Insurance underwriting method and device, computer equipment and storage medium

Abstract

The application relates to the technical field of artificial intelligence, and discloses an insurance underwriting method and device, computer equipment and a storage medium, which comprise the following steps: obtaining sample insurance business data of multiple different data source types, performing coding processing on the sample insurance business data of each data source type through an original time sequence feature extraction model to obtain a sample time sequence coding feature sequence of each data source type; training the original time sequence feature extraction model based on the sample time sequence coding feature sequence to obtain mask prediction loss data and time sequence alignment loss data; performing model optimization on the original time sequence feature extraction model based on the mask prediction loss data and the time sequence alignment loss data to obtain a target time sequence feature extraction model; performing feature extraction on policy time sequence data through the target time sequence feature extraction model to obtain policy feature data; and performing underwriting based on the policy feature data. The application can be applied to the insurance underwriting scene of financial technology and improves the accuracy of insurance underwriting.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and is applicable to the financial technology field, particularly to an insurance underwriting method, apparatus, computer equipment, and storage medium. Background Technology

[0002] Insurance underwriting methods can be used to uniformly model and analyze multi-source time-series data such as customer behavior logs, policy lifecycle data, claims records, and financial indicators in order to conduct risk review and assessment of insurance applications.

[0003] Currently, insurance underwriting primarily relies on human experience to perform feature engineering on multi-source time-series data, followed by analysis and prediction of the extracted key information. However, in insurance scenarios, data sources such as customer behavior logs, policy data, and claims records are diverse and have significantly different time granularities. Typically, data of different granularities needs to be forcibly aligned to a unified time scale before modeling. This modeling process may result in the loss of crucial time-dynamic features, such as the difference between short-term fluctuations and long-term trends. This makes it difficult to accurately extract key information related to insurance risks from multi-source time-series data, impacting the accuracy of insurance underwriting. Summary of the Invention

[0004] This application provides an insurance underwriting method, apparatus, computer equipment, and storage medium to solve the technical problem that feature engineering relying on human experience makes it difficult to accurately extract key information related to insurance risks from multi-source time-series data, thus affecting the accuracy of insurance underwriting.

[0005] Firstly, an insurance underwriting method is provided, including: Multiple sample insurance business data from different data sources are acquired, and the sample insurance business data from each data source type are encoded using a preset original time-series feature extraction model to obtain a sample time-series encoded feature sequence for each data source type. Based on the sample temporal coding feature sequence, the original temporal feature extraction model is trained for temporal feature prediction to obtain mask prediction loss data; Based on the sample temporal coding feature sequence, the original temporal feature extraction model is trained for temporal alignment to obtain temporal alignment loss data; Based on the mask prediction loss data and the temporal alignment loss data, the original temporal feature extraction model is optimized to obtain the target temporal feature extraction model. The target time-series feature extraction model is used to extract features from the pre-acquired policy time-series data to obtain policy feature data. Underwriting is performed based on the policy feature data.

[0006] Secondly, an insurance underwriting device is provided, comprising: The business data acquisition module is used to acquire sample insurance business data from multiple different data sources. It encodes the sample insurance business data from each data source type using a preset original time-series feature extraction model to obtain a sample time-series encoded feature sequence for each data source type. The temporal feature prediction training module is used to train the original temporal feature extraction model for temporal feature prediction based on the sample temporal coding feature sequence, and to obtain mask prediction loss data. The temporal alignment training module is used to perform temporal alignment training on the original temporal feature extraction model based on the sample temporal coding feature sequence to obtain temporal alignment loss data. The model optimization module is used to optimize the original temporal feature extraction model based on the mask prediction loss data and the temporal alignment loss data to obtain the target temporal feature extraction model. The policy feature extraction module is used to extract features from the pre-acquired policy time series data through the target time series feature extraction model to obtain policy feature data; The underwriting module is used to perform underwriting based on the policy feature data.

[0007] 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 above-described insurance underwriting method.

[0008] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described insurance underwriting method.

[0009] The above-mentioned insurance underwriting method, device, computer equipment, and storage medium can be applied to insurance underwriting scenarios in financial technology. First, sample insurance business data from multiple different data sources are acquired. Then, the sample insurance business data from each data source type is encoded using a preset original time-series feature extraction model to obtain a sample time-series encoded feature sequence for each data source type. This provides a rich and structured feature foundation for subsequent training, which helps to accurately capture the dependencies across time steps during subsequent processing and improves the ability of the time-series encoded feature sequence to express the dynamic changes in insurance business. Next, the original temporal feature extraction model is trained for temporal feature prediction based on the sample temporal encoded feature sequence to obtain mask prediction loss data. The original temporal feature extraction model is then trained for temporal alignment based on the sample temporal encoded feature sequence to obtain temporal alignment loss data. Based on the mask prediction loss data and the temporal alignment loss data, the original temporal feature extraction model is optimized to obtain the target temporal feature extraction model. This enables the model to learn the inherent patterns and contextual dependencies of temporal data rather than memorizing surface patterns, and to understand the same risk event across time scales. It avoids feature ambiguity caused by different granularities and improves the usability, generalization ability, and expressive power of features in different scenarios. Finally, the pre-acquired policy time-series data is extracted using a target time-series feature extraction model to obtain policy feature data. Underwriting is then performed based on this policy feature data. Since the obtained policy feature data has both strong semantic expression and multi-source time-series alignment characteristics, it can comprehensively and accurately depict the risk profile of the policyholder. This solves the technical problem that the feature engineering relying on human experience is unable to accurately extract key information related to insurance risks from multi-source time-series data, which affects the accuracy of insurance underwriting. This helps to improve the accuracy and efficiency of underwriting decisions and reduce the cost of manual review and the risk of misjudgment. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application 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.

[0011] Figure 1 This is a schematic diagram of an application environment for an insurance underwriting method according to one embodiment of this application; Figure 2 This is a flowchart illustrating an insurance underwriting method in one embodiment of this application; Figure 3 yes Figure 2 A schematic diagram of a specific implementation method for step S10; Figure 4 yes Figure 2A schematic diagram of a specific implementation method for step S20; Figure 5 yes Figure 2 A schematic diagram of a specific implementation method for step S30; Figure 6 yes Figure 5 A flowchart illustrating a specific implementation of step S31; Figure 7 This is a schematic diagram of the structure of an insurance underwriting device in one embodiment of this application; Figure 8 This is a schematic diagram of the structure of a computer device according to one embodiment of this application. Detailed Implementation

[0012] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0013] First, let's analyze some of the terms used in this application: Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0014] Natural Language Processing (NLP): NLP uses computers to process, understand, and utilize human language (such as Chinese and English). It is a branch of artificial intelligence and an interdisciplinary field of computer science and linguistics, often referred to as computational linguistics. NLP includes syntactic analysis, semantic analysis, and discourse understanding. It is commonly used in machine translation, handwritten and printed character recognition, speech recognition and text-to-speech conversion, intent recognition, information extraction and filtering, text classification and clustering, sentiment analysis, and opinion mining. It involves data mining, machine learning, knowledge acquisition, knowledge engineering, artificial intelligence research, and linguistic research related to language computation.

[0015] Insurance underwriting methods can be used to uniformly model and analyze multi-source time-series data such as customer behavior logs, policy lifecycle data, claims records, and financial indicators in order to conduct risk review and assessment of insurance applications.

[0016] Currently, insurance underwriting primarily relies on human experience to perform feature engineering on multi-source time-series data, followed by analysis and prediction of the extracted key information. However, in insurance scenarios, data sources such as customer behavior logs, policy data, and claims records are diverse and have significantly different time granularities. Typically, data of different granularities needs to be forcibly aligned to a unified time scale before modeling. This modeling process may result in the loss of crucial time-dynamic features, such as the difference between short-term fluctuations and long-term trends. This makes it difficult to accurately extract key information related to insurance risks from multi-source time-series data, impacting the accuracy of insurance underwriting.

[0017] Based on this, embodiments of this application provide an insurance underwriting method, apparatus, computer equipment, and storage medium to solve the technical problem that feature engineering relying on human experience makes it difficult to accurately extract key information related to insurance risks from multi-source time-series data, thus affecting the accuracy of insurance underwriting.

[0018] The insurance underwriting method, apparatus, computer equipment, and storage medium provided in this application are specifically described through the following embodiments. First, the insurance underwriting method in this application embodiment is described.

[0019] The insurance underwriting method provided in this application embodiment can be applied to, for example, Figure 1In this application environment, the client communicates with the server via a network. The server can obtain sample insurance business data from multiple different data sources through the client. It then encodes the sample insurance business data from each data source type using a pre-defined original time-series feature extraction model, obtaining a sample time-series encoded feature sequence for each data source type. Based on the sample time-series encoded feature sequences, the server trains the original time-series feature extraction model for time-series feature prediction, obtaining masked prediction loss data. It then trains the original time-series feature extraction model for time-series alignment, obtaining time-series aligned loss data. Based on the masked prediction loss data and the time-series aligned loss data, the server optimizes the original time-series feature extraction model, obtaining a target time-series feature extraction model. The target time-series feature extraction model is used to extract features from the pre-acquired policy time-series data, obtaining policy feature data. Underwriting is performed based on the policy feature data, and the underwriting results are fed back to the client for insurance agents to monitor and process.

[0020] In this application, sample insurance business data from multiple different data sources are first acquired. A pre-defined original temporal feature extraction model is used to encode the sample insurance business data for each data source type, resulting in a sample temporal encoded feature sequence for each data source type. This provides a rich and structured feature foundation for subsequent training, helping to accurately capture dependencies across time steps and improving the expressive power of the temporal encoded feature sequence for the dynamic changes in insurance business. Next, the original temporal feature extraction model is trained for temporal feature prediction based on the sample temporal encoded feature sequences, yielding masked prediction loss data. Then, the original temporal feature extraction model is trained for temporal alignment based on the sample temporal encoded feature sequences, yielding temporal aligned loss data. Finally, the original temporal feature extraction model is optimized based on the masked prediction loss data and the temporal aligned loss data to obtain the target temporal feature extraction model. This enables the model to learn the inherent patterns and contextual dependencies of temporal data rather than memorizing surface patterns, and to understand the same risk event across time scales, avoiding feature ambiguity caused by different granularities and improving the usability, generalization ability, and expressive power of features in different scenarios. Finally, the pre-acquired policy time-series data is extracted using a target time-series feature extraction model to obtain policy feature data. Underwriting is then performed based on this policy feature data. Since the obtained policy feature data has both strong semantic expression and multi-source time-series alignment characteristics, it can comprehensively and accurately depict the risk profile of the policyholder. This solves the technical problem that the feature engineering relying on human experience is unable to accurately extract key information related to insurance risks from multi-source time-series data, which affects the accuracy of insurance underwriting. This helps to improve the accuracy and efficiency of underwriting decisions and reduce the cost of manual review and the risk of misjudgment.

[0021] The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The following detailed description uses specific embodiments to illustrate this application.

[0022] Please see Figure 2 As shown, Figure 2 A flowchart illustrating the insurance underwriting method provided in this application embodiment includes the following steps: S10. Obtain sample insurance business data from multiple different data sources, and encode the sample insurance business data from each data source type using a preset original time-series feature extraction model to obtain the sample time-series encoded feature sequence for each data source type. First, sample insurance business data from multiple different data sources are obtained. The data source type refers to different business systems within the insurance company. The sample insurance business data includes, but is not limited to, claims records, policy underwriting data, customer behavior data, and customer financial indicators. Each type of sample insurance business data has a different data structure and business meaning, and the sample size of sample insurance business data from each data source type is at least one.

[0023] It should be noted that the preset original time-series feature extraction model refers to a model that is randomly initialized before training begins and has not yet been optimized with insurance business data. It is used to extract information features related to insurance underwriting from multi-source business data, such as claims time information, insurance type and amount information, user behavior information, etc. It helps to accurately extract the time-series dependencies of the data itself from multi-source data, providing complete and structured information for subsequent insurance underwriting.

[0024] Specifically, the original time-series feature extraction model includes at least an encoding layer, a transformer encoder, a prediction layer, and a fusion layer. The encoding layer encodes the sample insurance business data using time granularity (time stamps of the sample insurance business data), resulting in business-coded data. This transforms the discrete sample insurance business data into a continuous feature space that the original time-series feature extraction model can process, possessing temporal attributes and enabling the model to distinguish between short-term fluctuations and long-term trends. The transformer encoder independently processes the sample time-series encoded feature sequences from various data sources, extracting long-range temporal dependencies within the sequences to capture the dynamic patterns behind continuous business behaviors (such as the evolution of claims risk). The prediction layer receives contextual representations during the pre-training phase and reconstructs predictions from masked data values, forcing the model to learn deep correlations and potential temporal patterns within the data under unlabeled conditions. The fusion layer introduces configurable business weights for different data source types, compressing and integrating the target time-series features from different data source types output by the transformer encoder into a unified global representation vector through weighted summation. This enables the nonlinear convergence and dimensional merging of multi-source heterogeneous information within the feature space, effectively compensating for the information blind spots in risk profiling from a single data source (such as claims data alone), and significantly enhancing the overall representation's coverage of complex business scenarios and its comprehensive discrimination capability for downstream prediction tasks.

[0025] Among them, such as Figure 3 As shown, in step S10, that is, encoding the sample insurance business data of each data source type using a preset original time-series feature extraction model to obtain the sample time-series encoded feature sequence of each data source type, the following steps are included: S11. For each data source type, obtain the business timestamp of the sample insurance business data; use the original time series feature extraction model to perform feature encoding on the sample insurance business data and business timestamp to obtain the sample business encoding features; S12. Sort the sample business coding features by time based on the business timestamp to obtain the sample time-series coding feature sequence for each data source type.

[0026] Specifically, each data source type is processed separately. In detail, for each sample insurance business data of a specific data source type, the business timestamp of that sample insurance business data is first obtained. The business timestamp represents the time point when the business occurred. For example, if an insurance claim occurred on January 1, 2025, the business timestamp is January 1, 2025; if a customer's inquiry about an insurance product occurred at 10:31:25 on January 31, 2025, the business timestamp for that action is also January 31, 2025, at 10:31:25.

[0027] Next, the sample insurance business data and business timestamps are encoded using the encoding layer of the original time-series feature extraction model to obtain the sample business encoding features.

[0028] In step S11, which involves encoding the sample insurance business data and business timestamps using the encoding layer of the original time-series feature extraction model to obtain the sample business encoding features, the steps include: The sample insurance business data is feature-encoded using the original time-series feature extraction model to obtain business-encoded data; The business timestamps are time-encoded using the original time-series feature extraction model to obtain business time-encoded data; Feature fusion is performed on business code data and business time code data to obtain sample business code features.

[0029] Specifically, the sample insurance business data is feature-encoded through the encoding layer of the original time-series feature extraction model to obtain business-encoded data, thereby transforming the discrete sample insurance business data into a continuous feature space that the original time-series feature extraction model can process. The business timestamps are then time-encoded through the encoding layer of the original time-series feature extraction model to obtain business time-encoded data, thus mapping the business timestamps to scales such as day, month, and time interval. Furthermore, feature concatenation can be used to fuse the business-encoded data and the business time-encoded data to obtain the sample business-encoded features corresponding to the sample insurance business data. These sample business-encoded features retain both business-related information and time information, which helps to more accurately distinguish between short-term abnormal fluctuations and long-term risk trends in subsequent feature extraction processes, improving the precision of underwriting decisions.

[0030] It is understandable that by introducing time coding, the original time-series feature extraction model can distinguish behavioral patterns at different time scales, such as short-term fluctuations and long-term trends, thus making it easier to consider the temporal correlation when extracting features.

[0031] It should be noted that the sample business coding features include business time coding data with business timestamps, which can be understood as containing time information. Therefore, for multiple sample business coding features in each data source type, the multiple sample business coding features are sorted by time based on the chronological order of the business timestamps to form a sequence, resulting in a sample time-series coding feature sequence for each data source type. This eliminates the time-series disorder caused by data extraction, storage, and querying, ensuring that the feature sequence is consistent with the timeline of actual business occurrences. This allows the original time-series feature extraction model to accurately learn the dependencies across time steps, avoiding risk misjudgments caused by time sequence reversal, thereby improving the underwriting model's ability to capture the evolution trajectory of customer risk.

[0032] For example, in a car insurance underwriting scenario, a customer's accident records may be returned as [2024 bump-and-run accident 5000 yuan, 2022 rear-end collision accident 16000 yuan, 2023 bump-and-run accident 3000 yuan] due to the database query order. By encoding and sorting, a sample time-series encoded feature sequence of [2022 rear-end collision accident 16000 yuan, 2023 bump-and-run accident 3000 yuan, 2024 bump-and-run accident 5000 yuan] can be obtained.

[0033] It is understandable that the sample time-series coding feature sequence contains multiple elements, which are the sample business coding features.

[0034] For steps S11 to S12 above, firstly, feature encoding is performed on the sample insurance business data to obtain business-coded data, and time encoding is performed on the business timestamps to obtain business-time-coded data. This avoids semantic confusion caused by forcibly fusing business numerical and time information in the early stages, allowing the two types of information to independently retain their respective features. Next, feature fusion is performed on the business-coded data and business-time-coded data to obtain sample business-coded features. This enables the time-series feature extraction model to simultaneously perceive changes in business data and their temporal relationships, thereby effectively distinguishing between short-term fluctuations and long-term trends. Finally, the sample business-coded features are sorted by time based on the business timestamps to obtain a sample time-series-coded feature sequence for each data source type. This helps to accurately capture cross-time step dependencies in subsequent processing, avoids feature distortion caused by temporal disorder, and ultimately improves the ability of the time-series-coded feature sequence to express the dynamic changes in insurance business.

[0035] S20. Based on the sample temporal coding feature sequence, train the original temporal feature extraction model for temporal feature prediction to obtain mask prediction loss data; It is important to understand that temporal feature prediction training here is one of the tasks of model training. Specifically, temporal feature prediction training refers to using masked self-supervised learning to enable the original temporal feature extraction model to predict the content of the masked part based on the unmasked temporal segments. This allows the model to learn the inherent rules and contextual dependencies of temporal data rather than memorizing surface patterns, thereby improving the generalization ability and robustness of features and effectively enhancing the feature representation ability of downstream tasks.

[0036] Among them, such as Figure 4 As shown, step S20, which involves training the original temporal feature extraction model for temporal feature prediction based on the sample temporal coding feature sequence to obtain mask prediction loss data, includes the following steps: S21. Mask the sample temporal coding feature sequence to obtain the sample masked coding feature sequence; It should be noted that masking and mask prediction can be performed on the temporal coding feature sequences of samples from one or several data source types, or masking and mask prediction can be performed on the temporal coding feature sequences of samples from all data source types. The specific choice should be made based on the actual business scenario, and is not limited to this.

[0037] Specifically, for a sample temporal encoded feature sequence, certain elements in the sample temporal encoded feature sequence can be masked using methods such as random masking (randomly covering specific time steps, suitable for learning fine-grained local patterns), continuous masking (randomly selecting several consecutive time periods for overall masking, suitable for learning long-range dependencies and periodic patterns), and hybrid masking (combining random masking and continuous masking to enable the model to capture both local details and understand the global structure), to obtain the sample masked encoded feature sequence.

[0038] It should be noted that masking also needs to be performed according to a specific masking ratio, usually set between 15% and 30%. The masking ratio needs to balance two goals: if the ratio is too low, the learning task will be too simple and the model will have difficulty learning effective features; if the ratio is too high, there will be too little contextual information and the reconstruction will be too difficult.

[0039] S22. Perform mask prediction on the sample mask encoded feature sequence using the original temporal feature extraction model to obtain the sample mask predicted feature sequence. Specifically, the sample mask-encoded feature sequence is temporally correlated and encoded by the transformer encoder of the original temporal feature extraction model. The dependencies between consecutive time steps within the sample mask-encoded feature sequence are extracted and captured to obtain a global representation vector, which is the mask correlation feature. This mask correlation feature integrates information from other uncovered positions in the sample mask-encoded feature sequence and reflects the dynamic patterns, short-term fluctuations, and local patterns within the comparison temporal encoded feature sequence.

[0040] Next, the mask-related features are masked and predicted by the prediction layer of the original temporal feature extraction model to obtain the sample mask prediction feature sequence. It should be understood that the sample mask prediction feature sequence contains the predicted values ​​of the masked elements in the sample mask encoded feature sequence.

[0041] S23. Based on the sample mask prediction feature sequence and the sample temporal coding feature sequence, calculate the loss to obtain the mask prediction loss data.

[0042] After obtaining the sample masked prediction feature sequences corresponding to all sample time-series encoded feature sequences that have undergone masking processing, loss calculations are performed on the predicted value of the masked element in each sample masked prediction feature sequence and the true value of the masked element in the corresponding sample time-series encoded feature sequence. Multiple masked prediction loss data are obtained, and all masked prediction loss data are summed to obtain the final masked prediction loss data. Specifically, if the masked element is a continuous value (such as compensation amount or premium), the mean squared error loss function (MSE) can be used; if the masked element is a discrete value (such as behavior type or insurance category), cross-entropy loss can be used.

[0043] S30. Based on the sample temporal coding feature sequence, the original temporal feature extraction model is trained for temporal alignment to obtain temporal alignment loss data; It is important to understand that time-series alignment training here is one of the tasks of model training. Specifically, time-series alignment training refers to extracting features from business data of the same data source type at different time granularities, and then constraining the model to make the output features of the two time granularities as close as possible. This enables the model to understand the same risk event across time scales, avoids feature ambiguity caused by different granularities, and improves the usability of features in different scenarios.

[0044] Among them, such as Figure 5 As shown, step S30, which involves training the original temporal feature extraction model for temporal alignment based on the sample temporal coding feature sequence to obtain temporal alignment loss data, includes the following steps: S31. Construct samples based on sample time-series coding feature sequences to obtain control time-series coding feature sequences; It is important to understand that sample construction here refers to constructing control time-series encoded feature sequences at different time granularities based on the sample time-series encoded feature sequences. The sample time-series encoded feature sequences and the control time-series encoded feature sequences both express the same event, but at different time granularities, providing data for subsequent cross-granularity alignment.

[0045] Among them, such as Figure 6 As shown, step S31, which involves constructing samples from the sample temporal coding feature sequence to obtain the control temporal coding feature sequence, includes the following steps: S311. Obtain the reference time-series unit of the sample temporal coding feature sequence, and determine the target time-series unit based on the reference time-series unit; wherein, the target time-series unit is larger than the reference time-series unit; S312. Based on the target time series unit, perform time series segmentation on the sample time series coding feature sequence to obtain multiple segmented time series coding feature sequences, and aggregate the segmented time series coding feature sequences to obtain multiple target time series coding features; S313. Summarize the time-series coding features of multiple targets to obtain the comparison time-series coding feature sequence.

[0046] Specifically, the first step is to obtain the reference time-series unit for the temporal coding features of the samples. The reference time-series unit is the time granularity, which includes days, weeks, months, quarters, years, etc.

[0047] It should be noted that the benchmark time series unit is the smallest time granularity extracted from the sample insurance business data. Therefore, after obtaining the benchmark time series unit, a larger time granularity is set as the target time series unit based on the benchmark time series unit, realizing the transformation from fine granularity to coarse granularity. This ensures that the constructed control time series coding feature sequence and the sample time series coding feature sequence differ in time span but describe the same business object.

[0048] It is understandable that the target time series unit is larger than the base time series unit. For example: if the base time series unit is day, then the target time series unit can be week, month, quarter, or year; if the base time series unit is week, then the target time series unit can be month, quarter, or year; if the base time series unit is month, then the target time series unit can be quarter or year; if the base time series unit is quarter, then the target time series unit can be year. Specifically, the choice needs to be made according to the actual business scenario, and no restrictions are imposed here.

[0049] Having identified the target temporal unit, the sample temporal coding feature sequence is segmented into multiple continuous segments based on this unit. Each segment within the sample temporal coding feature sequence comprises elements that form a segmented temporal coding feature sequence. Next, for each segmented temporal coding feature sequence, the feature values ​​of each element are merged (either by averaging or summing) to obtain a coarse-grained representation, which is the target temporal coding feature. This process preserves the local semantic information within each segment and avoids information loss caused by direct aggregation. For example, fine-grained representations of multiple days within a month are merged into a coarse-grained representation at the month level.

[0050] Finally, all target temporal coding features are summarized in chronological order to form a complete coarse-grained temporal representation sequence, resulting in the reference temporal coding feature sequence.

[0051] It should be noted that the comparison time-series encoded feature sequence and the sample time-series encoded feature sequence constitute positive sample pairs of the same business data at different time granularities. The comparison time-series encoded feature sequence and the sample time-series encoded feature sequence correspond one-to-one in sequence structure, ensuring that the model can accurately constrain the consistency between different time granularities and improve the model's ability to understand customer risks at multiple scales in insurance underwriting.

[0052] S32. Extract features from the sample time-series encoded feature sequence using the original time-series feature extraction model to obtain sample feature data; S33. Extract features from the control time-series encoded feature sequence using the original time-series feature extraction model to obtain control feature data; Specifically, the time-series encoded feature sequence of the sample is temporally correlated by the transformer encoder of the original time-series feature extraction model. The dependency relationship between the time steps before and after the sample time-series encoded feature sequence is extracted and captured to obtain the global representation vector, which is the sample feature data. This sample feature data is used to characterize the dynamic law, short-term fluctuation and local pattern inside the sample time-series encoded feature sequence.

[0053] The transformer encoder of the original time series feature extraction model performs time series correlation encoding on the reference time series coded feature sequence, extracts and captures the dependencies between consecutive time steps within the reference time series coded feature sequence, and obtains a global representation vector, which is the reference feature data. This reference feature data is used to characterize the dynamic patterns, short-term fluctuations and local patterns within the reference time series coded feature sequence.

[0054] It is understandable that by using the transformer encoder of the original temporal feature extraction model to extract features from the sample temporal encoded feature sequence and the control temporal encoded feature sequence respectively, it is possible to ensure that the extracted sample feature data and control feature data truly reflect the current encoding capability of the model, and ensure that the sample feature data and control feature data are in the same feature space and the same parameter system, so that the difference between the two is caused only by the difference in granularity rather than the difference in the model, ensuring that the calculation of alignment loss has a clear semantic meaning, and promoting consistent semantic expression under different granularities.

[0055] S34. Based on the sample feature data and the control feature data, the alignment loss is calculated to obtain the temporal alignment loss data.

[0056] Specifically, the square of the Euclidean distance (i.e., the square of the L2 norm) between the sample feature data and the control feature data can be calculated to obtain the temporal alignment loss data. This temporal alignment loss data can measure the degree of inconsistency between the representations at the two granularities. As a supervision signal for model optimization, it encourages the model to map representations at different granularities to similar positions in the same semantic space, thereby enhancing the model's adaptability to data at different temporal granularities and improving the generalization and stability of risk assessment in underwriting scenarios.

[0057] S40. Based on the mask prediction loss data and the temporal alignment loss data, the original temporal feature extraction model is optimized to obtain the target temporal feature extraction model. It is important to understand that model optimization here refers to updating the parameters of the original temporal feature extraction model based on masked prediction loss data and temporal alignment loss data through model optimization algorithms (such as gradient descent and Adam algorithm) to obtain a target temporal feature extraction model with temporal prediction capability and cross-granularity alignment capability, so as to output a more comprehensive and stable feature representation for subsequent insurance underwriting tasks.

[0058] In step S40, the original temporal feature extraction model is optimized based on the mask prediction loss data and the temporal alignment loss data to obtain the target temporal feature extraction model, including the following steps: The target loss data is obtained by weighted fusion of the mask prediction loss data and the temporal alignment loss data. The original time-series feature extraction model is adjusted based on the target loss data to obtain the target time-series feature extraction model.

[0059] Specifically, the mask prediction loss data and the temporal alignment loss data are weighted and summed according to preset weight coefficients to obtain the target loss data. This avoids the single task loss dominating the optimization direction, allowing the model to simultaneously consider local prediction accuracy and global consistency. For example, the weight coefficients are set to 0.6 and 0.4, with the weight coefficient for the mask prediction loss data being 0.6 and the weight coefficient for the temporal alignment loss data being 0.4.

[0060] After obtaining the target loss data, optimization algorithms (such as gradient descent, Adam algorithm, etc.) are used to calculate the gradient based on the target loss data. Then, the model parameters of the original time-series feature extraction model are updated along the opposite direction of the gradient. After multiple iterations, the target loss data is gradually reduced until a preset stopping condition is met, resulting in the target time-series feature extraction model. By adjusting the parameters, the original time-series feature extraction model can be continuously optimized. When applying the trained model, the extracted time-series features possess both local prediction accuracy and multi-scale consistency. In insurance underwriting scenarios, this enables more accurate extraction of underwriting-related information from multi-source data, while ensuring logical consistency between short-term indicators and long-term risk assessment conclusions, thereby improving the reliability of underwriting decisions.

[0061] It should be noted that this application trains the original temporal feature extraction model through temporal feature prediction and temporal alignment training. This allows the model to learn the potential temporal patterns of data such as insurance claims and financial expenses without requiring a large amount of manually labeled data. By allowing data at different time granularities to express the same trend information, the model can improve its stability and generalization ability across different business scenarios. Compared to traditional single-task training methods, the target temporal feature extraction model trained in this application can form a unified temporal representation, enabling it to uniformly process premiums, claims, and expenses at multiple time granularities, including daily, monthly, and event-level data. This supports reuse for subsequent tasks such as expense prediction and claims prediction, reducing repetitive modeling costs and improving model transferability.

[0062] S50. Extract features from the pre-acquired policy time-series data using the target time-series feature extraction model to obtain policy feature data; It should be noted that the policy time-series data obtained in advance is the relevant data of a certain customer during the application phase. This policy time-series data comes from different data sources, including but not limited to claims records, policy underwriting data, customer behavior data, and customer financial indicators.

[0063] The policy time-series data is extracted using a trained target time-series feature extraction model to obtain policy feature data. Specifically, the process includes the following steps: Step 1: For each data source type, obtain the policy timestamp of the policy time series data; encode the policy time series data using the encoding layer of the target time series feature extraction model to obtain policy coded data; encode the policy timestamp using the encoding layer of the target time series feature extraction model to obtain policy time coded data; perform feature fusion on the policy coded data and policy time coded data to obtain policy business coded features. Step 2: Sort the policy business coding features by time based on the policy timestamp to obtain the target time-series coding feature sequence for each data source type; Step 3: For each data source type, the target temporal encoded feature sequence is temporally correlated and encoded using the transformer encoder of the target temporal feature extraction model to obtain the target temporal features; Step 4: Obtain the business weights for each data source type; Step 5: Through the fusion layer of the target time series feature extraction model, the target time series features are weighted and fused based on business weights to obtain policy feature data.

[0064] The specific implementation methods for steps one and two are basically the same as those for steps S11 and S12, and will not be repeated here.

[0065] For step three, the Ttransformer encoder of the trained target temporal feature extraction model independently extracts the time dependency patterns within the target temporal encoded feature sequence of each data source type, avoiding mutual interference between temporal signals from different sources, thereby preserving the unique risk evolution characteristics of each source. Specifically, the implementation of step three is basically the same as the implementation of steps S32 and S33, and will not be repeated here.

[0066] It should be noted that business weight refers to the importance coefficient assigned to each data source type based on professional experience or historical data statistics in the insurance underwriting field, reflecting the degree of contribution of the source to the underwriting risk assessment. For example, the weight of claims data is 0.5, the weight of policy data is 0.3, and the weight of behavioral data is 0.2, and all weights are non-negative and normalizable.

[0067] For step four, the business weights of each data source type are read from the insurance business rules base. The business weight refers to the importance coefficient assigned to each data source type based on professional experience or historical data statistics in the insurance underwriting field, reflecting the degree of contribution of the source to the underwriting risk assessment. For example, the weight of claims data is 0.5, the weight of policy data is 0.3, and the weight of behavioral data is 0.2, so that high-risk signals (such as claims records) receive a higher integration weight.

[0068] In step five, within the fusion layer of the target time-series feature extraction model, the target time-series features from each data source type are weighted and summed according to their corresponding business weights to obtain policy feature data. This policy feature data integrates multi-scale time-series information of the customer across all data source types. While integrating multi-source time-series information, business weights ensure that key information dominates the feature representation, preventing low-value information from diluting high-value information. This results in a unified representation that covers multi-dimensional risks and highlights core risk factors, effectively improving the accuracy and stability of underwriting decisions.

[0069] Understandably, by using the target time series feature extraction model as a tool for extracting policy time series data, it is not necessary to model each new policy from scratch, which can significantly reduce underwriting computational overhead and reduce the workload of manual feature engineering, thus helping to improve model transferability.

[0070] S60. Underwriting based on policy feature data.

[0071] It should be noted that insurance underwriting refers to the decision-making process by which insurance companies review and assess the risk characteristics of the insured (extracted policy characteristics) to determine whether to underwrite and under what conditions. Specifically, underwriting requires the use of pre-trained underwriting decision models (such as machine learning models, large language models, etc.) provided by the insurance company to output underwriting results. These results include, but are not limited to, normal risk (underwriting at the standard rate); slightly higher risk (underwriting after an additional premium); no liability for specific diseases or risks, but normal underwriting for the remaining portion; deferred underwriting; and rejection of the application due to risk exceeding the insurable range.

[0072] Understandably, this application extracts high-quality policy feature data through a well-trained target time-series feature extraction model, which can more comprehensively depict the risk profile of the policyholder, improve underwriting accuracy, reduce claims losses caused by misjudgments, and at the same time improve underwriting efficiency and reduce manual review costs.

[0073] In some embodiments, after step S50, in addition to underwriting the policy feature data, customer behavior analysis, financial product marketing, claims fraud detection, and financial health analysis can also be performed on the policy feature data. The specific use depends on the actual application scenario and is not limited here.

[0074] As can be seen, the above solution can be applied to insurance underwriting scenarios in fintech. First, sample insurance business data from multiple different data sources are acquired. Then, a pre-defined original temporal feature extraction model is used to encode the sample insurance business data from each data source type, obtaining a sample temporal encoded feature sequence for each data source type. This provides a rich and structured feature foundation for subsequent training, helping to accurately capture dependencies across time steps and improving the ability of the temporal encoded feature sequence to express the dynamic changes in insurance business. Next, the original temporal feature extraction model is trained for temporal feature prediction based on the sample temporal encoded feature sequences, obtaining masked prediction loss data. Then, the original temporal feature extraction model is trained for temporal alignment based on the sample temporal encoded feature sequences, obtaining temporal alignment loss data. Finally, the original temporal feature extraction model is optimized based on the masked prediction loss data and the temporal alignment loss data to obtain the target temporal feature extraction model. This enables the model to learn the inherent patterns and contextual dependencies of temporal data rather than memorizing surface patterns, and to understand the same risk event across time scales, avoiding feature ambiguity caused by different granularities, and improving the usability, generalization ability, and expressive power of features in different scenarios. Finally, the pre-acquired policy time-series data is extracted using a target time-series feature extraction model to obtain policy feature data. Underwriting is then performed based on this policy feature data. Since the obtained policy feature data has both strong semantic expression and multi-source time-series alignment characteristics, it can comprehensively and accurately depict the risk profile of the policyholder. This solves the technical problem that the feature engineering relying on human experience is unable to accurately extract key information related to insurance risks from multi-source time-series data, which affects the accuracy of insurance underwriting. This helps to improve the accuracy and efficiency of underwriting decisions and reduce the cost of manual review and the risk of misjudgment.

[0075] 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 this application.

[0076] In one embodiment, an insurance underwriting device is provided, which corresponds one-to-one with the insurance underwriting methods described in the above embodiments. For example... Figure 7 As shown, the insurance underwriting device includes a business data acquisition module 101, a time-series feature prediction and training module 102, a time-series alignment training module 103, a model optimization module 104, a policy feature extraction module 105, and an underwriting module 106. Detailed descriptions of each functional module are as follows: The business data acquisition module 101 is used to acquire sample insurance business data from multiple different data sources. It encodes the sample insurance business data from each data source type using a preset original time-series feature extraction model to obtain the sample time-series encoded feature sequence for each data source type. The temporal feature prediction training module 102 is used to train the original temporal feature extraction model for temporal feature prediction based on the sample temporal encoded feature sequence, and to obtain mask prediction loss data. The temporal alignment training module 103 is used to train the original temporal feature extraction model on temporal alignment based on the sample temporal coding feature sequence to obtain temporal alignment loss data. The model optimization module 104 is used to optimize the original temporal feature extraction model based on mask prediction loss data and temporal alignment loss data to obtain the target temporal feature extraction model. The policy feature extraction module 105 is used to extract features from the pre-acquired policy time series data through the target time series feature extraction model to obtain policy feature data; The underwriting module 106 is used for underwriting based on policy feature data.

[0077] In one embodiment, the business data acquisition module 101 is specifically used for: For each data source type, obtain the business timestamp of the sample insurance business data; use the original time series feature extraction model to perform feature encoding on the sample insurance business data and business timestamp to obtain the sample business encoding features; Based on the business timestamp, the sample business coding features are sorted by time to obtain the sample time-series coding feature sequence for each data source type.

[0078] In one embodiment, the business data acquisition module 101 is specifically used for: The sample insurance business data is feature-encoded using the original time-series feature extraction model to obtain business-encoded data; The business timestamps are time-encoded using the original time-series feature extraction model to obtain business time-encoded data; Feature fusion is performed on business code data and business time code data to obtain sample business code features.

[0079] In one embodiment, the temporal feature prediction training module 102 is specifically used for: The sample temporal encoded feature sequence is masked to obtain the sample masked encoded feature sequence; The sample mask-encoded feature sequence is predicted by performing mask prediction on the sample mask encoding feature sequence using the original temporal feature extraction model; Loss calculation is performed based on the sample mask prediction feature sequence and the sample temporal coding feature sequence to obtain mask prediction loss data.

[0080] In one embodiment, the time-series alignment training module 103 is specifically used for: Samples are constructed based on the sample time-series coding feature sequences to obtain the control time-series coding feature sequences; The sample time-series encoded feature sequence is extracted using the original time-series feature extraction model to obtain sample feature data; Feature extraction is performed on the control time-series encoded feature sequence using the original time-series feature extraction model to obtain control feature data; Alignment loss is calculated based on sample feature data and control feature data to obtain temporal alignment loss data.

[0081] In one embodiment, the time-series alignment training module 103 is specifically used for: Obtain the baseline time-series unit of the sample temporal coding feature sequence, and determine the target time-series unit based on the baseline time-series unit; wherein, the target time-series unit is larger than the baseline time-series unit; Based on the target time series unit, the sample time series coding feature sequence is temporally segmented to obtain multiple segmented time series coding feature sequences, and the segmented time series coding feature sequences are aggregated to obtain multiple target time series coding features; By summarizing the temporal coding features of multiple targets, a comparative temporal coding feature sequence is obtained.

[0082] In one embodiment, the model optimization module 104 is specifically used for: The target loss data is obtained by weighted fusion of the mask prediction loss data and the temporal alignment loss data. The original time-series feature extraction model is adjusted based on the target loss data to obtain the target time-series feature extraction model.

[0083] This application provides an insurance underwriting device applicable to insurance underwriting scenarios in fintech. First, it acquires sample insurance business data from multiple data sources. Then, it encodes the sample insurance business data from each data source type using a pre-defined original time-series feature extraction model, obtaining a sample time-series encoded feature sequence for each data source type. This provides a rich and structured feature foundation for subsequent training, helping to accurately capture dependencies across time steps and improving the expressive power of the time-series encoded feature sequence for the dynamic changes in insurance business. Next, it trains the original time-series feature extraction model for time-series feature prediction based on the sample time-series encoded feature sequences, obtaining masked prediction loss data. It then trains the original time-series feature extraction model for time-series alignment based on the sample time-series encoded feature sequences, obtaining time-series aligned loss data. Finally, it optimizes the original time-series feature extraction model based on the masked prediction loss data and the time-series aligned loss data, obtaining a target time-series feature extraction model. This enables the model to learn the inherent patterns and contextual dependencies of time-series data rather than memorizing surface patterns, and to understand the same risk event across time scales, avoiding feature ambiguity caused by different granularities and improving the usability, generalization ability, and expressive power of features in different scenarios. Finally, the pre-acquired policy time-series data is extracted using a target time-series feature extraction model to obtain policy feature data. Underwriting is then performed based on this policy feature data. Since the obtained policy feature data has both strong semantic expression and multi-source time-series alignment characteristics, it can comprehensively and accurately depict the risk profile of the policyholder. This solves the technical problem that the feature engineering relying on human experience is unable to accurately extract key information related to insurance risks from multi-source time-series data, which affects the accuracy of insurance underwriting. This helps to improve the accuracy and efficiency of underwriting decisions and reduce the cost of manual review and the risk of misjudgment.

[0084] Specific limitations regarding the insurance underwriting device can be found in the limitations of the insurance underwriting method described above, and will not be repeated here. Each module in the aforementioned insurance underwriting 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 corresponding operations of each module.

[0085] Please see Figure 8 , Figure 8 The hardware structure of a computer device according to another embodiment is illustrated. The computer device includes: The processor 801 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 802 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 802 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 802 and is called and executed by the processor 801 to implement the insurance underwriting method of the embodiments of this application, including: Acquire sample insurance business data from multiple different data sources, and encode the sample insurance business data from each data source type using a pre-defined raw time-series feature extraction model to obtain the sample time-series encoded feature sequence for each data source type. The original temporal feature extraction model is trained for temporal feature prediction based on the sample temporal coding feature sequence to obtain mask prediction loss data; The original temporal feature extraction model is trained on temporal alignment based on the sample temporal coding feature sequence to obtain temporal alignment loss data; The original temporal feature extraction model is optimized based on mask prediction loss data and temporal alignment loss data to obtain the target temporal feature extraction model. The policy feature data is obtained by extracting features from the pre-acquired policy time series data using the target time series feature extraction model. Underwriting is based on policy feature data.

[0086] The 803 input / output interface is used to implement information input and output. The communication interface 804 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 805 transmits information between various components of the device (e.g., processor 801, memory 802, input / output interface 803, and communication interface 804); The processor 801, memory 802, input / output interface 803, and communication interface 804 are connected to each other within the device via bus 805.

[0087] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Acquire sample insurance business data from multiple different data sources, and encode the sample insurance business data from each data source type using a pre-defined raw time-series feature extraction model to obtain the sample time-series encoded feature sequence for each data source type. The original temporal feature extraction model is trained for temporal feature prediction based on the sample temporal coding feature sequence to obtain mask prediction loss data; The original temporal feature extraction model is trained on temporal alignment based on the sample temporal coding feature sequence to obtain temporal alignment loss data; The original temporal feature extraction model is optimized based on mask prediction loss data and temporal alignment loss data to obtain the target temporal feature extraction model. The policy feature data is obtained by extracting features from the pre-acquired policy time series data using the target time series feature extraction model. Underwriting is based on policy feature data.

[0088] 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 relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0089] 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.

[0090] 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.

[0091] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0092] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 this application, and should all be included within the protection scope of this application.

Claims

1. An insurance underwriting method, characterized in that, The method includes: Multiple sample insurance business data from different data sources are acquired, and the sample insurance business data from each data source type are encoded using a preset original time-series feature extraction model to obtain a sample time-series encoded feature sequence for each data source type. Based on the sample temporal coding feature sequence, the original temporal feature extraction model is trained for temporal feature prediction to obtain mask prediction loss data; Based on the sample temporal coding feature sequence, the original temporal feature extraction model is trained for temporal alignment to obtain temporal alignment loss data; Based on the mask prediction loss data and the temporal alignment loss data, the original temporal feature extraction model is optimized to obtain the target temporal feature extraction model. The target time-series feature extraction model is used to extract features from the pre-acquired policy time-series data to obtain policy feature data. Underwriting is performed based on the policy feature data.

2. The insurance underwriting method according to claim 1, characterized in that, The step of training the original temporal feature extraction model based on the sample temporal coding feature sequence to obtain mask prediction loss data includes: The sample temporal coding feature sequence is masked to obtain the sample masked coding feature sequence; The sample mask encoding feature sequence is predicted by performing mask prediction on the sample mask encoding feature sequence using the original temporal feature extraction model; The loss is calculated based on the sample mask prediction feature sequence and the sample temporal coding feature sequence to obtain the mask prediction loss data.

3. The insurance underwriting method according to claim 1, characterized in that, The step of performing time-series alignment training on the original time-series feature extraction model based on the sample time-series encoded feature sequence to obtain time-series alignment loss data includes: Based on the sample temporal coding feature sequence, a sample is constructed to obtain the control temporal coding feature sequence; The sample time-series encoded feature sequence is extracted using the original time-series feature extraction model to obtain sample feature data; The original time-series feature extraction model is used to extract features from the control time-series encoded feature sequence to obtain control feature data; Alignment loss is calculated based on the sample feature data and the control feature data to obtain the temporal alignment loss data.

4. The insurance underwriting method according to claim 3, characterized in that, The step of constructing a sample based on the sample time-series coding feature sequence to obtain a control time-series coding feature sequence includes: Obtain the reference time-series unit of the sample time-series encoded feature sequence, and determine the target time-series unit based on the reference time-series unit; wherein the target time-series unit is larger than the reference time-series unit; Based on the target time-series unit, the sample time-series coding feature sequence is temporally segmented to obtain multiple segmented time-series coding feature sequences, and the segmented time-series coding feature sequences are aggregated to obtain multiple target time-series coding features; The target temporal coding features are summarized to obtain the comparison temporal coding feature sequence.

5. The insurance underwriting method according to any one of claims 1 to 4, characterized in that, The step involves encoding the sample insurance business data of each data source type using a preset original time-series feature extraction model to obtain a sample time-series encoded feature sequence for each data source type, including: For each data source type, obtain the business timestamp of the sample insurance business data; use the original time-series feature extraction model to perform feature encoding on the sample insurance business data and the business timestamp to obtain sample business encoding features; Based on the business timestamp, the sample business coding features are sorted by time to obtain the sample time-series coding feature sequence for each data source type.

6. The insurance underwriting method according to claim 5, characterized in that, The step of encoding the sample insurance business data and the business timestamp using the original time-series feature extraction model to obtain sample business encoding features includes: The original time-series feature extraction model is used to perform feature encoding on the sample insurance business data to obtain business coded data; The business timestamp is time-encoded using the original time-series feature extraction model to obtain business time-encoded data. The service coding data and the service time coding data are fused to obtain the sample service coding features.

7. The insurance underwriting method according to any one of claims 1 to 4, characterized in that, The optimization of the original temporal feature extraction model based on the mask prediction loss data and the temporal alignment loss data to obtain the target temporal feature extraction model includes: The target loss data is obtained by weighted fusion of the mask prediction loss data and the temporal alignment loss data. The original temporal feature extraction model is adjusted based on the target loss data to obtain the target temporal feature extraction model.

8. An insurance underwriting device, characterized in that, The device includes: The business data acquisition module is used to acquire sample insurance business data from multiple different data sources. It encodes the sample insurance business data from each data source type using a preset original time-series feature extraction model to obtain a sample time-series encoded feature sequence for each data source type. The temporal feature prediction training module is used to perform temporal feature prediction training on the original temporal feature extraction model based on the sample temporal coding feature sequence to obtain mask prediction loss data. The temporal alignment training module is used to perform temporal alignment training on the original temporal feature extraction model based on the sample temporal coding feature sequence to obtain temporal alignment loss data. The model optimization module is used to optimize the original temporal feature extraction model based on the mask prediction loss data and the temporal alignment loss data to obtain the target temporal feature extraction model. The policy feature extraction module is used to extract features from the pre-acquired policy time series data through the target time series feature extraction model to obtain policy feature data; The underwriting module is used to perform underwriting based on the policy feature data.

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 insurance underwriting 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 insurance underwriting method as described in any one of claims 1 to 7.

Images