Personalized property insurance underwriting risk assessment method and system based on big data

Abstract

The invention provides a personalized property insurance underwriting risk assessment method and system based on big data. The method comprises the steps of obtaining multi-source heterogeneous data; the multi-source heterogeneous data comprises customer property insurance policy data, claim settlement records and property insurance industry risk data based on big data analysis; performing multi-dimensional feature extraction on the multi-source heterogeneous data based on a dynamic heterogeneous fusion module; the dynamic heterogeneous fusion module is formed by connecting a spatial convolution unit, a time sequence memory unit and a business rule unit in parallel; the spatial convolution unit extracts spatial distribution features, and the time sequence memory unit extracts time sequence distribution features; the business rule unit extracts features conforming to business logic; and inputting the multi-dimensional features into a risk assessment model for assessment, and outputting an assessment result including a risk level. According to the invention, the final property insurance underwriting risk assessment result is more accurate.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a personalized property insurance underwriting risk assessment method and system based on big data. Background Technology

[0002] Traditional property insurance risk assessment methods rely heavily on limited historical data and relatively fixed experience models. Their limitations become increasingly apparent when faced with massive, complex, and constantly changing modern risk factors. Traditional methods for feature extraction from property insurance data lack in-depth data mining and effective integration. Customer property insurance policy data, claims records, and property insurance industry risk data each possess distinct structures and characteristics, making it difficult for traditional assessment methods to fully leverage the potential correlations between these data. For example, there are complex spatiotemporal relationships and business logic connections between the distribution of customer policies in different regions, local claims records, and the overall risk situation of the property insurance industry in those regions. Past simplistic data processing methods fail to extract these deep-seated features, resulting in assessment results that do not accurately reflect the actual risk level and ultimately lead to inaccurate risk assessments. Summary of the Invention The main objective of this invention is to provide a personalized property insurance underwriting risk assessment method and system based on big data, aiming to overcome the inaccuracy of current property insurance underwriting risk assessment results.

[0003] To achieve the above objectives, this invention provides a personalized property insurance underwriting risk assessment method based on big data, comprising the following steps: Acquire multi-source heterogeneous data; the multi-source heterogeneous data includes customer property insurance policy data, claims records, and property insurance industry risk data based on big data analysis; The multi-source heterogeneous data is subjected to multi-dimensional feature extraction based on a dynamic heterogeneous fusion module. The dynamic heterogeneous fusion module is composed of a spatial convolution unit, a temporal memory unit, and a business rule unit connected in parallel. The spatial convolution unit extracts spatial distribution features, the temporal memory unit extracts temporal distribution features, and the business rule unit extracts features that conform to business logic. The multi-dimensional features are input into the risk assessment model for evaluation, and the output includes the assessment results containing the risk level.

[0004] Furthermore, acquire multi-source heterogeneous data, including: Obtain multiple raw data sets; The original data is cleaned to remove duplicate and abnormal data, and then standardized to obtain the multi-source heterogeneous data. Furthermore, based on the dynamic heterogeneous fusion module, multi-dimensional feature extraction is performed on the multi-source heterogeneous data, including: By using multi-layer convolution kernels of spatial convolution units, convolution operations are performed on the geographical location information of insured property in the customer property insurance policy data and the regional risk distribution information in the property insurance industry risk data to extract risk clustering features and geographical correlation features at different spatial scales as spatial distribution features. By using the long short-term memory network structure of the temporal memory unit, the time sequence of claims events in the claims records and the insurance time sequence in the customer property insurance policy data are analyzed to extract the risk occurrence frequency characteristics and risk change trend characteristics in the time dimension, and the feature information of key time nodes is dynamically selected as the temporal distribution characteristics. Based on the pre-built property insurance underwriting business rule library, the business rule unit matches and analyzes the policyholder qualification information and property attribute information in the customer property insurance policy data, and extracts features that meet the business access conditions and features related to the underwriting limit as features that conform to the business logic. The features extracted by the spatial convolutional unit, temporal memory unit, and business rule unit are weighted and fused, and the weight values ​​are dynamically determined based on the risk contribution of each feature in historical evaluation cases.

[0005] Furthermore, the size and number of convolution kernels in the spatial convolution unit are dynamically adjusted according to the type of property insurance.

[0006] Furthermore, the multi-dimensional features are input into the risk assessment model for evaluation, and the output is an assessment result including the risk level, including: The multi-dimensional features are input into the risk assessment model; wherein, the risk assessment model includes a basic assessment network and a dynamic correction network; the basic assessment network is used to perform preliminary risk quantification on the multi-dimensional features; the dynamic correction network accesses real-time updated industry risk warning data and regional property insurance payout fluctuation index, captures real-time risk factors related to the current assessment object through an attention mechanism, and dynamically corrects the output of the basic assessment network to obtain the risk quantification value. Based on the configured adaptive risk threshold range, the risk quantification value is mapped to a preset multi-level risk level system, and an assessment report containing risk level and key risk point annotations is output.

[0007] Furthermore, the system dynamically corrects the real-time updated industry risk warning data and regional property insurance payout volatility index via the network access, captures real-time risk factors relevant to the current assessment object through an attention mechanism, and dynamically corrects the output of the basic assessment network to obtain a quantified risk value, including: The industry risk warning data and regional property insurance claims volatility index that are accessed are structured to form a risk factor matrix; A three-level screening process based on an attention mechanism is constructed to screen the risk factor matrix; when screening by spatial correlation in the first level, a geographical risk diffusion coefficient is introduced, and the geographical risk diffusion coefficient is dynamically adjusted according to the protection level of the property of the assessed object. When screening by time correlation in the second level, the risk exposure cycle of the insured property is taken into account, and the time attention decay mechanism is used to assign decreasing weights to factors that have exceeded the risk exposure cycle but still have potential impact. When filtering by business relevance at the third level, knowledge graph reasoning is used to supplement implicit business relevance, and the latest property insurance industry standards and policy provisions are accessed in real time to automatically update the relevance rules. The number of factors retained is dynamically adjusted according to the risk complexity of the assessed object. For the core risk factors after screening, a bivariate coupling model is used to calculate the correction coefficient, which is: correction coefficient = main coefficient × (1 + adjustment coefficient × risk hedging coefficient). Substitute the correction coefficient into the output of the basic assessment network, and perform graded corrections according to the methods of immediate correction and delayed correction to obtain the risk quantification value.

[0008] Furthermore, the principal coefficient is obtained by multiplying the risk factor intensity value by the historical influence coefficient. The historical influence coefficient introduces a decay factor, and the decay rate is related to the type of risk factor. The adjustment coefficient is dynamically adjusted based on the current market risk sentiment index, and the risk hedging coefficient is calibrated in real time in conjunction with the risk reserve balance.

[0009] Furthermore, the multi-dimensional features are input into the risk assessment model for evaluation, and the output is an assessment result including the risk level, including: The multi-dimensional features are input into the risk assessment model. The risk assessment model adopts a deep belief network architecture. It first performs unsupervised pre-training using historical unlabeled multi-source heterogeneous data to extract the potential distribution features of the data. Then, it performs supervised fine-tuning by combining labeled risk level data. An adversarial training mechanism is introduced during the fine-tuning process to enhance the model's ability to identify abnormal risk features by generating adversarial examples. The generation of adversarial examples is based on the risk mutation pattern in the property insurance field. The risk assessment model uses a hierarchical attention mechanism to assign dynamic attention weights to spatial distribution characteristics, temporal distribution characteristics, and characteristics that conform to business logic, and obtains a comprehensive risk score through weighted calculation; the weight values ​​are updated in real time according to the type of property being assessed and the risk environment in which it is located. The preset risk level threshold is calibrated using a reinforcement learning algorithm. During the calibration process, the initial calibration is performed by combining the matching degree between the risk level and the actual loss in the claims data of the past year, and the sudden risk factors of the property's location are incorporated to perform a second verification and correction of the initial risk level, thus obtaining the calibrated risk level threshold. Based on the comprehensive risk score and the calibrated risk level threshold, an assessment result including the risk level is generated.

[0010] This invention also provides a personalized property insurance underwriting risk assessment system based on big data, comprising: The acquisition module is used to acquire multi-source heterogeneous data; the multi-source heterogeneous data includes customer property insurance policy data, claims records, and property insurance industry risk data based on big data analysis. An extraction module is used to extract multi-dimensional features from the multi-source heterogeneous data based on a dynamic heterogeneous fusion module. The dynamic heterogeneous fusion module is composed of a spatial convolutional unit, a temporal memory unit, and a business rule unit connected in parallel. The spatial convolutional unit extracts spatial distribution features, the temporal memory unit extracts temporal distribution features, and the business rule unit extracts features that conform to business logic. The assessment module is used to input the multi-dimensional features into the risk assessment model for assessment and output assessment results including risk levels.

[0011] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0012] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the methods described above.

[0013] This invention provides a personalized property insurance underwriting risk assessment method and system based on big data, comprising: acquiring multi-source heterogeneous data; the multi-source heterogeneous data includes customer property insurance policy data, claims records, and property insurance industry risk data based on big data analysis; extracting multi-dimensional features from the multi-source heterogeneous data based on a dynamic heterogeneous fusion module; the dynamic heterogeneous fusion module is composed of a spatial convolutional unit, a temporal memory unit, and a business rule unit connected in parallel; the spatial convolutional unit extracts spatial distribution features, the temporal memory unit extracts temporal distribution features; the business rule unit extracts features that conform to business logic; and inputting the multi-dimensional features into a risk assessment model for evaluation, outputting an assessment result including risk level. In this invention, spatial distribution features are extracted by the spatial convolutional unit, temporal distribution features are extracted by the temporal memory unit, and features that conform to business logic are extracted by the business rule unit; deep features are extracted from multiple dimensions, making the final property insurance underwriting risk assessment result more accurate. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the steps of a personalized property insurance underwriting risk assessment method based on big data in one embodiment of the present invention; Figure 2 This is a block diagram of a personalized property insurance underwriting risk assessment system based on big data, according to one embodiment of the present invention. Figure 3 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention.

[0015] The implementation, functional features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] Reference Figure 1 One embodiment of the present invention provides a personalized property insurance underwriting risk assessment method based on big data, comprising the following steps: Step S1: Obtain multi-source heterogeneous data; the multi-source heterogeneous data includes customer property insurance policy data, claims records, and property insurance industry risk data based on big data analysis. Step S2: Multi-dimensional feature extraction is performed on the multi-source heterogeneous data based on the dynamic heterogeneous fusion module; the dynamic heterogeneous fusion module is composed of a spatial convolution unit, a temporal memory unit, and a business rule unit connected in parallel; the spatial convolution unit extracts spatial distribution features, the temporal memory unit extracts temporal distribution features, and the business rule unit extracts features that conform to business logic; Step S3: Input the multi-dimensional features into the risk assessment model for evaluation, and output the assessment result including the risk level.

[0018] In this embodiment, as described in step S1 above, this step aims to construct a comprehensive and diverse risk assessment data foundation to provide data support for subsequent feature extraction and model evaluation. The multi-source heterogeneous data covers three core types of data: customer property insurance policy data, including key information such as policyholder identity information, type, quantity, value, policy term, and underwriting terms, which serves as the basis for assessing individual risk; claims records, including detailed data such as the time, location, cause, payout amount, and processing result of historical claims, reflecting the historical risk exposure and payout patterns of the insured property; and property insurance industry risk data based on big data analysis, sourced from industry statistical platforms, third-party data service providers, and publicly available information channels, covering macro and meso-level information such as the overall loss ratio, risk event incidence rate, impact of industry policy changes, and the role of the macroeconomic environment in the property insurance market across different regions and types of property insurance.

[0019] Through a multi-channel data collection mechanism, data from different structures (structured policy fields, semi-structured claims reports, and unstructured industry analysis texts) and sources (internal business systems of insurance companies and external cooperation platforms) are aggregated. During the collection process, data integrity and consistency are ensured through technologies such as data interface standardization, format conversion, and redundancy verification, laying a high-quality data foundation for subsequent multi-dimensional feature extraction.

[0020] As described in step S2 above, this step utilizes the parallel processing architecture of the dynamic heterogeneous fusion module to achieve in-depth mining of potential risk features in multi-source data, overcoming the limitations of traditional single feature extraction. This module consists of three functional units connected in parallel, each focusing on extracting features of a specific dimension, and forming a comprehensive feature set through a collaborative mechanism.

[0021] Spatial Convolution Unit: This unit extracts features from the geospatial information contained in the data. Using a pre-defined convolution kernel, it performs sliding calculations on spatial coordinate data such as the location of the insured property, the location of historical claims events, and high-risk areas within the industry. This identifies risk distribution patterns at different spatial scales, such as the clustering risk of a certain type of property in a specific geographical area, the differences in loss ratios across different areas of a city, and the spatial correlation between high-risk areas of natural disasters and property insurance risks. During the convolution process, the kernel parameters are dynamically adjusted according to the property type (e.g., residential, commercial buildings, industrial plants) to adapt to the spatial risk characteristics of different properties.

[0022] Temporal Memory Unit: Employing a temporal deep learning architecture (such as Long Short-Term Memory networks), this unit captures the temporal features of the data. For time series of claims records, it extracts trends in the frequency of risk events as a function of seasons and years, such as the cyclical impact of flood season on property insurance claims. For the time distribution of policy data, it analyzes the correlation between insurance behavior and risk exposure at different times, such as the risk aversion motivation behind the year-end peak in corporate property insurance purchases. Simultaneously, by combining time series of industry risk data, it identifies the long-term impact and short-term fluctuations of macroeconomic risk factors (such as economic cycles and policy adjustments). Through dynamic control of memory and forget gates, this unit effectively retains the characteristic information of key time nodes and filters out noise interference.

[0023] Business Rules Unit: Based on the company's underwriting business rules system and general standards in the property insurance industry, this unit extracts features from data at the business logic level. It includes a built-in structured business rules library, containing standards for underwriter qualification review (such as credit rating thresholds for policyholders and property valuation standards), underwriting risk grading conditions (such as the correspondence between property insurance coverage and risk levels), and rules for considering special risk factors (such as depreciation coefficients for aged property and additional underwriting conditions for high-risk industries). Through rule matching and logical reasoning, it extracts features from policy data that meet business access requirements (such as whether the policyholder meets qualification conditions), features from claims records that are related to underwriting liability (such as whether the cause of the claim falls within the scope of insurance liability), and features from industry data that are compatible with business strategies (such as the matching degree between industry risk levels and the company's underwriting limits).

[0024] Ultimately, the features extracted from the three units are combined through a dynamic weighted fusion mechanism to form a multi-dimensional feature set. The weights are dynamically adjusted according to the contribution of different features in historical risk assessment, ensuring that the feature set can cover multi-dimensional information such as space, time series, and business, and highlight key features that have a significant impact on risk assessment.

[0025] As described in step S3 above, this step uses a risk assessment model to comprehensively analyze the fused multi-dimensional features, thereby achieving a quantitative assessment and classification of personalized underwriting risks. The risk assessment model is built based on machine learning algorithms, and its training data comes from historical underwriting cases of insurance companies and corresponding risk level labels. The model structure has undergone multiple iterations and optimizations to adapt to property insurance risk assessment scenarios.

[0026] First, a multi-dimensional feature set is input into the model. The model then uses nonlinear transformations and feature combinations to quantify the potential risks of the insured property, generating a preliminary risk quantification value. This quantification value comprehensively reflects the combined impact of spatial distribution characteristics (such as geographical risk), temporal distribution characteristics (such as risk change trends), and business logic characteristics (such as business access matching degree) on underwriting risk.

[0027] Subsequently, the model maps the quantified risk value to the corresponding risk level based on a pre-set multi-level risk rating system (such as low, low-medium, medium, medium-high, and high). The threshold for level classification is dynamically adjusted through historical data verification to ensure that the correspondence between the level boundaries and the actual compensation risk remains stable.

[0028] The final assessment results include a specific risk level and an analysis of the contribution of each dimension of characteristics to that level (such as the proportion of spatial factors and the proportion of temporal factors). It also marks key risk points (such as the risk of frequent natural disasters in a certain region recently, and the reasons for high-frequency risks in the policyholder's historical claims records), providing a clear and explainable risk basis for underwriting decisions.

[0029] In one embodiment, acquiring multi-source heterogeneous data includes: Obtain multiple raw data sets; The original data is cleaned to remove duplicate and abnormal data, and then standardized to obtain the multi-source heterogeneous data. In this embodiment, a duplicate data detection algorithm is used to identify and remove redundant records in the original dataset. For example, duplicate insurance application records of the same policyholder and duplicate claim reports generated due to operational errors in the claims system are identified by hash comparison and similarity calculation of key fields (such as policyholder ID and case number) to ensure the uniqueness of each data record. For abnormal data, a dual verification mechanism of statistical analysis and business rules is adopted: outliers in numerical data (such as property value and compensation amount) are identified based on statistical methods such as standard deviation and interquartile range; for the cleaned data, dimensional unification and format standardization are adopted to eliminate heterogeneity between different data sources.

[0030] In one embodiment, multi-dimensional feature extraction is performed on the multi-source heterogeneous data based on a dynamic heterogeneous fusion module, including: By using multi-layer convolution kernels of spatial convolution units, convolution operations are performed on the geographical location information of insured property in the customer property insurance policy data and the regional risk distribution information in the property insurance industry risk data to extract risk clustering features and geographical correlation features at different spatial scales as spatial distribution features. By using the long short-term memory network structure of the temporal memory unit, the time sequence of claims events in the claims records and the insurance time sequence in the customer property insurance policy data are analyzed to extract the risk occurrence frequency characteristics and risk change trend characteristics in the time dimension, and the feature information of key time nodes is dynamically selected as the temporal distribution characteristics. Based on the pre-built property insurance underwriting business rule library, the business rule unit matches and analyzes the policyholder qualification information and property attribute information in the customer property insurance policy data, and extracts features that meet the business access conditions and features related to the underwriting limit as features that conform to the business logic. The features extracted by the spatial convolutional unit, temporal memory unit, and business rule unit are weighted and fused, and the weight values ​​are dynamically determined based on the risk contribution of each feature in historical evaluation cases.

[0031] In this embodiment, the spatial convolution unit utilizes hierarchical operations with multiple convolutional kernels to achieve in-depth mining of geospatial information in the data, thereby accurately extracting spatial distribution features. Specifically, for the geographical location information of insured property recorded in customer property insurance policy data (such as latitude and longitude coordinates and administrative division codes), and the regional risk distribution information contained in property insurance industry risk data (such as risk level classification of different regions and geographical distribution of risk event incidence rates), the spatial convolution unit uses multiple convolutional kernels to perform convolution operations.

[0032] The convolutional kernel settings are dynamically adaptable, with their size and number adjusted according to the type of property insurance (e.g., residential property insurance, commercial property insurance, industrial property insurance): for residential property insurance, smaller kernels are used to capture fine spatial features at the community level; for industrial property insurance, larger kernels are used to cover the broad-area risk correlations surrounding industrial zones. Through multi-layer convolutional operations, two types of core spatial features are extracted at different spatial scales (e.g., street level, district / county level, city level): first, risk clustering features, which identify the concentrated distribution pattern of a certain type of property insurance risk in a specific geographical area, such as the clustering of commercial building property insurance claims in a specific business district of a city; second, geographical correlation features, which explore the risk transmission relationship between different geographical areas, such as the geographical correlation between pollution risks in upstream industrial areas and property insurance risks in downstream residential areas. These extracted features collectively constitute spatial distribution features, providing key geographical dimensions for subsequent risk assessment.

[0033] The temporal memory unit utilizes the unique structure of Long Short-Term Memory (LSTM) networks to capture and filter dynamic risk characteristics in time-series data, forming temporal distribution features. This unit focuses on two types of time-series data: first, the time series of claims events in claims records, including the specific occurrence time and interval of each claims event; and second, the insurance time series in customer property insurance policy data, covering the policyholder's historical insurance application dates, renewal periods, and insurance intervals.

[0034] Long Short-Term Memory (LSTM) networks analyze time series segment by segment through the synergistic action of input gates, forget gates, and output gates: the input gate filters new time-series information and incorporates it into memory; the forget gate selectively forgets outdated or irrelevant historical information to avoid redundant information interference; and the output gate outputs valid memory information based on current evaluation needs. Through this mechanism, the temporal memory unit can extract two core temporal features: risk occurrence frequency features, which statistically analyze the number and frequency changes of claims or insurance activities within a specific time period (e.g., monthly, quarterly, annual), such as the claim frequency characteristics of a certain type of property during the flood season; and risk change trend features, which identify the direction of evolution of risk events or insurance activities over time, such as the year-on-year increase or decrease trend of property insurance coverage in a certain region. Simultaneously, this unit can dynamically filter feature information from key time nodes, such as the timing of major natural disasters or abrupt changes in risk characteristics before and after policy adjustments, ensuring that the temporal distribution features accurately reflect risk patterns over time.

[0035] The business rules unit is based on a comprehensive property insurance underwriting business rules library. Through rule matching and logical reasoning, it extracts features from data that conform to business logic. The property insurance underwriting business rules library contains multi-dimensional business rules, covering rules for policyholder qualification review (such as policyholder credit score thresholds and business qualification requirements), rules for property attribute assessment (such as property value assessment standards and the correspondence between property usage years and risk coefficients), and rules for determining underwriting limits (such as the maximum underwriting amount for different property types and limit adjustment rules linked to the policyholder's risk level). The rules library is updated in real time according to business adjustments and changes in industry management policies.

[0036] For policyholder qualification information (such as credit reports, income statements, and business licenses) and property attribute information (such as property type, purchase date, appraised value, and intended use) in customer property insurance policy data, the business rules unit conducts a two-way matching analysis: On the one hand, it compares the policyholder qualification information with the access conditions in the business rules database to extract features that meet the business access conditions, such as whether the policyholder's credit score reaches the access threshold and whether the business license is valid; on the other hand, based on the association rules between property attribute information and underwriting limits, it extracts features related to underwriting limits, such as the ratio of property appraised value to the maximum underwriting limit and the limit discount coefficient corresponding to the property's useful life. In this way, it ensures that the extracted features are highly consistent with the insurance company's underwriting business logic, providing feature support for risk assessment that aligns with actual business practices.

[0037] After extracting spatial distribution features, temporal distribution features, and features consistent with business logic from each unit, a unified multi-dimensional feature set needs to be formed through weighted fusion to achieve synergistic effects of features from different dimensions. The weights in the fusion process are not fixed but dynamically determined based on the risk contribution of each feature in historical assessment cases. Specifically, a feature contribution evaluation model is constructed to quantitatively analyze the correlation between each feature in historical assessment cases and the final risk assessment result: features that significantly impact risk level classification in historical cases (such as the strong correlation between risk clustering characteristics in a region and the property insurance loss rate in that region) are assigned higher weights; features with weaker impacts are assigned lower weights. This dynamic adjustment mechanism of weights ensures that the fused feature set can adapt to the needs of different assessment scenarios. For example, during periods of frequent natural disasters, the weight of spatial distribution features will increase accordingly; during periods of business policy adjustments, the weight of features consistent with business logic will be enhanced. Through this weighted fusion, the final multi-dimensional feature set retains the unique value of each dimension while highlighting key risk factors through weight allocation, providing comprehensive and focused feature inputs for the accurate calculation of the subsequent risk assessment model.

[0038] In one embodiment, the size and number of convolution kernels of the spatial convolution unit are dynamically adjusted according to the type of property insurance.

[0039] In one embodiment, the multi-dimensional features are input into a risk assessment model for evaluation, and an assessment result including risk level is output, including: The multi-dimensional features are input into the risk assessment model; wherein, the risk assessment model includes a basic assessment network and a dynamic correction network; the basic assessment network is used to perform preliminary risk quantification on the multi-dimensional features; the dynamic correction network accesses real-time updated industry risk warning data and regional property insurance payout fluctuation index, captures real-time risk factors related to the current assessment object through an attention mechanism, and dynamically corrects the output of the basic assessment network to obtain the risk quantification value. Based on the configured adaptive risk threshold range, the risk quantification value is mapped to a preset multi-level risk level system, and an assessment report containing risk level and key risk point annotations is output.

[0040] In this embodiment, firstly, the multi-dimensional features extracted by the dynamic heterogeneous fusion module are input into the risk assessment model. Through the collaborative operation of the basic assessment network and the dynamic correction network, the accurate quantification of underwriting risk is achieved.

[0041] Preliminary Risk Quantification by the Basic Assessment Network: The basic assessment network employs a deep neural network architecture, with training data derived from the risk levels and corresponding multi-dimensional features marked in historical underwriting cases. Through multi-layered nonlinear transformations, the network comprehensively analyzes the spatial distribution characteristics (such as risk clustering and geographical correlation), temporal distribution characteristics (such as risk frequency and trends), and features consistent with business logic (such as business access matching degree and underwriting limit correlation), mapping these features to preliminary risk quantification values. These quantification values ​​reflect the risk level based on historical data patterns, providing a benchmark reference for subsequent adjustments.

[0042] Dynamic correction of the dynamic correction network: As a real-time risk adjustment mechanism, the dynamic correction network accesses two types of real-time data: First, industry risk early warning data, covering policy change early warnings and major risk event reports in the property insurance field (such as regional natural disaster early warnings and industry systemic risk warnings); second, regional property insurance payout volatility index, including real-time statistical indicators such as the payout ratio and payout amount volatility over the past 72 hours divided by administrative region.

[0043] During the correction process, the dynamic correction network uses an attention mechanism to filter features from real-time data: First, it calculates the correlation between each real-time risk factor (such as a regional rainstorm warning or a sudden increase in the loss ratio of a specific industry) and the current assessment object (insured property type, geographical location, industry, etc.). The correlation calculation integrates multi-dimensional parameters such as spatial distance, business attributes, and risk transmission paths. Then, it assigns higher attention weights to real-time risk factors with high correlation, focusing on key influencing factors. Based on the selected real-time risk factors, the dynamic correction network adjusts the preliminary risk quantification value output by the basic assessment network through a preset correction algorithm (such as extrapolation based on the impact coefficient of historical related cases), ultimately obtaining a risk quantification value that integrates historical patterns and real-time dynamics, ensuring that the risk assessment can respond to the latest risk situation.

[0044] Next, by matching the risk quantification value with the adaptive threshold range, the risk level is classified, and an assessment report containing key information is generated, providing a direct basis for underwriting decisions.

[0045] A pre-defined multi-level risk rating system (e.g., low, low-medium, medium, medium-high, high) corresponds to an adaptive risk threshold range. The upper and lower limits of this range are not fixed values ​​but are dynamically adjusted based on the actual claims data of similar property insurance business from the insurance company over the past three months: when the loss ratio of similar business increases, the threshold range shifts upwards to raise the risk level classification standard; when the loss ratio decreases, the threshold range shifts downwards accordingly to avoid over-assessment. The dynamically corrected risk quantification value is compared with the current threshold range; if it falls into a certain range, it is mapped to the corresponding risk level, achieving an objective risk level classification.

[0046] The output assessment report contains two core components: first, a clearly defined risk level, which intuitively reflects the level of underwriting risk; and second, the identification of key risk points. By tracing the constituent factors of the quantified risk value, the report identifies characteristics that significantly affect the risk level, such as high-risk clustering in a certain region, a recent surge in the claims volatility index, and business access characteristics that do not meet optimal standards. The report also marks the impact weight of each key risk point. Furthermore, the report integrates real-time risk factors incorporated into the dynamic correction network, explaining the moderating effect of real-time factors on the risk level (e.g., a rainstorm warning in a certain region leading to a one-level upgrade in the risk level), ensuring the interpretability and business relevance of the assessment results.

[0047] In one embodiment, the dynamic correction network accesses real-time updated industry risk warning data and regional property insurance payout volatility index, captures real-time risk factors related to the current assessment object through an attention mechanism, and dynamically corrects the output of the basic assessment network to obtain a risk quantification value, including: The industry risk warning data and regional property insurance claims volatility index that are accessed are structured to form a risk factor matrix; A three-level screening process based on an attention mechanism is constructed to screen the risk factor matrix; when screening by spatial correlation in the first level, a geographical risk diffusion coefficient is introduced, and the geographical risk diffusion coefficient is dynamically adjusted according to the protection level of the property of the assessed object. When screening by time correlation in the second level, the risk exposure cycle of the insured property is taken into account, and the time attention decay mechanism is used to assign decreasing weights to factors that have exceeded the risk exposure cycle but still have potential impact. When filtering by business relevance at the third level, knowledge graph reasoning is used to supplement implicit business relevance, and the latest property insurance industry standards and policy provisions are accessed in real time to automatically update the relevance rules. The number of factors retained is dynamically adjusted according to the risk complexity of the assessed object. For the core risk factors after screening, a bivariate coupling model is used to calculate the correction coefficient, which is: correction coefficient = main coefficient × (1 + adjustment coefficient × risk hedging coefficient). Substitute the correction coefficient into the output of the basic assessment network, and perform graded corrections according to the methods of immediate correction and delayed correction to obtain the risk quantification value.

[0048] In this embodiment, the dynamic correction network first standardizes the accessed real-time data, transforming unstructured and semi-structured information into computable structured data to construct a risk factor matrix. Specifically, for industry risk warning data (such as policy change texts, natural disaster warning announcements, and industry risk event reports), core risk elements are extracted using natural language processing technologies (such as entity recognition, keyword extraction, and sentiment analysis), including risk event type, scope of impact, severity, and release time, and converted into numerical feature vectors. For regional property insurance payout volatility indices (such as time-series data such as payout rates, payout amount volatility, and case growth rates divided by administrative regions), trend, periodic, and random features are extracted using time-series decomposition algorithms and converted into fixed-dimensional feature vectors.

[0049] After aligning the feature vectors of the two types of data along a unified dimension, they are combined to form a risk factor matrix. Each element in the matrix represents the quantified value of a real-time risk factor in a specific dimension, providing a structured input basis for subsequent screening and correction.

[0050] Next, a three-level screening process is constructed based on the attention mechanism to screen the risk factor matrix. Through the three-level progressive screening, core risk factors that are highly relevant to the current assessment object are accurately extracted from the risk factor matrix, ensuring the pertinence and effectiveness of the correction.

[0051] The first level, screening by spatial correlation: Based on the spatial distance between the location of the assessed property and the area affected by risk factors, spatial correlation is calculated, and a geographical risk diffusion coefficient is introduced to correct the correlation. The geographical risk diffusion coefficient has a preset baseline value based on the natural diffusion characteristics of risk types (such as floods, typhoons, and fires), and is dynamically adjusted in conjunction with the protection level of the assessed property (such as building seismic resistance level, fire protection facility configuration level, flood control dike height, etc.): the higher the protection level, the faster the diffusion and attenuation rate of the risk factor, and the smaller the coefficient value; the lower the protection level, the larger the coefficient value. For example, for the same typhoon warning, among two types of property located at the same distance, the geographical risk diffusion coefficient of an industrial plant with a higher protection level is less than that of an ordinary residential building with a lower protection level, resulting in a lower weight for the typhoon factor in the spatial correlation calculation of the former. Through this mechanism, risk factors with a high correlation to the assessed property within the spatial impact range are screened out.

[0052] The second level involves screening based on time relevance: This involves considering the risk exposure cycle of the insured property (e.g., the usage cycle of seasonal properties, the construction cycle of engineering properties) to determine the temporal effectiveness of risk factors and handling cross-cycle impacts through a time-attention decay mechanism. Real-time risk factors within the risk exposure cycle (e.g., drought warnings during crop growth cycles) are assigned higher time weights. For risk factors exceeding the risk exposure cycle but still potentially having lagged effects (e.g., residual risks during post-disaster reconstruction), decreasing weights are assigned according to a time decay function (e.g., exponential decay), with the decay rate dynamically adjusted based on the sustained impact characteristics of the risk factors (e.g., the long-term persistence of environmental pollution, the sustained timeliness of policy changes). For example, if a chemical industrial park leak occurs outside the risk exposure cycle of the insured property, but its pollution risk has a long-term impact, the time-attention decay mechanism will cause its weight to decrease slowly over time, rather than immediately returning to zero, ensuring a reasonable consideration of lagged risks.

[0053] The third level involves filtering by business relevance: Based on the business attributes of the assessed object, such as its property type, industry, and underwriting terms, explicit and implicit relationships between risk factors and the assessed object are uncovered through knowledge graph reasoning. The knowledge graph contains entities and relationships such as property type, risk factors, industry classification, and underwriting rules. Implicit business relationships are supplemented through rule reasoning (e.g., the implicit relationship between a certain type of precision instrument property and the risk of power supply stability, though not reflected in explicit rules, can be derived through indirect relationships in the knowledge graph). Simultaneously, it connects in real-time to the property insurance industry standard database and policy release platform. When new industry standards (such as updates to property security protection standards) or policy provisions (such as adjustments to coverage scope or new rate management regulations) are released, the association rules in the knowledge graph are automatically updated (e.g., adjusting the association weight between a certain type of property and a specific risk factor). The final number of retained risk factors is dynamically adjusted based on the risk complexity of the assessed object (determined comprehensively by property value, usage complexity, historical risk records, etc.): the higher the risk complexity, the higher the upper limit of the number of retained factors to ensure coverage of multi-dimensional risk impacts; the lower the risk complexity, the fewer the retained factors to avoid redundant information interference.

[0054] Then, for the core risk factors after screening, the correction coefficient is calculated by a bivariate coupling model to quantify the impact of the risk factors on the underwriting risk. The formula is: Correction coefficient = main coefficient × (1 + adjustment coefficient × risk hedging coefficient).

[0055] The main coefficient is obtained by multiplying the risk factor intensity value by the historical impact coefficient. The risk factor intensity value is a quantitative representation of the real-time risk factor (such as the value corresponding to the natural disaster warning level, or the standardized value of the payout volatility); the historical impact coefficient is based on historical data, statistically calculating the average impact weight of this type of risk factor on the final payout result in similar insured cases in the past, and introducing a time decay factor (which decreases as the time since the historical case increases) to ensure that the impact weight of recent cases is higher.

[0056] Adjustment coefficient: Reflects the correction of the impact of current market risk sentiment on risk factors. It is generated based on industry sentiment analysis (such as media reporting trends and market participant risk expectation surveys). The value range is [-0.3, 0.3]. A positive value indicates that market sentiment amplifies the impact of risk, while a negative value indicates that market sentiment weakens the impact of risk.

[0057] Risk hedging coefficient: Determined by combining internal risk tolerance indicators such as the insured scale, reinsurance arrangements, and risk reserve balance in the region and property type of the assessed object. The value range is [0.5, 1.5]. When the company's risk exposure in this area is low, the coefficient value is smaller (e.g., 0.5-0.8), reducing the adjustment range; when the risk exposure is high, the coefficient value is larger (e.g., 1.2-1.5), enhancing the adjustment sensitivity and making the adjustment more in line with the company's actual risk tolerance.

[0058] Finally, the calculated correction coefficients are substituted into the preliminary risk quantification value output by the basic assessment network, and a hierarchical correction mechanism is adopted to achieve dynamic adjustment, balancing real-time performance and stability.

[0059] Immediate correction: For risk factors with urgent and high certainty (such as regional disasters that have occurred or policy changes that have taken effect), an immediate correction of 60% of the correction coefficient is made to quickly respond to sudden risks and ensure that the risk quantification value can reflect the current risk situation in a timely manner.

[0060] Delayed Adjustment: For risk factors with uncertain or lagging effects (such as medium- and long-term weather forecasts and policy drafts), the remaining 40% of the adjustment coefficient is included in the delayed adjustment queue. Adjustments are applied in stages based on the changing trends of the risk factor within the next 24 hours (such as warning level upgrades / downgrades, volatility index convergence / divergence). For example, if a regional rainstorm warning is upgraded during the delay period, the corresponding delayed adjustment magnitude is increased proportionally; if the warning is lifted, this portion of the adjustment is cancelled.

[0061] By using tiered correction, we can avoid excessive volatility that may result from a single correction mode, while ensuring dynamic tracking of risk changes. Ultimately, we obtain a quantitative risk value that integrates historical patterns and real-time dynamics, providing an accurate quantitative basis for subsequent risk level classification.

[0062] In one embodiment, after obtaining the risk quantification value, the process includes: The risk quantification value is verified. During the verification, it is determined whether the risk quantification value is within the preset range, and the rationality of the risk level jump before and after the correction is compared. At the same time, the risk entropy value is calculated. When the entropy value is too high, the additional verification process is automatically initiated. If it is still unreasonable, intelligent backtracking is initiated. The preset range is determined based on the risk quantification value distribution characteristics of similar property insurance business in history.

[0063] In this embodiment, the risk quantification value obtained after dynamic correction undergoes an initial verification, and its basic validity is evaluated through a preset range. This preset range is determined based on the distribution characteristics of risk quantification values ​​of similar historical property insurance businesses. The upper and lower limits of the range are the historical data mean plus or minus three standard deviations (or adjusted to two standard deviations according to business accuracy requirements), covering the quantification results under normal risk scenarios. For example, if the historical data mean is 50 and the standard deviation is 8, the aforementioned preset range is [26, 74].

[0064] Based on a pre-defined multi-level risk rating system, this study analyzes whether the risk level transitions before and after the risk quantification value correction conform to business logic. First, the initial risk quantification value output by the basic assessment network and the dynamically corrected risk quantification value are mapped to the corresponding risk levels (e.g., the initial quantification value corresponds to "medium risk", and the corrected value corresponds to "high risk"), and the direction and magnitude of the level transitions are identified (e.g., transitions across one level, transitions across two or more levels).

[0065] For risk level jumps, the following logic is used to determine their reasonableness: First, check whether the jump is supported by sufficient real-time risk factors. For example, when the risk level jumps from "low" to "high," it is necessary to verify whether there are highly correlated major real-time risk factors (such as a sudden major natural disaster warning in the area where the assessed object is located). Second, refer to the verification results of similar historical jump cases. If the current jump pattern has been proven to be reasonable in historical cases (such as similar risk factor combinations leading to the same magnitude of level change and consistent with the actual compensation results), it is deemed reasonable. If the jump magnitude exceeds the maximum record of similar historical cases and lacks sufficient risk factor support (such as jumping directly from "low risk" to "high risk" without significant real-time risk factors), it is deemed unreasonable and further verification is required.

[0066] The risk entropy value is introduced to quantify the degree of uncertainty behind the risk quantification value. The information entropy algorithm is used to calculate the disorder of the current combination of risk factors: the lower the entropy value, the more consistent the direction of the influence of each risk factor on the quantification result, the clear the weight allocation, and the higher the certainty of the risk assessment; the higher the entropy value, the more conflicting the influence of risk factors (such as some factors pointing to high risk and some factors pointing to low risk) or the ambiguity of the weight allocation, and the greater the uncertainty of the risk assessment.

[0067] A preset entropy threshold (determined based on the correlation analysis between risk entropy and assessment accuracy in historical data) is set up. When the calculated risk entropy exceeds this threshold, an additional verification process is automatically initiated: the historical high-entropy risk case library is called up, and a case matching algorithm is used to find historical cases similar to the current risk factor combination. The correction strategies and actual compensation results of historical cases are referenced to conduct a secondary assessment of the current risk quantification value. At the same time, the collection dimensions of real-time risk factors are increased (such as supplementing micro-risk data within 3 kilometers of the assessment object). By expanding the data sample size, the information uncertainty is reduced, the risk quantification value is recalculated, and the verification is repeated until the entropy value drops below the threshold or the risk itself is confirmed to have high uncertainty characteristics.

[0068] If, after the above verification steps, the risk quantification value is still deemed unreasonable (e.g., exceeding a reasonable range without a reasonable explanation, or the risk level transition not conforming to business logic and failing to be corrected even after additional verification), then the intelligent backtracking mechanism is activated to locate the root cause of the problem and re-optimize the correction process. Through a complete verification and backtracking mechanism, it is ensured that the risk quantification value not only responds to real-time risk dynamics but also conforms to business logic and historical patterns, providing a reliable quantitative basis for subsequent risk level classification.

[0069] In one embodiment, the principal coefficient is obtained by multiplying the risk factor intensity value by the historical influence coefficient, wherein the historical influence coefficient introduces a decay factor, and the decay rate is related to the type of risk factor; The adjustment coefficient is dynamically adjusted based on the current market risk sentiment index, and the risk hedging coefficient is calibrated in real time in conjunction with the risk reserve balance.

[0070] In this embodiment, the risk factor intensity value is an objective quantitative representation of real-time risk factors, and a differentiated quantification method is adopted according to the type of risk factor. For example, for natural disaster early warning factors, an intensity value of 1-4 is assigned according to the early warning level (such as blue, yellow, orange, and red); for regional payout fluctuation factors, quantification is carried out by the deviation of the payout rate from the benchmark value (such as a 10% deviation corresponding to an intensity value of 0.5, and a 30% deviation corresponding to an intensity value of 1.5).

[0071] The historical impact coefficient is calculated based on historical underwriting data and reflects the average weight of the impact of this type of risk factor on the final payout outcome in past cases. For example, if a certain type of rainstorm warning factor has led to an average increase of 20% in the payout ratio in historical data, its historical impact coefficient can be set to 0.2. To avoid interference from outdated data, a decay factor is introduced into the historical impact coefficient, and the decay rate is related to the type of risk factor.

[0072] The adjustment coefficient is used to correct the amplification or weakening effect of market risk sentiment on risk factors, and its value is dynamically determined by the current market risk sentiment index. The market risk sentiment index is calculated using a combination of multi-dimensional data: first, public opinion data within the property insurance industry, such as the sentimental tone of industry news reports and the intensity of discussions about property insurance risks on social media; second, data on the behavior of market participants, such as the frequency of underwriting policy adjustments by insurers and rate fluctuations in the reinsurance market; and third, macroeconomic indicators, such as the business climate index and consumer confidence index. The index typically ranges from -0.3 to 0.3. Positive values ​​indicate a cautious market sentiment (e.g., increased risk warnings in public opinion), amplifying the impact of risk factors; negative values ​​indicate an optimistic market sentiment (e.g., clear signals of industry expansion), weakening the impact of risk factors.

[0073] The risk hedging coefficient is used to adjust the correction range based on the insurance company's own risk tolerance. Its value is calibrated in real time according to the company's dynamically updated risk reserve balance. Risk reserve balance is the fund reserve set aside by an insurance company to cope with potential claims, directly reflecting its risk-bearing capacity. The risk hedging coefficient is positively correlated with the risk reserve balance: when the reserve balance is sufficient (e.g., reaching or exceeding 120% of the management requirement), the company has a stronger risk tolerance, and the risk hedging coefficient is higher (e.g., 1.2-1.5), allowing for a more appropriate adjustment range and a more sensitive response to risk changes; when the reserve balance is tight (e.g., below 100% of the management requirement), the company needs to control its risk exposure, and the risk hedging coefficient is lower (e.g., 0.5-0.8), compressing the adjustment range and avoiding over-assessment leading to business contraction. During the calibration process, the reserve balance data from the company's financial system is connected in real time, and the risk hedging coefficient is dynamically updated according to preset mapping rules (e.g., the coefficient is adjusted by 0.1 for every 10% change in the reserve balance), ensuring that the adjustment result reflects both objective risk and the company's actual operational capabilities.

[0074] In one embodiment, the multi-dimensional features are input into a risk assessment model for evaluation, and an assessment result including risk level is output, including: The multi-dimensional features are input into the risk assessment model. The risk assessment model adopts a deep belief network architecture. It first performs unsupervised pre-training using historical unlabeled multi-source heterogeneous data to extract the potential distribution features of the data. Then, it performs supervised fine-tuning by combining labeled risk level data. An adversarial training mechanism is introduced during the fine-tuning process to enhance the model's ability to identify abnormal risk features by generating adversarial examples. The generation of adversarial examples is based on the risk mutation pattern in the property insurance field. The risk assessment model uses a hierarchical attention mechanism to assign dynamic attention weights to spatial distribution characteristics, temporal distribution characteristics, and characteristics that conform to business logic, and obtains a comprehensive risk score through weighted calculation; the weight values ​​are updated in real time according to the type of property being assessed and the risk environment in which it is located. The preset risk level threshold is calibrated using a reinforcement learning algorithm. During the calibration process, the initial calibration is performed by combining the matching degree between the risk level and the actual loss in the claims data of the past year, and the sudden risk factors of the property's location are incorporated to perform a second verification and correction of the initial risk level, thus obtaining the calibrated risk level threshold. Based on the comprehensive risk score and the calibrated risk level threshold, an assessment result including the risk level is generated.

[0075] In this embodiment, the spatial distribution features, temporal distribution features, and features consistent with business logic extracted by the dynamic heterogeneous fusion module are first input into a risk assessment model employing a deep belief network architecture. The training process of this model is divided into two stages: unsupervised pre-training and supervised fine-tuning, to fully exploit the value of the data and improve the accuracy of risk assessment.

[0076] Unsupervised pre-training: The model first learns autonomously using historical, unlabeled, multi-source heterogeneous data (including policy data, claims records, and industry risk data without risk level labels). Through multi-layer nonlinear transformations of deep belief networks, it extracts latent distribution features from the data layer by layer, such as clustering patterns of risk characteristics of different types of assets and implicit correlations between cross-dimensional features. This stage does not require manual data labeling and can uncover complex risk patterns that are difficult for humans to identify from massive amounts of raw data, laying the foundation for subsequent fine-tuning.

[0077] Supervised fine-tuning: Building upon pre-training, the model undergoes supervised learning by incorporating data labeled with risk levels (i.e., multi-dimensional features with determined risk levels from historical underwriting cases). The network parameters are adjusted using backpropagation to ensure the model's risk assessment results align with actual risk levels. To enhance the model's ability to identify abnormal risks, an adversarial training mechanism is introduced during fine-tuning: Based on risk mutation patterns in the property insurance field (such as a sudden surge in risk due to a major natural disaster in a region, or a sudden change in the nature of risk for a certain type of property due to policy changes), adversarial examples are generated (i.e., perturbation information simulating risk mutations is added to normal feature data). By training the model on these adversarial examples, its sensitivity and discriminative power against extreme risks and abnormal fluctuations are improved, preventing assessment failures due to rare risk events.

[0078] Furthermore, the risk assessment model uses a hierarchical attention mechanism to differentiate and weight the multi-dimensional features of the input, thereby achieving accurate consideration of risk factors in different dimensions.

[0079] The hierarchical attention mechanism assigns dynamic attention weights to three types of features: For spatial distribution features (such as the risk concentration in the area where the insured property is located and geographically related risks), weights are assigned based on the sensitivity of the property type (such as residential, industrial, and commercial buildings) to the geographical environment. For example, commercial buildings in coastal areas have a higher spatial weight for typhoon risk than residential buildings in inland areas. For temporal distribution features (such as the frequency of risk events and risk change trends), weights are dynamically adjusted based on the temporal characteristics of the risk environment in which the property is located (such as increasing the weight of seasonal temporal features for property insurance during the flood season). For features that conform to business logic (such as the matching degree of the insured's qualifications and the correlation of the underwriting limit), weights are assigned based on the current business policy orientation (such as increasing the weight of the corresponding feature if qualification review is emphasized in a certain period).

[0080] The weighting update mechanism is linked to the real-time risk environment: when a risk warning is issued for the area where the assessed property is located (such as a blue rainstorm warning), the weight of the spatial distribution characteristics of that area is automatically increased; when the claim records of a certain type of property show temporal anomalies (such as a sudden increase in the frequency of claims in the past three months), the weight of the temporal distribution characteristics is immediately adjusted upwards. Through weighted calculation, the three types of characteristics are integrated into a comprehensive risk score, which comprehensively reflects the cumulative impact of multi-dimensional risk factors, providing a quantitative basis for risk level classification.

[0081] The preset multi-level risk levels (such as low, low-medium, medium, medium-high, and high) correspond to fixed initial level thresholds. To ensure that the thresholds fit the actual risk scenarios, a reinforcement learning algorithm is used for dynamic calibration to ensure the matching accuracy between the risk level and the actual loss.

[0082] Initial calibration: Using claims data from the past year as a sample, the matching degree between each risk level and the corresponding actual loss is calculated (e.g., whether the actual payout ratio corresponding to the "high-risk" level is significantly higher than that of the "medium-risk" level). The reinforcement learning algorithm uses "improving the matching degree" as the target reward, adjusting the level threshold through multiple iterations: if the actual loss of a certain level is generally higher than expected, the lower limit of the threshold for that level is raised to reduce the probability of misclassification as a low level; if the actual loss is generally lower than expected, the lower limit of the threshold is lowered to avoid over-evaluation. Through initial calibration, the level thresholds are made to conform to the statistical patterns of historical risk losses.

[0083] Secondary verification and correction: Based on the initial calibration, dynamic adjustments are made by incorporating sudden risk factors in the area where the property is located (such as earthquakes, regional fires, and other events that significantly impact losses in the short term). For areas experiencing sudden risk events, the algorithm automatically lowers the corresponding threshold (e.g., lowering the "medium-high risk" threshold to classify more cases as high-risk) to reflect the amplifying effect of real-time risks on losses. For areas where the risk event has subsided, the threshold is gradually restored to the initial calibration level according to a preset attenuation curve. Through secondary verification, the level thresholds can reflect both long-term patterns and respond to the impact of short-term sudden risks, ultimately resulting in calibrated level thresholds that are both stable and timely.

[0084] Finally, the comprehensive risk score is compared with the calibrated risk level threshold. The corresponding risk level is determined based on the threshold range in which the score falls, and an assessment result is generated. For example, if the comprehensive risk score is 85 points and the calibrated threshold range for "high risk" is 80-100 points, then the assessment result is determined to be "high risk".

[0085] The assessment results can include not only the risk level, but also the contribution percentage of each dimension's characteristics to the score (e.g., spatial characteristics 30%, temporal characteristics 40%, business characteristics 30%), as well as annotations of key risk points (e.g., "A 20% increase in the claims rate in a certain region over the past month has led to an increase in the weight of temporal characteristics"). This structured presentation allows the assessment results to both intuitively reflect the risk level and trace the causes of the risk, providing an interpretable and actionable basis for underwriting decisions.

[0086] In one embodiment, after outputting the assessment results including the risk level, the process includes: Keyword extraction is performed on the evaluation results, and the keywords are converted into distinct characters; Obtain multiple industries associated with the property insurance industry as associated industries; construct an undirected graph of industry associations based on the property insurance industry and the associated industries; wherein each industry is a node in the undirected graph, and the industries are connected by edges, with the edge value being the degree of association between the two industries; The characters are sequentially replaced onto each node of the industry association undirected graph to obtain the character association undirected graph; Obtain the customer's type and match the corresponding encoding table based on the customer's type; Based on the character association undirected graph, the encoding table is mutated to obtain a mutated encoding table; The multi-source heterogeneous data is encoded based on the aforementioned variant encoding table and stored in the insurance company's database.

[0087] In this embodiment, the output assessment results, including risk levels, are first structurally parsed. Keyword extraction algorithms from natural language processing are used to filter out keywords reflecting core risk characteristics. These keywords include, but are not limited to, risk level, risk triggers, asset characteristics, and geographic identifiers. After extraction, character mapping rules are used to convert each keyword into unique and non-repeating characters. Character selection follows industry-standard encoding conventions, such as using a combination of uppercase letters (AZ) and numbers (0-9) to achieve uniqueness. Furthermore, character length is positively correlated with keyword importance (e.g., the character length for risk level is 3 characters, and the character length for geographic identifier is 2 characters), ensuring character distinguishability in subsequent association operations.

[0088] Next, through supply chain analysis and industry data mining, multiple industries directly or indirectly related to the property insurance industry are identified as related industries. The selection criteria for related industries include the frequency of business interactions (e.g., the frequency of underwriting cooperation between property insurance and the construction industry), risk transmission paths (e.g., production accidents in the manufacturing industry may trigger property insurance claims), and the degree of data sharing (e.g., the level of customer information exchange between the insurance and banking industries). Typical related industries include construction, manufacturing, finance, transportation, and meteorological services.

[0089] Using the property insurance industry and selected related industries as nodes, an undirected graph of industry relationships is constructed. Edges between nodes represent the relationships between industries, with each edge assigned a correlation degree. This correlation degree is calculated using quantitative indicators: based on data such as the proportion of business cooperation amount between industries over the past three years, the number of jointly participated projects, and the frequency of co-occurrence of risk events, weights are determined using the analytic hierarchy process (AHP) and then comprehensively calculated. The value ranges from [0,1], with higher values ​​indicating a stronger correlation (e.g., the correlation between the property insurance industry and the construction industry might reach 0.8, while the correlation with the education industry might be 0.2). The undirected graph is constructed using an adjacency matrix storage method from graph theory, with matrix elements corresponding to the correlation degrees of the edges, ensuring a structured expression of industry relationships.

[0090] Furthermore, the nodes (industries) in the undirected graph of industry associations are ranked according to the degree of correlation between the industry and the evaluation results. The ranking rule is as follows: industries directly related to the keywords in the evaluation results (such as logistics industry ranking first when the evaluation results involve "warehousing facilities") are ranked first, and industries with weaker correlations are ranked later.

[0091] Based on the sorting results, the generated distinct characters are sequentially replaced to their corresponding nodes: characters corresponding to core keywords such as risk level and risk trigger are prioritized for allocation to the property insurance industry node and the top N industry nodes with the highest relevance (N being the number of keywords). The remaining characters are then allocated to other related industry nodes in descending order of relevance. After the replacement, an undirected graph of character associations is formed, where nodes are characters and edge assignments maintain the original industry-to-industry relevance, realizing the mapping from "industry association" to "character association" and providing the association structure foundation for subsequent encoding variations.

[0092] Then, the customer type characteristics of the assessment targets are extracted through the customer information system. Customer type classification is based on customer attributes (e.g., individual customers / corporate customers), company size (e.g., micro-enterprises / large groups), insurance history (e.g., first-time policyholders / renewing customers), and asset size (e.g., high-value asset customers / ordinary asset customers), etc., with classification standards consistent with the insurance company's business classification system. Based on the classified customer type, the corresponding basic coding table is matched from a pre-set coding table library. The coding table library contains standardized coding rules for different customer types.

[0093] Next, using the structural features of the character-associated undirected graph as mutation rules, the matched basic encoding table is adaptively adjusted. Specific mutation methods include: Field order adjustment: Based on the connection relationship (edge ​​correlation) of character nodes in the undirected graph, the field positions of the coding table are reordered. The field positions corresponding to characters with higher correlation are closer in the coding table (e.g., the characters corresponding to the property insurance industry and the construction industry have high correlation, so the "property type" field and the "construction qualification" field are adjacent in the coding table). Field length adjustment: Adjust the length of the corresponding field based on the degree (number of edges connected) of the character node. For characters with higher degree (i.e. more related industries), the length of the corresponding field will be increased by 1-2 bits to accommodate more related information. Check bit generation rule adjustment: The character corresponding to the edge with the maximum correlation in the undirected graph is used as the seed character for check bit generation. A new check bit rule is generated through a hash algorithm (such as CRC32) to ensure the uniqueness and integrity of the mutated encoding table.

[0094] During the mutation process, the core identification function of the coding table remains unchanged, and only industry-related characteristics are incorporated through structural adjustments to form a mutated coding table adapted to the current evaluation object.

[0095] Finally, the multi-source heterogeneous data (including customer property insurance policy data, claims records, industry risk data, etc.) are transformed using variation coding tables according to data type (structured data / unstructured data). After coding, the coding results are classified according to customer type and risk level and stored in the insurance company's distributed database. The database adopts a partitioned storage strategy, with each partition corresponding to one type of variation coding table, and a metadata index of a character association undirected graph is established. This allows for tracing back the original data and industry relationships through the coding results, while also meeting the security requirements of data encryption and access control.

[0096] In one embodiment, after outputting the assessment results including the risk level, the process includes: Key information is extracted from the evaluation results and converted into a unique feature code; Based on the influence range characteristics of the dominant risk type, a risk feature tree is constructed. The dominant risk type is taken as the root node, and secondary risk factors and risk impact dimensions are derived sequentially downwards to form a multi-level tree structure. Each node is marked with the probability range of the occurrence of the risk feature. The feature codes are sequentially mapped to the corresponding nodes of the risk feature tree. During the mapping process, the association relationship is recorded through node path encoding to form a feature code-risk feature mapping table. Obtain customer insurance preference data and historical claims data from the past three years, extract multiple key behavioral dimensions, and construct a customer behavior correlation matrix; Using the hierarchical weights of the risk feature tree as adjustment parameters, the customer behavior correlation matrix is ​​updated with weights to generate a variable behavior correlation matrix; Using the eigenvectors of the correlation matrix of mutation behavior as encryption keys, customer sensitive information in multi-source heterogeneous data is segmented and encrypted. The encrypted data is classified and stored in the insurance company's distributed database according to the hierarchical structure of the risk feature tree, and each data segment is accompanied by a feature code index.

[0097] In this embodiment, the assessment results, including risk levels, are subjected to in-depth analysis, and key information is extracted using information extraction algorithms (such as BERT-based entity recognition models). After extraction, each key piece of information is converted into a unique feature code using an encrypted hash algorithm (such as SHA-256). The feature code generation process incorporates a timestamp and a unique identifier of the assessed object (such as a policy number), ensuring that even the same key information can generate differentiated feature codes in different assessment scenarios. Furthermore, each feature code has a fixed length (such as 256 bits), satisfying the requirements of uniqueness and collision resistance, providing a reliable identification foundation for subsequent mapping and encryption.

[0098] Next, based on the dominant risk types identified in the assessment results, a risk feature tree is constructed by combining their impact range characteristics. The impact range characteristics of the dominant risk types include the radius of influence (e.g., the radius of influence of typhoon risk can reach hundreds of kilometers), the speed of diffusion (e.g., the diffusion speed of fire risk increases with wind speed), and the duration (e.g., the duration of flood risk can reach several weeks). These characteristics determine the hierarchical division logic of the risk feature tree. The risk feature tree adopts a multi-level tree structure: the root node is the selected dominant risk type (e.g., "typhoon risk"); the first-level child nodes are secondary risk factors, that is, the specific risk manifestations derived from the dominant risk type (e.g., "typhoon risk" derives to "strong wind damage," "heavy rain flooding," and "storm surge"); the second-level child nodes are risk impact dimensions, that is, the specific impacts that each secondary risk factor may cause (e.g., "strong wind damage" derives to "structural damage," "equipment overturning," and "external object impact"); and so on, until a complete tree structure containing 3-5 levels of nodes is formed. Each node is labeled with the probability range of the corresponding risk feature (e.g., the node for "strong wind damage" is labeled "0.3-0.5", indicating that the probability of this risk occurring within the assessment period is between 30% and 50%). The probability range is determined based on the statistical data of similar risk events from Guoren Insurance over the past five years.

[0099] Then, based on the correlation between key information and each node in the risk feature tree, the feature codes generated in step one are sequentially mapped to their corresponding nodes. The correlation is quantified through semantic similarity calculation (such as cosine similarity algorithm). For example, the feature code corresponding to "industrial equipment worth over ten million" has a semantic similarity of 0.85 with the node "equipment overturning," so it is preferentially mapped to that node. The feature code corresponding to "high risk" has the highest correlation with the root node "typhoon risk," so it is directly mapped to the root node. During the mapping process, the correlation between the feature code and the risk feature is recorded through node path encoding. The node path encoding adopts a hierarchical sequence number combination method (e.g., the root node is "0," the first-level child nodes are "0-1" and "0-2," and the second-level child nodes are "0-1-1" and "0-1-2"). Each feature code corresponds to a unique path code, clearly reflecting its position in the risk feature tree. Finally, a feature code-risk feature mapping table is formed, which contains fields such as feature code, corresponding risk feature description, node path encoding, and correlation probability, realizing the structured association between the feature code and the risk feature tree.

[0100] Next, customer insurance preference data and historical claims data for the past three years were collected. Insurance preference data included purchase channels (e.g., online self-service purchase, agent-based purchase), purchase periods (e.g., annual purchase, quarterly purchase), and optional riders (e.g., whether to purchase "waiver of deductible" or "extended liability insurance"). Historical claims data included the cause of claims (e.g., "loss due to natural disaster" or "loss due to human error"), claim processing time (e.g., "reporting within 24 hours" or "reporting after 72 hours"), and claim amount range (e.g., "below 10,000 yuan" or "above 100,000 yuan"). Feature engineering was performed on the above data to extract 3-5 key behavioral dimensions (e.g., "risk aversion tendency," "timeliness of claims response," and "efficiency of insurance decision-making"). Each dimension is a comprehensive indicator (range [0,1]) after standardization of the original data. A customer behavior correlation matrix was constructed using the key behavioral dimensions as row and column indicators. The elements in the matrix are the correlation coefficients between two behavioral dimensions (calculated using the Pearson correlation coefficient, range [-1,1]).

[0101] Then, based on the hierarchical weights of the risk feature tree, the customer behavior correlation matrix is ​​updated with weights. The hierarchical weights are determined according to the importance of each node in the risk assessment: the root node (dominant risk type) has the highest weight (e.g., 0.4), the secondary risk factor nodes have the next highest weight (e.g., 0.3), the risk impact dimension nodes have a lower weight (e.g., 0.2), and the bottom-level nodes have the lowest weight (e.g., 0.1). Furthermore, the weights of nodes within the same level are dynamically allocated according to the probability interval of occurrence (the higher the probability, the greater the weight).

[0102] The weighted update is performed by multiplying the element values ​​(correlation coefficients) in the customer behavior correlation matrix that are related to high-weight risk characteristics by the corresponding level's weight coefficient. For example, if the "risk aversion tendency" dimension is closely related to "typhoon risk" (root node, weight 0.4), then the correlation coefficients of this dimension with other dimensions are multiplied by 0.4; if "timeliness of claims response" is closely related to "rainstorm flooding" (secondary node, weight 0.25), then the corresponding correlation coefficients are multiplied by 0.25. The updated matrix is ​​the variant behavior correlation matrix, whose element values ​​retain the inherent correlation patterns of customer behavior while incorporating the hierarchical influence of risk characteristics, making the matrix more closely aligned with the risk scenario of the current assessment object.

[0103] Finally, the correlation matrix of mutation behavior is decomposed to extract its principal feature vector, which is then used as the encryption key. The encryption process employs a segmented encryption strategy: sensitive customer information (such as customer ID number, property address, bank account information, income status, etc.) from multi-source heterogeneous data is segmented according to information type (such as identity information, property information, and financial information). Each segment is encrypted using a corresponding feature vector sub-vector (e.g., identity information corresponds to the first principal feature vector, and property information corresponds to the second principal feature vector), ensuring the encryption strength meets financial industry security standards. After encryption, the data is categorized and stored in the insurance company's distributed database according to the hierarchical structure of the risk feature tree: the encrypted data corresponding to the root node is stored as a first-level partition, the encrypted data corresponding to the secondary node is stored as a second-level partition, and so on. Each data segment is accompanied by a corresponding feature code index, which contains the association information of the feature code, node path code, and encryption key. This allows for reverse lookup of the risk features corresponding to the data segment through the feature code, enabling data tracing and risk correlation analysis. Simultaneously, the database is configured with an access control mechanism, ensuring that personnel in different positions can only access the data partitions relevant to their responsibilities, thus guaranteeing the security of sensitive information.

[0104] Reference Figure 2 In another embodiment of the present invention, a personalized property insurance underwriting risk assessment system based on big data is also provided, comprising: The acquisition module is used to acquire multi-source heterogeneous data; the multi-source heterogeneous data includes customer property insurance policy data, claims records, and property insurance industry risk data based on big data analysis. An extraction module is used to extract multi-dimensional features from the multi-source heterogeneous data based on a dynamic heterogeneous fusion module. The dynamic heterogeneous fusion module is composed of a spatial convolutional unit, a temporal memory unit, and a business rule unit connected in parallel. The spatial convolutional unit extracts spatial distribution features, the temporal memory unit extracts temporal distribution features, and the business rule unit extracts features that conform to business logic. The assessment module is used to input the multi-dimensional features into the risk assessment model for assessment and output assessment results including risk levels.

[0105] In this embodiment, the specific implementation of each module in the above system embodiment is described in the above method embodiment, and will not be repeated here.

[0106] Reference Figure 3 This invention also provides a computer device, which can be a server, and its internal structure can be as follows: Figure 3As shown, the computer device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0107] Those skilled in the art will understand that Figure 3 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the computer devices on which the present invention is applied.

[0108] An embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0109] In summary, the personalized property insurance underwriting risk assessment method and system based on big data provided in this embodiment of the invention includes: acquiring multi-source heterogeneous data; the multi-source heterogeneous data includes customer property insurance policy data, claims records, and property insurance industry risk data based on big data analysis; extracting multi-dimensional features from the multi-source heterogeneous data based on a dynamic heterogeneous fusion module; the dynamic heterogeneous fusion module is composed of a spatial convolution unit, a temporal memory unit, and a business rule unit connected in parallel; the spatial convolution unit extracts spatial distribution features, the temporal memory unit extracts temporal distribution features; the business rule unit extracts features that conform to business logic; and the multi-dimensional features are input into a risk assessment model for evaluation, outputting an assessment result including risk level. In this invention, spatial distribution features are extracted by the spatial convolution unit, temporal distribution features are extracted by the temporal memory unit, and features that conform to business logic are extracted by the business rule unit; deep features are extracted from multiple dimensions, making the final property insurance underwriting risk assessment result more accurate.

[0110] Those skilled in the art will understand that all or part of the processes in 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 present invention and embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM, etc.

[0111] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0112] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A personalized property insurance underwriting risk assessment method based on big data, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data; the multi-source heterogeneous data includes customer property insurance policy data, claims records, and property insurance industry risk data based on big data analysis; Multi-dimensional feature extraction is performed on the multi-source heterogeneous data based on the dynamic heterogeneous fusion module; The dynamic heterogeneous fusion module is composed of a spatial convolutional unit, a temporal memory unit, and a business rule unit connected in parallel; the spatial convolutional unit extracts spatial distribution features, the temporal memory unit extracts temporal distribution features, and the business rule unit extracts features that conform to business logic. The multi-dimensional features are input into the risk assessment model for evaluation, and the output includes the assessment results containing the risk level.

2. The personalized property insurance underwriting risk assessment method based on big data according to claim 1, characterized in that, Acquiring multi-source heterogeneous data, including: Obtain multiple raw data sets; The original data is cleaned to remove duplicate and abnormal data, and then standardized to obtain the multi-source heterogeneous data.

3. The personalized property insurance underwriting risk assessment method based on big data according to claim 1, characterized in that, The multi-dimensional feature extraction of the multi-source heterogeneous data is performed based on the dynamic heterogeneous fusion module, including: By using multi-layer convolution kernels of spatial convolution units, convolution operations are performed on the geographical location information of insured property in the customer property insurance policy data and the regional risk distribution information in the property insurance industry risk data to extract risk clustering features and geographical correlation features at different spatial scales as spatial distribution features. By using the long short-term memory network structure of the temporal memory unit, the time sequence of claims events in the claims records and the insurance time sequence in the customer property insurance policy data are analyzed to extract the risk occurrence frequency characteristics and risk change trend characteristics in the time dimension, and the feature information of key time nodes is dynamically selected as the temporal distribution characteristics. Based on the pre-built property insurance underwriting business rule library, the business rule unit matches and analyzes the policyholder qualification information and property attribute information in the customer property insurance policy data, and extracts features that meet the business access conditions and features related to the underwriting limit as features that conform to the business logic. The features extracted by the spatial convolutional unit, temporal memory unit, and business rule unit are weighted and fused, and the weight values ​​are dynamically determined based on the risk contribution of each feature in historical evaluation cases.

4. The personalized property insurance underwriting risk assessment method based on big data according to claim 3, characterized in that, The size and number of convolution kernels of the spatial convolution unit are dynamically adjusted according to the type of property insurance.

5. The personalized property insurance underwriting risk assessment method based on big data according to claim 1, characterized in that, The multi-dimensional features are input into the risk assessment model for evaluation, and the output includes an assessment result containing the risk level, including: The multi-dimensional features are input into the risk assessment model; wherein, the risk assessment model includes a basic assessment network and a dynamic correction network; the basic assessment network is used to perform preliminary risk quantification on the multi-dimensional features; the dynamic correction network accesses real-time updated industry risk warning data and regional property insurance payout fluctuation index, captures real-time risk factors related to the current assessment object through an attention mechanism, and dynamically corrects the output of the basic assessment network to obtain the risk quantification value. Based on the configured adaptive risk threshold range, the risk quantification value is mapped to a preset multi-level risk level system, and an assessment report containing risk level and key risk point annotations is output.

6. The personalized property insurance underwriting risk assessment method based on big data as described in claim 5, characterized in that, The system dynamically adjusts the output of the basic assessment network by using real-time updated industry risk warning data and regional property insurance claims volatility index. It captures real-time risk factors relevant to the current assessment object through an attention mechanism, and then dynamically adjusts these factors to obtain a quantified risk value, including: The industry risk warning data and regional property insurance claims volatility index that are accessed are structured to form a risk factor matrix; A three-level screening process based on an attention mechanism is constructed to screen the risk factor matrix; when screening by spatial correlation in the first level, a geographical risk diffusion coefficient is introduced, and the geographical risk diffusion coefficient is dynamically adjusted according to the protection level of the property of the assessed object. When screening by time correlation in the second level, the risk exposure cycle of the insured property is taken into account, and the time attention decay mechanism is used to assign decreasing weights to factors that have exceeded the risk exposure cycle but still have potential impact. When filtering by business relevance at the third level, knowledge graph reasoning is used to supplement implicit business relevance, and the latest property insurance industry standards and policy provisions are accessed in real time to automatically update the relevance rules. The number of factors retained is dynamically adjusted according to the risk complexity of the assessed object. For the core risk factors after screening, a bivariate coupling model is used to calculate the correction coefficient, which is: correction coefficient = main coefficient × (1 + adjustment coefficient × risk hedging coefficient). Substitute the correction coefficient into the output of the basic assessment network, and perform graded corrections according to the methods of immediate correction and delayed correction to obtain the risk quantification value.

7. The personalized property insurance underwriting risk assessment method based on big data as described in claim 6, characterized in that, The principal coefficient is obtained by multiplying the risk factor intensity value by the historical impact coefficient. The historical impact coefficient introduces a decay factor, and the decay rate is related to the type of risk factor. The adjustment coefficient is dynamically adjusted based on the current market risk sentiment index, and the risk hedging coefficient is calibrated in real time in conjunction with the risk reserve balance.

8. A personalized property insurance underwriting risk assessment system based on big data, characterized in that, include: The acquisition module is used to acquire heterogeneous data from multiple sources. The multi-source heterogeneous data includes customer property insurance policy data, claims records, and property insurance industry risk data based on big data analysis; The extraction module is used to extract multi-dimensional features from the multi-source heterogeneous data based on the dynamic heterogeneous fusion module; The dynamic heterogeneous fusion module is composed of a spatial convolutional unit, a temporal memory unit, and a business rule unit connected in parallel; the spatial convolutional unit extracts spatial distribution features, the temporal memory unit extracts temporal distribution features, and the business rule unit extracts features that conform to business logic. The assessment module is used to input the multi-dimensional features into the risk assessment model for assessment and output assessment results including risk levels.

Images