, to obtain a risk assessment score for the target object; the risk assessment model is trained based on a preset federated learning architecture.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the insurance risk assessment method based on sports health data and federated learning as described in any one of claims 1-8.

Images