Model updating method and device for insurance data drift, equipment and medium

Abstract

The present application relates to the technical field of machine learning, and provides a model updating method and device for insurance data drift, equipment and medium. The method determines the latest distribution characteristics of the correlation index group between the input data group, the output data group and the preset insurance business field pair based on the latest insurance data sample, determines the distribution offset degree of each group by comparing the difference between the latest distribution characteristics and the historical distribution characteristics, and finally determines the insurance data drift level in combination with the distribution offset degree of each group of data, and triggers model updating. The present application can be applied to the financial field. By focusing on the distribution offset of the three dimensions of the input data, the output data and the correlation between different insurance business fields in the financial insurance business, the insurance data drift level is determined, so that the model updating is triggered in time, thereby ensuring the effectiveness and stability of the model when applied in the insurance business.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a model update method, apparatus, device, and medium for insurance data drift. Background Technology

[0002] With the deepening of digitalization, machine learning models have been widely applied in numerous fields such as finance, healthcare, and transportation. The performance of machine learning models is highly dependent on the distribution characteristics of the training data. When the distribution of the actual input data deviates significantly from the baseline distribution during model training, the model's prediction accuracy will continuously decline, and may even completely lose its effectiveness. Therefore, existing model update systems typically employ offset detection schemes that directly identify data drift by monitoring changes in the distribution of single or multiple input features, and then adapt the model to the changed data distribution through updates.

[0003] However, existing technologies mostly focus on general internet scenarios such as transportation and industry. In fields like finance and insurance, which have specific business characteristics and high-risk attributes, such offset detection methods have significant limitations and cannot meet the needs of specific scenarios. For example, in financial scenarios (such as insurance claims and credit risk control), existing detection methods judge drift by focusing on changes in the distribution of input features (such as changes in the age distribution of policyholders and the income distribution of loan applicants). However, in some fraud patterns, the distribution of many input features does not show significant shifts. The misjudgment of such drifts by existing technologies prevents the model from updating in a timely manner to address these fraud patterns, thus affecting the model's risk prediction performance.

[0004] Therefore, there is an urgent need to propose a new model update method for insurance data drift in financial scenarios, so as to accurately capture insurance data drift, improve model update efficiency, and ensure the stability and effectiveness of the model when applied in insurance business. Summary of the Invention

[0005] This invention provides a model update method, apparatus, device, and medium for addressing insurance data drift, to accurately capture insurance data drift, improve model update efficiency, and ensure the stability and effectiveness of the model when applied in insurance business.

[0006] Firstly, a model update method for insurance data is provided, including: Obtain multiple latest insurance data samples within the latest sliding time window. Each latest insurance data sample includes input data and output data of the student model, as well as correlation indicators between preset insurance business field pairs. The preset business field pairs exist in the input data and / or the output data. The student model is a task processing model deployed on the edge for insurance business. Based on all the latest insurance data samples, determine the latest distribution characteristics of the input data group, the latest distribution characteristics of the output data group, and the latest distribution characteristics of the correlation index group; The distribution offset is calculated based on the latest and historical distribution characteristics of the same group of data to obtain the distribution offset of each group of data. The historical distribution characteristics are determined based on historical insurance data samples within the historical sliding time window. Based on the distribution offset of each group of data, the insurance data drift level is determined, and the student model is updated according to the model update strategy corresponding to the insurance data drift level.

[0007] Secondly, a model update device for insurance data drift is provided, comprising: The acquisition module is used to acquire multiple latest insurance data samples within the latest sliding time window. Each latest insurance data sample includes the input data and output data of the student model, as well as the correlation index between preset insurance business field pairs. The preset insurance business field pairs exist in the input data and / or the output data. The student model is a task processing model deployed on the edge for insurance business. The determination module is used to determine the latest distribution characteristics of the input data group, the latest distribution characteristics of the output data group, and the latest distribution characteristics of the correlation index group based on all the latest insurance data samples. The calculation module is used to calculate the distribution offset based on the latest distribution characteristics and historical distribution characteristics of the same group of data, respectively, to obtain the distribution offset of each group of data. The historical distribution characteristics are determined based on historical insurance data samples within the historical sliding time window. The update module is used to determine the insurance data drift level based on the distribution offset of each group of data, and update the student model according to the model update strategy corresponding to the insurance data drift level.

[0008] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the model update method provided in the first aspect.

[0009] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the model update method provided in the first aspect.

[0010] The aforementioned scheme for model updates addressing insurance data drift first acquires multiple latest insurance data samples within the latest sliding time window. Each latest insurance data sample includes the input data, output data, and correlation indicators between preset insurance business field pairs of the edge-end student model. Then, based on all the latest insurance data samples, the latest distribution characteristics of the input data group, the output data group, and the correlation indicator group are determined. Next, by comparing the differences between the latest and historical distribution characteristics, the distribution offset of the input data group, the output data group, and the correlation indicator group is determined. Finally, the insurance data drift level is determined by combining the distribution offsets of each data group, and the corresponding model update strategy is triggered to update the student model. This application focuses on the data distribution offsets in three dimensions—input data, output data, and the correlation between preset insurance business field pairs of the edge-end student model—in the financial insurance scenario to determine the current insurance data drift level in the financial insurance scenario. When updating the model, we not only focus on the numerical changes in the input and output data, but also on the changes in the relationships between the preset insurance business field pairs in the input and / or output data. This allows us to identify changes in real insurance business scenarios and new financial risk patterns earlier and more accurately, triggering model updates to address these new changes and ensuring the effectiveness and stability of the model when applied in insurance business. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of an application environment for a model update method for insurance data drift in one embodiment of the present invention; Figure 2 This is a flowchart illustrating a model update method for insurance data drift in one embodiment of the present invention; Figure 3 This is a flowchart illustrating a model update method for insurance data drift in another embodiment of the present invention; Figure 4 This is a flowchart illustrating a model update method for insurance data drift in another embodiment of the present invention; Figure 5 This is a flowchart illustrating a model update method for insurance data drift according to another embodiment of the present invention. Figure 6This is a schematic diagram of a model update device for insurance data drift in another embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] This invention provides a model update method for insurance data drift, which can be applied to applications such as... Figure 1 In this application environment, the edge device communicates with the cloud via a network. A lightweight student model is deployed on the edge, while a complex teacher model is deployed on the cloud. Both the student and teacher models are task processing models deployed for insurance business. The edge device acquires multiple samples of the latest insurance data within the latest sliding time window. Each sample includes the input and output data of the edge's student model, as well as correlation indicators between preset insurance business field pairs. Based on all the latest insurance data samples, the edge device determines the latest distribution characteristics of the input data group, the output data group, and the correlation indicator group. Then, the edge device calculates the distribution offset based on the latest and historical distribution characteristics of the same data group. This yields the distribution offset for each data group. Based on the distribution offset of each data group, the insurance data drift level is determined, and the student model is updated according to the model update strategy corresponding to the insurance data drift level. When updating the student model, if edge-cloud collaborative distillation is required, the edge device will select abnormal insurance data samples from the latest insurance data samples and upload them to the cloud. The cloud then guides the edge device to perform model distillation based on these abnormal insurance data samples.

[0015] It should be noted that this is merely an illustrative example of the interaction process between the edge and the cloud, and is not intended to limit it.

[0016] The edge device can include, but is not limited to, various personal computers, laptops, and other devices. The cloud can be implemented using a dedicated server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.

[0017] Please see Figure 2 As shown, Figure 2A flowchart illustrating a model update method for insurance data drift provided in an embodiment of the present invention includes the following steps: S201, Obtain multiple latest insurance data samples within the latest sliding time window. Each latest insurance data sample includes the input data and output data of the student model, as well as the correlation index between preset insurance business field pairs. The preset insurance business field pairs exist in the input data and / or output data. The student model is a task processing model deployed on the edge for insurance business.

[0018] For example, a sliding time window can be set. The window length and sliding step size of the sliding time window can be set as needed, and this application does not impose any restrictions.

[0019] Within the latest sliding time window, the edge collects the input data that needs to be input into the student model, and collects the output data obtained by the student model after processing these input data, as well as calculates the correlation index between preset insurance business field pairs in the input data and / or output data.

[0020] Each input data set, along with its corresponding output data and correlation indicators, constitutes a sample of the latest insurance data.

[0021] For example, within the latest sliding time window (e.g., the last 30 days, with a sliding step of 1 day), the edge device collects claim application-related data that needs to be input into the student model. This data may include basic information about the policyholder, policy information, incident information, and medical treatment information. After processing the above input data, the student model outputs data, which may include the claim risk level or fraud probability value. The specific content of the output data is determined by the functionality of the student model; this is merely an example. For instance, the student model could be a risk prediction model or a claim amount prediction model deployed for insurance business, etc.

[0022] Pre-defined insurance business field pairs that need attention can be determined in advance based on the financial insurance scenario, such as the sum insured and the amount of medical expenses, the location of the accident and the permanent residence of the waiver, the severity of the accident and the amount of compensation, etc.

[0023] The correlation indicators for the pre-defined insurance business field pairs may include the Pearson correlation coefficient, the ratio between the two, or the degree of correlation, etc., which are not limited in this application.

[0024] S202, based on all the latest insurance data samples, determine the latest distribution characteristics of the input data group, the latest distribution characteristics of the output data group, and the latest distribution characteristics of the correlation index group.

[0025] For example, the input data in all the latest insurance data samples constitutes the input data group, the output data in all the latest insurance data samples constitutes the output data group, and the correlation indicators in all the latest insurance data samples constitute the correlation indicator group.

[0026] Perform statistical distribution analysis on each data set to determine the latest distribution characteristics of each data set.

[0027] In one example, the latest distribution features may include distribution parameters of continuous distributions such as mean and variance, and / or distribution parameters of discrete distributions such as class frequency and number of classes.

[0028] For example, when statistically analyzing the distribution of continuous data, variance, mean, and quantiles can be used as distribution parameters, or parameters of distribution models such as positive distribution and uniform distribution can be selected; when statistically analyzing the distribution of discrete data, category probability or proportion can be used as distribution parameters.

[0029] The input data set may include multiple insurance business fields, such as transaction amount, transaction type, etc. For example, after performing statistical distribution analysis on the input data set, the distribution parameters for the continuous data "transaction amount" are a mean of 500 yuan and a variance of 10000; the distribution parameters for the discrete data "transaction type" are 0.4 for transfer, 0.1 for wealth management, and 0.5 for consumption.

[0030] It should be noted that the analytical methods used in statistical analysis, such as the mean, distribution model, or frequency, can be set as needed, and this application does not impose any restrictions.

[0031] S203, calculate the distribution offset based on the latest and historical distribution characteristics of the same group of data respectively, and obtain the distribution offset of each group of data. The historical distribution characteristics are determined based on historical insurance data samples within the historical sliding time window.

[0032] For example, the historical sliding time window can be the window preceding the latest sliding time window. Historical insurance data samples are collected from the historical sliding time window. The distribution characteristics of the input data in these historical insurance data samples are the historical distribution characteristics of the input data group. Similarly, the distribution characteristics of the output data in these historical insurance data samples are the historical distribution characteristics of the output data group. Likewise, the distribution characteristics of the correlation indicators in these historical insurance data samples are the historical distribution characteristics of the correlation indicator group.

[0033] For each data set, calculate the distribution offset by examining the latest and historical distribution characteristics of the same field. Then, determine the overall distribution offset of the data set based on the distribution offsets of all fields within that data set.

[0034] In one example, the distribution offset can be calculated using any of the following methods: KL divergence, JS divergence, Wasserstein distance, or PSI (population stability index). This application does not impose any limitation on these methods.

[0035] S204. Based on the distribution offset of each group of data, determine the insurance data drift level, and update the student model according to the model update strategy corresponding to the insurance data drift level.

[0036] For example, the distribution offset of the input data group, the distribution offset of the output data group, and the numerical offset of the correlation index group are weighted and summed, and the degree of insurance data drift is determined according to the result of the weighted summation, that is, the above-mentioned insurance data drift level.

[0037] In the weighted summation process, the weights corresponding to the input data group, the output data group, and the correlation index group can be obtained by training the cloud-based teacher model and then sent to the edge to inform the edge to focus on which data distribution offset is the primary focus and which data distribution offset is the secondary focus in the data offset detection.

[0038] In some embodiments, the model update strategy includes: The strategy for determining whether to enable edge-cloud collaborative distillation, and / or the strategy for adjusting the number of abnormal insurance data samples uploaded, and / or the strategy for adjusting the update intensity when the edge performs local updates to the student model.

[0039] In one example, insurance data drift levels can be categorized as stable, slight drift, significant drift, and severe drift.

[0040] When the system is in a stable state, edge-cloud collaborative distillation is not enabled, the number of abnormal insurance data samples uploaded is zero, and the student model is updated locally and incrementally at the edge.

[0041] When the data drift is at a mild level, edge-cloud collaborative distillation is enabled. A small number of anomalous insurance data samples (e.g., samples with low confidence or large residuals in correlation indicators) are uploaded to the cloud. This allows the cloud-based teacher model to output high-precision prediction results as pseudo-labels for the anomalous insurance data samples, and these pseudo-labels and anomalous insurance data samples are included in a "local distillation set." The edge-based student model uses the local distillation set distributed from the cloud to perform a local distillation update, thereby improving the student model's sensitivity to the new business environment after insurance data drift.

[0042] When the model exhibits significant drift, edge-cloud collaborative distillation is initiated. All anomalous insurance data samples are uploaded to the cloud, while the edge processor reduces its local learning rate or pauses regular online updates to avoid continuing to use outdated prediction strategies for self-reinforcement updates when the model is severely shifted. During edge-cloud collaborative distillation, the cloud directly pushes new model parameters trained based on the anomalous insurance samples to the edge processor, which then performs batch fine-tuning in the background. This fine-tuning can utilize a dual-model structure: the old model is served online, while the new model is fine-tuned in the background. A smooth switch between the old and new models is performed after adjustments to ensure uninterrupted business operations.

[0043] When the system is in a state of severe drift, initiate edge-cloud collaborative distillation, upload all the latest insurance data samples to the cloud, pause the update of local model parameters on the edge, and wait for the cloud to complete large-scale retraining before distributing the new model.

[0044] To make it easier to understand, here are some examples of situations that may trigger insurance data drift: In the financial sector, new large-scale fraud patterns emerge, such as fraudsters using a combination of low coverage with high payouts and insurance policies issued at temporary addresses in different locations. Within the latest sliding time window, the data input into the edge risk control model shows no significant shift in the latest and historical distributions of payout amounts and incident locations. Therefore, existing technologies, which only detect changes in the distribution of single features, struggle to identify these new fraud patterns. This application focuses on the relationships between different fields, specifically the correlation indicators between pre-defined business field pairs. It captures changes in the distribution of the payout amount to the coverage ratio and the correlation between the incident location and the policyholder's permanent residence. This allows for the identification of drift in the relationship between payout amount and coverage, as well as shifts in the relationship between the incident location and the policyholder's permanent residence. This triggers model updates, enabling the model to learn the identification logic for these new fraud patterns from sample data, thereby improving risk control accuracy. Existing technologies cannot capture the drift in relationships between the aforementioned business fields, which may lead to the incorrect determination that no data drift has occurred. Consequently, the model will not be updated, and the model will still rely on the old logic learned during the training phase to make risk control predictions. This makes it difficult to effectively identify fraudulent behavior and creates risk control vulnerabilities.

[0045] In summary, the model update method for insurance data drift provided in this application first obtains multiple latest insurance data samples within the latest sliding time window. Each latest insurance data sample includes the input data, output data, and correlation indicators between preset insurance business field pairs of the edge student model. Then, based on all the latest insurance data samples, the latest distribution characteristics of the input data group, the output data group, and the correlation indicator group are determined. Next, by comparing the differences between the latest distribution characteristics and historical distribution characteristics, the distribution offset of the input data group, the output data group, and the correlation indicator group is determined. Finally, the data drift level is determined by combining the distribution offsets of each group of data, and the corresponding model update strategy is triggered to update the student model. This application determines the insurance data drift level by focusing on the data distribution offsets of the input data, output data, and the correlation between preset insurance business field pairs of the edge student model. When updating the model, we not only focus on the numerical changes in the input and output data, but also on the changes in the relationships between the preset insurance business field pairs in the input and / or output data. This allows us to identify changes in real insurance business scenarios and new financial risk patterns earlier and more accurately, triggering the model to update in a timely manner in response to these new changes, thereby ensuring the effectiveness and stability of the model when applied in insurance business.

[0046] In some embodiments, such as Figure 3 As shown, step S203 above, "calculating the distribution offset based on the latest and historical distribution characteristics of the same group of data to obtain the distribution offset of each group of data," may include the following steps: S301, calculate the distribution offset between the latest distribution characteristics and historical distribution characteristics of the same insurance business field using a distance metric.

[0047] For example, the input data set includes multiple insurance business fields, and the latest distribution characteristics of the input data set include the latest distribution characteristics of all insurance business fields. The input data can be determined based on the processing task performed by the student model. For instance, if the student model performs a claims review task, the aforementioned insurance business fields may include: policyholder name, ID number, policy number, insured amount, number of claims, vehicle fault information required for vehicle insurance claims, etc.

[0048] Distance metrics can be used to quantify the difference or similarity between two objects (e.g., two distributions), and can be implemented by any method such as calculating the relative entropy between the two objects, calculating the JS divergence, or the Wasserstein distance, without limitation in this application.

[0049] For the same insurance business field, the distribution offset of the field is obtained by measuring the distance between the latest and historical distribution characteristics of the field.

[0050] S302, determine the distribution offset of the input data group based on the distribution offset of all insurance business fields.

[0051] For example, after determining the distribution offset of all insurance business fields in the input data group, the distribution offsets of all insurance business fields can be weighted and summed, and the summed result can be used as the distribution offset of the input data group.

[0052] The weights of the weighted summation can be preset manually according to business rules, or the importance of each field in the business rules can be determined through large model training to obtain the weight of that field.

[0053] In this embodiment, the distribution offset of each insurance business field in the input data is calculated first, and then the distribution offset of the entire input data group is determined based on the distribution offset of all insurance business fields. This allows attention to the offset of each insurance business field in the input data in terms of distribution, so that the overall offset of the input data can be taken into account in the subsequent determination of the insurance data offset level, thereby improving the accuracy of the determined data offset level.

[0054] In some embodiments, the latest distribution characteristics of the output data group include the latest distribution characteristics of the prediction task processing results output by the student model. Step S203 above, "calculating the distribution offset based on the latest and historical distribution characteristics of the same data group to obtain the distribution offset of each data group," includes: Using a distance metric, the distribution offset between the latest and historical distribution characteristics of the prediction task processing results is calculated to obtain the distribution offset of the output data set.

[0055] For example, the student model can be a categorical prediction model or a regression prediction model.

[0056] For example, a categorized prediction model can process input data and output a prediction result that is either low-risk, medium-risk, or high-risk.

[0057] When determining the latest distribution characteristics of the output data set, first calculate the distribution offset between the latest distribution characteristics and the historical distribution characteristics of the prediction task processing results, and use the distribution offset of the prediction task processing results as the distribution offset of the output data.

[0058] In some embodiments, such as Figure 4As shown, step S203 above, "calculating the distribution offset based on the latest and historical distribution characteristics of the same group of data to obtain the distribution offset of each group of data," may include the following steps: S401, for each preset insurance business field pair, calculate the absolute value of the difference between the latest distribution characteristics and the historical distribution characteristics of the correlation index.

[0059] S402, determine the distribution offset of the correlation indicator group based on the absolute values ​​of all preset insurance business fields.

[0060] For example, if there are multiple preset insurance business field pairs, first determine the distribution offset of the correlation distribution index between each preset insurance business field pair, and then integrate all the distribution offsets to obtain the distribution offset of the correlation index group.

[0061] For example, in an insurance claims scenario, the severity of the accident in the input data and the compensation amount in the output data are set as a preset insurance business field pair.

[0062] If the sliding time window is the last seven days, then the latest distribution characteristics of the preset insurance business field pairs are the mean of the correlation indicators of the field pairs within the last seven days; the historical distribution characteristics of the preset insurance business field pairs are the mean of the correlation indicators of the field pairs within the previous seven days. Calculating the absolute value of the difference between the two means yields the distribution offset of the correlation indicators between the preset insurance business field pairs.

[0063] Then, the distribution offset of the correlation indicators between all the preset insurance business fields is weighted and summed to obtain the distribution offset of the correlation indicator group.

[0064] The weights for the weighted summation can be preset manually according to insurance business rules, or the importance of each field in the business rules can be determined through large-scale model training to obtain the weight of that field.

[0065] In this embodiment, the distribution offset of the correlation index between each preset insurance business field pair is first calculated. Then, based on the distribution offset of the correlation index between all preset insurance business field pairs, the distribution offset of the correlation index group is determined. This allows attention to the distribution offset of the correlation index between each preset insurance business field pair, so that the offset of the correlation degree between each preset insurance business field pair can be taken into account in the subsequent determination of the insurance data offset level, thereby improving the accuracy of the determined insurance data offset level.

[0066] In some embodiments, such as Figure 5 As shown, step S204 above, "determining the data drift level based on the distribution offset of each group of data," may include the following steps: S501: Obtain the confidence scores of the student models corresponding to all the latest insurance data samples, and calculate the average confidence scores of all student models.

[0067] For example, input data from the latest insurance data sample is fed into the student model, which then outputs the prediction task processing result and the confidence level of that result. For ease of description, this application refers to this confidence level as the student model confidence level.

[0068] Each latest insurance data sample corresponds to a student model confidence score, and the average of all student model confidence scores is calculated.

[0069] S502, if the average value is lower than the preset reliability threshold, the latest insurance data sample with a student model confidence level lower than the preset reliability threshold is taken as an abnormal insurance data sample.

[0070] If the average confidence level of all student models is lower than the preset confidence threshold, then samples with student model confidence levels lower than the preset confidence threshold will be selected from all the latest insurance data samples and these samples will be regarded as abnormal insurance data samples.

[0071] S503 uploads abnormal insurance data samples to the cloud so that the cloud-based teacher model can process the input data in each abnormal insurance data sample and obtain the teacher model output distribution. The teacher model is a task processing model deployed in the cloud for insurance business.

[0072] For example, the edge device uploads abnormal insurance data samples to the cloud and inputs the input data of each abnormal insurance data sample into the teacher model deployed in the cloud. The teacher model processes the input data and obtains the teacher model output distribution.

[0073] The teacher model output distribution can be understood as the probability distribution of each prediction result output by the teacher model.

[0074] For example, if the teacher model is a classification model, the model may output a prediction result that is positive or negative. Therefore, after inputting an abnormal insurance data sample into the teacher model, the resulting output distribution includes the probability of a positive output and the probability of a negative output. For example, the probability of a positive output is 0.8, and the probability of a negative output is 0.2.

[0075] S504, Obtain the student model output distribution after processing the input data of the abnormal insurance data sample.

[0076] For example, the edge processor inputs the input data from the abnormal insurance data sample into the student model to obtain the student model's output distribution. For details on the student model's output distribution, please refer to the explanation of the teacher model's output distribution; it will not be repeated here. For instance, the student model's output distribution corresponding to an abnormal insurance data sample might include a positive output probability of 0.4 and a negative output probability of 0.6.

[0077] S505, determine the degree of difference between the student model output distribution and the teacher model output distribution corresponding to the abnormal insurance data samples.

[0078] In some embodiments, for the same abnormal insurance data sample, the relative entropy between the student model output distribution and the teacher model output distribution is calculated; the average value of the relative entropy corresponding to all abnormal insurance data samples is obtained to obtain the degree of difference.

[0079] Alternatively, other distance metrics mentioned above can be used to determine the degree of difference in the distribution of teacher and student outputs; this application does not impose any restrictions.

[0080] S506, the distribution offset of the input data group, the distribution offset of the output data group, the distribution offset of the correlation index group, and the difference are superimposed to determine the comprehensive offset index.

[0081] For example, the distribution difference of the input data group, the distribution offset of the output data group, the distribution offset of the correlation index group, and the difference are weighted and summed to obtain the comprehensive offset index.

[0082] In this embodiment, the method for obtaining the weights in the weighted summation can refer to the description of the weight acquisition method in the above embodiments, and will not be repeated here.

[0083] S507 determines the insurance data drift level based on the comprehensive offset index.

[0084] For example, offset threshold ranges can be predefined for different insurance data offset levels. After determining the comprehensive offset index, the insurance data drift level corresponding to the offset threshold range where the comprehensive offset index is located is determined as the insurance data drift level finally determined in this application embodiment, that is, the insurance data drift level corresponding to the latest insurance data sample.

[0085] In this embodiment, abnormal insurance data samples are screened using confidence levels and uploaded to the cloud. The teacher model deployed in the cloud determines the teacher model's output distribution based on these abnormal insurance data samples. Then, the edge computing unit calculates the difference between the teacher model's output distribution and the student model's output distribution to obtain the probability of various output results obtained by the student model and the teacher model. This quantifies the difference in predictive ability between the teacher model and the student model. When determining the level of insurance data drift, the difference between the teacher and student output distributions is added on top of the distribution offset of each group of data, thereby improving the detection accuracy of insurance data drift and effectively triggering model updates.

[0086] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0087] In one embodiment, a model update apparatus for insurance data drift is provided, which corresponds one-to-one with the model update method in the above embodiments. For example... Figure 6 As shown, the model update device 600 includes: an acquisition module 601, a determination module 602, a calculation module 603, and an update module 604. Detailed descriptions of each functional module are as follows: The acquisition module 601 is used to acquire multiple latest insurance data samples within the latest sliding time window. Each latest insurance data sample includes the input data and output data of the edge student model, as well as the correlation index between preset insurance business field pairs. The preset insurance business field pairs exist in the input data and / or output data. The student model is a task processing model deployed on the edge for insurance business. The determination module 602 is used to determine the latest distribution characteristics of the input data group, the latest distribution characteristics of the output data group, and the latest distribution characteristics of the correlation index group based on all the latest insurance data samples. The calculation module 603 is used to calculate the distribution offset based on the latest distribution characteristics and historical distribution characteristics of the same group of data, respectively, to obtain the distribution offset of each group of data. The historical distribution characteristics are determined based on historical insurance data samples within the historical sliding time window. The update module 604 is used to determine the insurance data drift level based on the distribution offset of each group of data, and update the student model according to the model update strategy corresponding to the insurance data drift level.

[0088] In some embodiments, the input data group includes multiple insurance business fields, and the latest distribution characteristics of the input data group include the latest distribution characteristics of all insurance business fields. The calculation module 603 is specifically used for: Calculate the distribution offset between the latest and historical distribution characteristics of the same insurance business field using a distance metric. Determine the distribution offset of the input data group based on the distribution offset of all insurance business fields.

[0089] In some embodiments, the latest distribution characteristics of the output data set include the latest distribution characteristics of the prediction task processing results output by the student model, and the calculation module 603 is specifically used for: Using a distance metric, the distribution offset between the latest and historical distribution characteristics of the prediction task processing results is calculated to obtain the distribution offset of the output data set.

[0090] In some embodiments, the calculation module 603 is specifically used for: For each preset insurance business field pair, calculate the absolute value of the difference between the latest distribution characteristics and the historical distribution characteristics of the correlation index; The distribution offset of the correlation indicator group is determined based on the absolute values ​​of all preset insurance business fields.

[0091] In some embodiments, the determining module 604 is specifically used for: Obtain the confidence scores of the student models corresponding to all the latest insurance data samples, and calculate the average confidence scores of all student models; If the average value is lower than the preset reliability threshold, the latest insurance data sample with a student model confidence level lower than the preset reliability threshold will be regarded as an abnormal insurance data sample. Abnormal insurance data samples are uploaded to the cloud so that the cloud-based teacher model can process the input data in each abnormal insurance data sample and obtain the output distribution of the teacher model. The teacher model is a task processing model deployed in the cloud for insurance business. The student model output distribution is obtained after processing the input data of the abnormal insurance data sample. Based on the student model output distribution and teacher model output distribution corresponding to the abnormal insurance data samples, determine the degree of difference between the teacher and student output distributions; The distribution offset of the input data group, the distribution offset of the output data group, the distribution offset of the correlation index group, and the difference are superimposed to determine the comprehensive offset index. The level of insurance data drift is determined based on the comprehensive offset index.

[0092] In some embodiments, the model update strategy includes: The strategy for determining whether to enable edge-cloud collaborative distillation, and / or the strategy for adjusting the number of abnormal insurance data samples uploaded, and / or the strategy for adjusting the update intensity when the edge performs local updates to the student model.

[0093] In summary, the model update device provided in this application first acquires multiple latest insurance data samples within the latest sliding time window. Each latest insurance data sample includes the input data, output data, and correlation indicators between preset insurance business field pairs of the edge student model. Then, based on all the latest insurance data samples, it determines the latest distribution characteristics of the input data group, the output data group, and the correlation indicator group. Next, by comparing the differences between the latest distribution characteristics and historical distribution characteristics, it determines the distribution offset of the input data group, the output data group, and the correlation indicator group. Finally, it combines the distribution offsets of each data group to determine the insurance data drift level and triggers the corresponding model update strategy to update the student model. This application determines the insurance data drift level in the current financial insurance scenario by focusing on the data distribution offsets in three dimensions: the input data, output data, and the correlation between preset insurance business field pairs of the edge student model. When updating the model, we not only focus on the numerical changes in the input and output data, but also on the changes in the relationships between the preset insurance business field pairs in the input and / or output data. This allows us to identify changes in real insurance business scenarios and new financial risk patterns earlier and more accurately, triggering model updates to address these new changes and ensuring the effectiveness and stability of the model when applied in insurance business.

[0094] Specific limitations regarding the model update device can be found in the limitations of the model update method above, and will not be repeated here. Each module in the aforementioned model update device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0095] In one embodiment, a computer device is provided, such as Figure 7 As shown, the computer device 700 includes a memory 701, a processor 702, and a computer program stored in the memory 701 and executable on the processor 702. When the processor 702 executes the computer program, it implements the steps of the model update method of any of the above embodiments.

[0096] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the model update method of any of the above embodiments.

[0097] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the description of each step of the data risk prediction method based on artificial intelligence in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0098] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0099] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0100] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A model update method for insurance data drift, characterized in that, The method includes: Obtain multiple latest insurance data samples within the latest sliding time window. Each latest insurance data sample includes input data and output data of the student model, as well as correlation indicators between preset insurance business field pairs. The preset insurance business field pairs exist in the input data and / or the output data. The student model is a task processing model deployed on the edge for insurance business. Based on all the latest insurance data samples, determine the latest distribution characteristics of the input data group, the latest distribution characteristics of the output data group, and the latest distribution characteristics of the correlation index group; The distribution offset is calculated based on the latest and historical distribution characteristics of the same group of data to obtain the distribution offset of each group of data. The historical distribution characteristics are determined based on historical insurance data samples within the historical sliding time window. Based on the distribution offset of each group of data, the insurance data drift level is determined, and the student model is updated according to the model update strategy corresponding to the insurance data drift level.

2. The method according to claim 1, characterized in that, The input data group includes multiple insurance business fields. The latest distribution feature of the input data group includes the latest distribution features of all the insurance business fields. The step of calculating the distribution offset based on the latest distribution feature and historical distribution feature of the same group of data to obtain the distribution offset of each group of data includes: Calculate the distribution offset between the latest distribution feature and the historical distribution feature of the same insurance business field using a distance metric. The distribution offset of the input data group is determined based on the distribution offset of all the insurance business fields.

3. The method according to claim 1, characterized in that, The latest distribution characteristics of the output data group include the latest distribution characteristics of the prediction task processing results output by the student model. The distribution offset is calculated based on the latest and historical distribution characteristics of the same data group to obtain the distribution offset of each data group, including: The distribution offset between the latest distribution features and the historical distribution features of the prediction task processing results is calculated using a distance metric to obtain the distribution offset of the output data group.

4. The method according to claim 1, characterized in that, The step of calculating the distribution offset based on the latest and historical distribution characteristics of the same group of data to obtain the distribution offset of each group of data includes: For each of the preset insurance business field pairs, calculate the absolute value of the difference between the latest distribution characteristics and the historical distribution characteristics of the correlation index; The distribution offset of the correlation index group is determined based on the absolute values ​​corresponding to all the preset insurance business fields.

5. The method according to claim 1, characterized in that, The determination of the insurance data drift level based on the distribution offset of each group of data includes: Obtain the confidence scores of the student models corresponding to all the latest insurance data samples, and calculate the average confidence scores of all the student models; If the average value is lower than a preset confidence threshold, the latest insurance data sample with a student model confidence level lower than the preset confidence threshold will be regarded as an abnormal insurance data sample. The abnormal insurance data samples are uploaded to the cloud so that the cloud teacher model can process the input data in each abnormal insurance data sample to obtain the teacher model output distribution. The teacher model is a task processing model deployed in the cloud for insurance business. The student model output distribution is obtained by processing the input data of the abnormal insurance data sample using the student model. Based on the student model output distribution and the teacher model output distribution corresponding to the abnormal insurance data samples, determine the degree of difference between the teacher and student output distributions; The distribution offset of the input data group, the distribution offset of the output data group, the distribution offset of the correlation index group, and the difference are superimposed to determine the comprehensive offset index; The insurance data drift level is determined based on the comprehensive offset index.

6. The method according to claim 5, characterized in that, The step of determining the difference between the student and teacher output distributions based on the student model output distribution and the teacher model output distribution corresponding to the abnormal insurance data samples includes: For the same abnormal insurance data sample, calculate the relative entropy between the student model output distribution and the teacher model output distribution; The average value of the relative entropy corresponding to all the abnormal insurance data samples is used to obtain the degree of difference.

7. The method according to any one of claims 1-6, characterized in that, The model update strategy includes: The strategy for determining whether to enable edge-cloud collaborative distillation, and / or the strategy for adjusting the number of abnormal insurance data samples uploaded, and / or the strategy for adjusting the update intensity when the edge performs local updates on the student model.

8. A model update device for insurance data drift, characterized in that, The device includes: The acquisition module is used to acquire multiple latest insurance data samples within the latest sliding time window. Each latest insurance data sample includes the input data and output data of the student model, as well as the correlation index between preset insurance business field pairs. The preset insurance business field pairs exist in the input data and / or the output data. The student model is a task processing model deployed on the edge for insurance business. The determination module is used to determine the latest distribution characteristics of the input data group, the latest distribution characteristics of the output data group, and the latest distribution characteristics of the correlation index group based on all the latest insurance data samples. The calculation module is used to calculate the distribution offset based on the latest distribution characteristics and historical distribution characteristics of the same group of data, respectively, to obtain the distribution offset of each group of data. The historical distribution characteristics are determined based on historical insurance data samples within the historical sliding time window. The update module is used to determine the insurance data drift level based on the distribution offset of each group of data, and update the student model according to the model update strategy corresponding to the insurance data drift level.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the model update method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the model update method as described in any one of claims 1 to 7.

Images