Consensus label generation method and device for multi-source heterogeneous risk scoring

Abstract

The application discloses a multi-source heterogeneous risk score consensus label generation method and device, wherein the method comprises the following steps: acquiring multi-source heterogeneous score data and performing preprocessing to obtain effective score data; based on the effective score data, performing global quantile normalization and score direction unification processing on each score source to obtain global risk quantiles of each score source in a unified scale and a unified risk direction; for each risk assessment scene, calculating a hybrid weight based on the characteristics of the scene; and according to the hybrid weight, fusing the global risk quantiles and local quantiles in the scene to generate a hybrid label of each score source in the scene; for each risk assessment scene, aggregating the hybrid labels of all score sources in the scene to generate a consensus label corresponding to the scene. The application can generate a unified, stable and reliable consensus label from multiple heterogeneous score sources, and the consensus label can be used as a training target of a machine learning model to construct a high-precision risk score model.

Description

Technical Field

[0001] This invention relates to the field of risk information assessment technology, specifically to a method and apparatus for generating consensus labels for multi-source heterogeneous risk scoring. Background Technology

[0002] In today's fields of insurance risk control, credit assessment, financial risk prediction, and e-commerce user profiling, multiple independent data sources or scoring agencies typically conduct risk assessments or score the same target. Taking auto insurance risk scoring as an example, the market includes multiple insurance companies, third-party data service providers, and vehicle inspection agencies, each providing its own risk score for the same vehicle, driver, and policy based on different data dimensions and algorithmic models. These scores reflect the risk characteristics of the target from different perspectives and have high reference value.

[0003] Currently, the most common multi-source scoring fusion methods in the industry mainly include the following categories: The first type is the simple averaging method. This method directly performs an arithmetic average of the scores from multiple rating sources to obtain the merged score. For example, if three rating sources give scores of 85, 90, and 80 respectively, the merged score is (85 + 90 + 80) divided by 3, which equals 85. This method is simple to implement, but it ignores the directional and scale differences between different rating sources. When the rating directions are inconsistent, the scores may cancel each other out, failing to accurately reflect the true risk level.

[0004] The second type is the weighted average method. This method assigns different weights to each rating source based on factors such as historical performance, data quality, and coverage, and then performs a weighted average. The weights can be set manually or obtained through training on historical data. This method is more reasonable than the simple average method, but it still cannot handle the problem of inconsistent rating directions, and the setting of weights often relies on expert experience, making it highly subjective.

[0005] The third type is the local quantile method. This method converts the scores from each rating source into quantile rankings within each sub-segment (such as the same car model, the same region, or the same time period), and then aggregates the quantiles. This method can eliminate differences in rating scales, but when only a few rating sources provide data for a certain scenario, local quantiles are greatly affected by noise, resulting in unstable results.

[0006] The fourth method involves selecting a single rating source. This method chooses a rating source considered the most reliable as the standard and directly uses its rating as the final result. While simple and direct, this method ignores information from other rating sources, potentially leading to information loss, and the choice of which rating source to use as the standard is often controversial.

[0007] The fifth category is the machine learning fusion method. This method uses scores from multiple rating sources as features and inputs them into a machine learning model (such as logistic regression, random forest, or neural network) for training. The model then automatically learns how to fuse these ratings. This method requires a large amount of labeled data as training data, but in practical applications, accurate labels (such as whether an accident occurred and the amount of compensation) are often difficult to obtain or require a long period of accumulation.

[0008] Relevant existing patent technologies include: Chinese patent application CN110826406A, which discloses a risk assessment method based on multi-source data fusion, by acquiring risk assessment results from multiple data sources and fusing them according to preset fusion rules; and Chinese patent application CN111914443A, which discloses a risk control model training method based on federated learning, by integrating data from multiple institutions through a federated learning framework for model training. However, these existing technologies have failed to effectively solve core problems such as inconsistency in direction, inconsistency in scale, and instability with small samples in multi-source heterogeneous scoring.

[0009] After in-depth analysis of the existing technology, the following defects and shortcomings were found: Defect 1: Inconsistent scoring direction. Different scoring sources employ different scoring semantics and directions. Some scoring sources use the semantics that "higher scores represent lower risk," meaning high scores indicate high-quality customers with low risk; while others use the semantics that "higher scores represent higher risk," meaning high scores indicate high-risk customers. For example, in a car insurance scoring scenario, Company A's scoring system might assign 80 points as low risk, while Company B's scoring system might assign 80 points as high risk. If these scores are directly averaged or weighted, the opposing scores will cancel each other out, causing the fusion result to fail to accurately reflect the true risk level. Current technologies typically require manual identification and adjustment of the scoring direction, which is inefficient and prone to errors.

[0010] Defect 2: Inconsistent scoring scales. Different scoring sources use different scoring ranges and distribution patterns. For example, Company A's scoring range is 0 to 100 points, Company B's is 0 to 1000 points, and Company C's is 0.00 to 1.00. Even if the scoring direction is consistent, these scores with different scales cannot be directly compared and merged. Furthermore, the distribution patterns of scores from different scoring sources also differ; some are close to a uniform distribution, some exhibit a normal distribution, and some show a long-tailed distribution. Simple linear normalization (such as min-max normalization) cannot handle these differences in distribution patterns.

[0011] Defect 3: Local quantile ranking is unstable in small sample scenarios. While it theoretically eliminates rating scale differences by calculating quantile rankings within each sub-segment, in practice, some scenarios may only have a few rating sources providing data. For example, for a rare car model or a remote area, only one or two rating sources might have data. In this case, the results of local quantile ranking are greatly affected by individual rating sources, amplifying noise and resulting in unstable and unreliable labels. When there is only one rating source in a scenario, local quantile ranking completely fails.

[0012] Defect 4: Ignoring differences in the credibility of scoring sources. Different scoring sources exhibit significant differences in data quality, algorithm level, and coverage. Some scoring sources may use advanced algorithms and abundant data, resulting in high scoring accuracy; while others may have limited data or outdated algorithms, leading to poor scoring quality. Existing methods such as simple averaging, weighted averaging, and median aggregation either ignore these quality differences or rely on manually set weights, failing to automatically identify and quantify the credibility of scoring sources. Low-quality scoring sources will lower the overall quality of the fusion result.

[0013] Defect 5: Lack of adaptability to scene characteristics. Different scenes have different characteristics, such as the number of rating sources, the dispersion of rating distribution, and sample coverage. An ideal fusion strategy should automatically adjust according to scene characteristics. For example, in scenes with a large number of rating sources, more local information can be relied upon; in scenes with a small number of rating sources, more global information should be relied upon. Existing fusion methods typically use fixed parameters and strategies, which cannot adaptively adjust according to scene characteristics, resulting in poor performance in some scenarios.

[0014] Defect 6: Inability to effectively handle missing data. In practical applications, some scoring sources may lack data in certain scenarios. For example, Company A may only cover passenger vehicles and not commercial vehicles; Company B may only cover certain regions. When some scoring sources are missing in a certain scenario, existing fusion methods often adopt simple ignore or fill strategies, which may lead to information loss or introduce bias.

[0015] Defect 7: Lack of interpretability. While some machine learning-based fusion methods (such as neural networks) may achieve good fusion results, they lack interpretability, making it difficult for business personnel to understand and trust the fusion results. This is a significant defect in fields such as insurance and finance, where interpretability is highly critical. Summary of the Invention

[0016] The main objective of this invention is to provide a consensus label generation method and apparatus for multi-source heterogeneous risk scoring, so as to solve the shortcomings of related technologies.

[0017] To achieve the above objectives, according to a first aspect of the present invention, a method is provided for acquiring multi-source heterogeneous scoring data and performing preprocessing to obtain effective scoring data; based on the effective scoring data, global quantile normalization and scoring direction unification processing are performed on each scoring source to obtain global risk quantiles with a unified scale and risk direction for each scoring source; wherein, the multi-source heterogeneous scoring data consists of risk scoring data of multiple scoring sources for the same batch of target objects; the global quantile normalization is used to eliminate differences in scoring scales among different scoring sources, and the scoring direction unification is used to unify the risk semantics of all scoring sources so that the larger the value, the higher the risk.

[0018] According to a second aspect of the present invention, a consensus label generation apparatus for multi-source heterogeneous risk scoring is provided, comprising: a data acquisition unit for acquiring multi-source heterogeneous scoring data and performing preprocessing to obtain effective scoring data, wherein the multi-source heterogeneous scoring data consists of risk scoring data from multiple scoring sources for the same batch of target objects; a data processing unit for performing global quantile normalization and scoring direction unification processing on each scoring source based on the effective scoring data to obtain global risk quantiles with unified scale and unified risk direction for each scoring source; wherein the global quantile normalization is used to eliminate differences in scoring scales among different scoring sources, and the scoring direction unification is used to unify the risk semantics of all scoring sources to the case where a larger value indicates higher risk; for each risk assessment scenario, a hybrid weight is calculated based on the features of the scenario; and the global risk quantile is fused with the local quantiles within the scenario according to the hybrid weight to generate a hybrid label for each scoring source in the scenario, wherein the scenario features include at least the number of effective scoring sources and the dispersion of the scoring distribution within the scenario; and a consensus label generation unit for aggregating the hybrid labels of all scoring sources within the scenario for each risk assessment scenario to generate a consensus label corresponding to the scenario.

[0019] According to a third aspect of the present invention, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the method described in any implementation of the first aspect.

[0020] This embodiment presents a consensus label generation method and apparatus for multi-source heterogeneous risk scoring. The method includes acquiring multi-source heterogeneous scoring data and performing preprocessing to obtain effective scoring data; based on the effective scoring data, performing global quantile normalization and scoring direction unification processing on each scoring source to obtain global risk quantiles with a unified scale and risk direction for each scoring source; for each risk assessment scenario, calculating a hybrid weight based on the features of that scenario; and fusing the global risk quantile with the local quantiles within that scenario according to the hybrid weight to generate a hybrid label for each scoring source in that scenario, wherein the scenario features include at least the number of effective scoring sources and the dispersion of the scoring distribution within that scenario; and for each risk assessment scenario, aggregating the hybrid labels of all scoring sources within that scenario to generate a consensus label corresponding to that scenario. This method can generate unified, stable, and reliable consensus labels from multiple heterogeneous scoring sources, and these consensus labels can be used as training targets for machine learning models to construct high-precision risk scoring models. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the consensus label generation method for multi-source heterogeneous risk scoring according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the application of the consensus label generation method for multi-source heterogeneous risk scoring in this invention. Figure 3 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of the invention described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0026] According to embodiments of the present invention, a consensus label generation method for multi-source heterogeneous risk scoring is provided, such as... Figure 1 As shown, steps 101 to 103 are included below: Step 101: Obtain multi-source heterogeneous scoring data and perform preprocessing to obtain effective scoring data; wherein, the multi-source heterogeneous scoring data is risk scoring data of multiple scoring sources for the same batch of target objects.

[0027] In this step, multi-source scoring data refers to scoring data from multiple scoring sources (such as different companies, different institutions, and different algorithms) on the same set of target objects (such as vehicles, insurance policies, and users). This data is usually stored in structured tabular form, such as CSV files or database tables.

[0028] As an optional implementation of this embodiment, the acquisition and preprocessing of multi-source scoring data includes: loading the corresponding multi-source scoring data from the original data source based on a predefined field for uniquely identifying a risk assessment scenario and a scoring field; performing data type conversion, invalid value processing, deduplication processing, and masking processing on the data corresponding to each field to obtain valid data, wherein, during masking processing, the scores of a specific scoring source in a specific dimension are set to missing values ​​according to a preset mask configuration.

[0029] In this optional implementation, preprocessing includes determining data fields, exemplarily involving two types of data fields: the first type is the scenario identifier field (scenario_cols), used to uniquely identify a risk assessment scenario. Scenarios are defined by business requirements, and different business scenarios may have different combinations of scenario identifiers.

[0030] For example, in the context of auto insurance pricing, the scenario identifier field might include: Vehicle Identification Number (VIN): Uniquely identifies a vehicle; Date: Scoring date or policy date; Damage Amount: Amount of vehicle damage compensation; Third-Party Liability Amount: Amount of third-party liability compensation; Driver Compensation Amount: Amount of personal injury compensation for the driver; Passenger Compensation Amount: Amount of personal injury compensation for the passenger. The choice of the scenario identifier field depends on business needs and can be a single field or a combination of multiple fields. Using the scenario identifier field, data can be divided into multiple evaluation scenarios, and the scores within each scenario will be merged to generate a consensus label.

[0031] The second category is scoring fields, including: Scoring Source Identifier: the ID of the company or institution from which the score originates; Commercial Score: a score reflecting the business dimension; Compulsory Risk Score: a score reflecting the compulsory risk dimension; and Comprehensive Score: a score reflecting the comprehensive risk. Scoring fields can be expanded according to business needs and are not limited to the three types mentioned above.

[0032] The data preprocessing step also includes data loading. Specifically, data to be processed is read from the original data file according to pre-determined data fields. The original data file may include, but is not limited to, CSV files, Parquet files, database tables, or other structured data sources.

[0033] Taking a CSV file as an example, the data loading process includes: first, reading the header information of the original data file to determine all fields contained in the original data file; then, filtering out the target fields from all fields based on the scene identifier field, the scoring source identifier field, and the scoring value field; finally, reading the original data file based on the target fields to obtain the data to be analyzed. By loading only the target fields related to scoring processing, the amount of data read from irrelevant fields can be reduced, improving data loading efficiency.

[0034] During data loading, missing values ​​are uniformly identified and standardized. These missing values ​​include, but are not limited to, empty strings, "NULL", "null", "None", "nan", and "NaN". The system uniformly converts these different forms of missing values ​​into preset missing value markers to avoid different missing value representations affecting subsequent data calculations.

[0035] Furthermore, the data preprocessing step also includes data type conversion. Specifically, the scoring source identifier field is converted to an integer type, and the scoring value field is converted to a floating-point type; for data that cannot be converted, it is set as a missing value. This ensures data type consistency in subsequent scoring calculations, group statistics, and anomaly identification processes, improving the accuracy and stability of the data processing results.

[0036] Preprocessing also includes handling invalid values. In risk assessment scenarios, score values ​​are typically positive. Zero or negative values ​​may indicate data errors or missing values. Records with score values ​​less than or equal to zero are considered invalid and set as missing values: For each score field, the following processing is performed: convert the score value to a numeric type, setting values ​​that fail to convert as missing values; records with score values ​​less than or equal to zero are set as missing values.

[0037] Preprocessing also includes deduplication, as the original data may contain duplicate records, such as multiple rating records for the same scene or from the same rating source. This invention removes duplicate records based on a combination of the scene identifier field, the rating source identifier, and the rating value, retaining only the last record. The purpose of deduplication is to ensure that each rating source provides only one valid rating for each scene.

[0038] Data preprocessing also includes masking specific scoring sources. In some business scenarios, the scores from certain scoring sources may be unreliable or unavailable for specific dimensions. For example, Company A may only provide business scores, not traffic scores and overall scores. This invention supports masking specific scoring sources for specific dimensions, setting them as missing values. The mask configuration is provided in the form of a mapping table, where the key is the scoring source identifier and the value is a list of scoring dimensions that need to be masked.

[0039] Step 102: Based on the effective scoring data, perform global quantile normalization and scoring direction unification on each scoring source to obtain global risk quantiles with unified scale and risk direction for each scoring source; wherein, the global quantile normalization is used to eliminate the differences in scoring scales between different scoring sources, and the scoring direction unification is used to unify the risk semantics of all scoring sources so that the larger the value, the higher the risk.

[0040] In this step, to eliminate directional and scale differences among different rating sources, the global quantile is calculated for each rating source across all its historical samples, and then normalized according to the rating direction to generate a uniform risk quantile.

[0041] In one optional implementation of this embodiment, the global quantile normalization and direction unification processing for each scoring source includes: pre-configuring the scoring direction of each scoring source; for each scoring source, calculating the global quantile ranking of each scoring value in all its historical scoring samples, wherein the global quantile ranking is the proportion of samples less than or equal to the scoring value to all samples of the scoring source; and generating risk quantiles according to the scoring direction configuration.

[0042] In this optional implementation, scoring directions can be pre-configured based on the scoring semantics of each scoring source. The scoring direction represents the correspondence between the score value and the risk level. For example, the scoring direction includes a first scoring direction and a second scoring direction. The first scoring direction indicates that a lower score value corresponds to higher sample quality or lower risk; that is, a higher score value corresponds to higher risk. The second scoring direction indicates that a higher score value corresponds to higher sample quality or lower risk; that is, a higher score value corresponds to lower risk.

[0043] Scoring sources with a second scoring direction can be stored in a preset scoring source set, denoted as HIGHER_IS_BETTER. This set represents the set of scoring sources where higher scores indicate better evaluation results. For example, scoring source 2019 can be configured in the HIGHER_IS_BETTER set; scoring sources such as scoring source 2011 and scoring source 2007, which are not configured in this set, can be processed according to the first scoring direction by default. The scoring direction configuration can be preset by business personnel or domain experts based on the business meaning of the scoring source, scoring rules, or evaluation logic, or it can be automatically determined based on historical sample data by analyzing the correlation between the scoring value and known risk labels, quality labels, or manual review results.

[0044] After obtaining the global quantile values ​​corresponding to each scoring source, the quantile values ​​are oriented and unified according to the scoring direction configuration to generate risk quantile values. These risk quantile values ​​are used to uniformly represent the risk level of a sample; their semantics are: the larger the value, the higher the risk. The conversion formula for oriented unification is as follows: ,in, This represents the hazard quantile of sample i in scoring source c. HIGHER_IS_BETTER represents the global quantile value of sample i under rating source c, and HIGHER_IS_BETTER represents the set of rating sources where a higher rating value indicates a better evaluation result.

[0045] As shown in the formula above, for rating sources where higher scores indicate better evaluation results, since high scores correspond to low risk, their global quantiles are reverse-transformed to obtain (1-p_{i,c}). For other rating sources, since higher scores correspond to higher risk, the global quantiles are directly used as risk quantiles. Through this processing, the rating results from different rating sources are converted into risk quantiles with a consistent direction, thus ensuring that all rating sources satisfy the unified semantics of "the larger the risk quantile, the higher the risk," facilitating subsequent cross-rating source comparisons, risk aggregation, and anomaly sample identification.

[0046] For each rating source, calculate the quantile rank of each rating across all its samples. The quantile rank represents the relative position of the rating within all samples of the rating source, and its value ranges from 0 to 1.

[0047] In one implementation, the standardization process described above includes quantile-based rating transformation. Specifically, for any rating source... c Retrieve all rating values ​​corresponding to the rating source and sort them in ascending order according to the rating values. For any rating value, calculate the proportion of samples with a value less than or equal to that rating value out of the total number of samples in the rating source, and use this proportion as the global quantile value corresponding to that rating value.

[0048] Furthermore, when multiple samples have the same score, the ranking of that score is determined by an average ranking method. That is, for multiple samples corresponding to the same score, their rankings in the sorting sequence are averaged, and a quantile value is calculated based on this average ranking. The calculation method for the quantile value can be expressed as follows: ,in, Indicates the source of the rating c The quantile value of the i-th sample. This indicates that the sample rating value is in the rating source. c Rank among all samples Indicates the source of the rating c The total number of samples.

[0049] After completing the global quantile calculation, the quantile directions of different scoring sources are uniformly processed according to the preset scoring direction configuration to generate risk quantile values. These risk quantile values ​​characterize the risk level of a sample; semantically, a larger value indicates a higher risk. Specifically, when a higher score from a scoring source indicates a higher risk, its global quantile value is directly used as the risk quantile value; conversely, when a lower score from a scoring source indicates a higher risk, its global quantile value is reversed to obtain risk quantile values ​​with consistent directions.

[0050] Through the above processing back The scoring results from different scoring sources are converted to a unified quantile scale, and the risk direction remains consistent, which facilitates subsequent cross-scoring source comparisons, risk aggregation calculations, and anomaly sample identification.

[0051] Step 103: For each risk assessment scenario, calculate the hybrid weight based on the features of the scenario; and fuse the global risk quantile with the local quantile within the scenario according to the hybrid weight to generate a hybrid label for each scoring source in the scenario, wherein the scenario features include at least the number of valid scoring sources and the dispersion of the scoring distribution within the scenario.

[0052] In this step, while global quantile normalization eliminates scale differences between rating sources, it also introduces a problem: the score distributions of different rating sources may differ, and some rating sources may be generally overestimated or underestimated. Using only global quantiles may lead to systematic overestimation or underestimation of scores in certain scenarios. On the other hand, local quantile methods calculate rankings within the scene, which can eliminate systematic biases between rating sources, but suffer from instability with small samples. To balance global consistency and local discriminative power, an adaptive local-global hybrid strategy is designed. This strategy automatically adjusts the mixing ratio of global and local quantiles based on scene characteristics (number of rating sources, score distribution dispersion).

[0053] As an optional implementation method in this embodiment, for each risk assessment scenario, the mixed weight calculation based on the characteristics of the scenario includes: calculating the sample size factor based on the number of valid scoring sources in the scenario, calculating the dispersion factor based on the interquartile range of the global risk quantiles in the scenario, and determining the corresponding mixed weight for the scenario based on the sample size factor and the dispersion factor.

[0054] In this optional implementation, the blending weights determine the proportion of global and local quantiles in the blended labels. The blending weights are obtained by multiplying two factors: a sample size factor and a dispersion factor. (See reference) Figure 2 The process of generating mixed tags is illustrated in the diagram.

[0055] The sample size factor reflects the impact of the number of rating sources within the scene on the mixed weights. When the number of rating sources is small, local quantiles are unstable, and more reliance should be placed on global quantiles; when the number of rating sources is large, local quantiles are more reliable, and more reliance can be placed on local quantiles.

[0056] The formula for calculating the sample size factor is: the sample size factor equals (number of rating sources in the scene minus one) divided by (number of rating sources in the scene minus one plus the sample size sensitivity coefficient), where the sample size sensitivity coefficient is 8.0 by default.

[0057] Characteristics of the sample size factor: When the number of rating sources is one, the sample size factor is zero, completely dependent on the global quantile; when the number of rating sources approaches infinity, the sample size factor approaches one, completely dependent on the local quantile; the sample size sensitivity coefficient controls the transition speed from global to local, the larger the value, the slower the transition. The dispersion factor reflects the impact of the dispersion of the rating distribution within a scene on the mixed weights. When the rating distribution is concentrated (low dispersion), it indicates that the differences between rating sources are small, and the additional information provided by local quantiles is limited; therefore, more reliance should be placed on the global quantile. When the rating distribution is dispersed (high dispersion), it indicates that there are differences between rating sources, and local quantiles can provide more discriminative information. Dispersion is measured using the interquartile range (IQR), which is the 75th quantile minus the 35th quantile.

[0058] The formula for calculating the dispersion factor is: dispersion factor equals interquartile range / (interquartile range plus dispersion sensitivity coefficient), where the dispersion sensitivity coefficient is 20.0 by default.

[0059] Characteristics of the dispersion factor: When the interquartile range is zero (scores are perfectly concentrated), the dispersion factor is zero, entirely dependent on the global quantile. As the interquartile range approaches infinity, the dispersion factor approaches one, entirely dependent on the local quantile. The dispersion sensitivity coefficient controls the transition speed from global to local.

[0060] The mixture weight equals the sample size factor multiplied by the dispersion factor. The mixture weight ranges from 0 to 1, with larger values ​​indicating a greater reliance on local quantiles and smaller values ​​indicating a greater reliance on global quantiles.

[0061] It should be understood that a simplified version of the mixed weight calculation can be used, considering only the sample size factor and ignoring the dispersion factor. In this case, the mixed weight is determined solely by the number of rating sources. The advantage of this approach is lower computational cost, making it suitable for scenarios with high real-time requirements. The disadvantage is that it cannot adjust the weights based on the rating distribution characteristics, and may not be the optimal solution in some scenarios.

[0062] As an optional implementation of this embodiment, the process of generating a mixed label for each scoring source in the scenario by fusing the global risk quantile with the local quantile in the scenario according to the mixed weight includes: ranking all risk quantiles in each risk assessment scenario in ascending order, and using a smooth quantile strategy to calculate the local quantile corresponding to each risk quantile; let the mixed label = (1 - mixed weight) × global risk quantile + mixed weight × local quantile.

[0063] For each scenario, a smoothing quantile transformation is performed to obtain local quantiles. First, a smoothing quantile function is defined. The traditional formula for calculating quantiles is: quantile equals rank divided by the number of samples. This calculation method is biased when the sample size is small.

[0064] As an optional implementation method in this embodiment, the smoothing quantile strategy is: smoothing quantile = (rank - a) / (number of samples + 1 - 2a), where a is the smoothing parameter and the number of samples is the number of valid rating sources in the scenario; when the number of samples ≤ 1, the local quantile is directly taken as 0.5.

[0065] An improved smoothing quantile formula is adopted, introducing a smoothing parameter 'a': the smoothing quantile equals (rank minus smoothing parameter 'a') divided by (sample number plus one minus twice the smoothing parameter 'a'), where the rank is the ascending ranking of the risk quantile within the scenario; the sample number is the number of valid rating sources within the scenario; the smoothing parameter 'a' defaults to 0.375 (corresponding to the quantile estimate of a normal distribution). When the sample number is less than or equal to one, 0.5 is directly returned as the local quantile; when the denominator is less than or equal to zero, it degenerates into the traditional formula: (rank minus one) divided by (sample number minus one).

[0066] Using a fixed smoothing parameter (e.g., 0.375) without dynamic adjustment based on sample size has the advantage of being simpler to implement and suitable for scenarios with relatively uniform sample size distribution. However, it may not be optimal in scenarios with significant sample size variations. In other words, a fixed smoothing parameter may not be the best choice when there are large differences in sample size. Small samples require a larger smoothing parameter to reduce noise, while large samples can use a smaller smoothing parameter to improve accuracy. An adaptive smoothing parameter function is provided, dynamically adjusting the smoothing parameter based on the sample size within the scenario. The adaptive smoothing parameter is calculated as follows: the adaptive smoothing parameter is equal to the base smoothing parameter multiplied by (one plus the value of the exponential function), where the parameter of the exponential function is the negative decay coefficient multiplied by (the number of samples minus the reference sample size) and then divided by the reference sample size. The default base smoothing parameter is 0.375, the default reference sample size is 50, and the default decay coefficient is 0.1.

[0067] The characteristics of the adaptive smoothing parameter are as follows: When the sample size is much smaller than the reference sample size, the exponential function value is close to one, and the adaptive smoothing parameter is close to twice the base smoothing parameter (maximum 0.5); when the sample size is equal to the reference sample size, the adaptive smoothing parameter is equal to the base smoothing parameter; when the sample size is much larger than the reference sample size, the exponential function value is close to zero, and the adaptive smoothing parameter is close to the base smoothing parameter. The adaptive smoothing parameter is limited to the range of 0-0.5.

[0068] Based on the mixed weight, the global risk quantile and local quantile are combined to obtain the mixed label. The formula for calculating the mixed label is: the mixed label equals (one minus the mixed weight) multiplied by the global risk quantile plus the mixed weight multiplied by the local quantile.

[0069] Characteristics of the hybrid label: When the hybrid weight is zero, the hybrid label is equal to the global risk quantile; when the hybrid weight is one, the hybrid label is equal to the local quantile; when the hybrid weight is between zero and one, the hybrid label is a weighted average of the global risk quantile and the local quantile; perform the above steps for each scoring dimension (business score, traffic score, comprehensive score) to obtain the hybrid label for the three dimensions.

[0070] As an optional implementation of this embodiment, the method further includes calculating the credibility weight of each rating source: constructing a scenario and rating source rating matrix, wherein each row of the matrix corresponds to a risk assessment scenario, and each column corresponds to a mixed label of a rating source; for each pair of rating sources in the scenario and rating source rating matrix, calculating their Spearman rank correlation coefficient on the common scenario, so as to determine the pairwise correlation through the Spearman rank correlation coefficient, wherein the Spearman rank correlation coefficient is used to measure the monotonic relationship between two variables, and the larger the value, the stronger the positive correlation; and determining the credibility weight of each rating source based on the pairwise correlation.

[0071] In this optional implementation, to improve the quality of consensus labels, this step calculates the credibility weight of each scoring source based on pairwise consistency among the scoring sources. Scoring sources with higher credibility weights will receive greater weight during aggregation, while scoring sources with lower credibility weights will have their weights reduced.

[0072] The specific implementation includes the following steps: Constructing a Scene-Rating Source Rating Matrix: Convert the data from a long table format to a wide table format and construct a scene-rating source rating matrix. Each row of the matrix corresponds to a scene, and each column corresponds to a mixed label for a rating source. This can be achieved using pivot tables: the index is the scene identifier field, the column is the rating source identifier, the value is the mixed label, and the aggregation function is the mean (for processing multiple records of the same scene and the same rating source).

[0073] Calculate the pairwise Spearman correlation between rating sources: For each pair of rating sources, calculate their Spearman rank correlation coefficient in the common scenario. The Spearman correlation coefficient measures the monotonic relationship between two variables, ranging from negative one to positive one, with a larger value indicating a stronger positive correlation.

[0074] The specific calculation method is as follows: For rating sources c1 and c2, extract these two columns from the scene-rating source rating matrix; delete rows with missing values ​​in any column to obtain paired valid data; calculate the number of rows with valid data as the number of overlapping samples; if the number of overlapping samples is greater than or equal to the minimum overlap threshold (default is 5,000), then calculate the Spearman correlation coefficient. For example, the Spearman correlation coefficient is calculated by ranking the values ​​of the two variables separately; then calculating the Pearson correlation coefficient for each ranking.

[0075] Calculate the weights of rating sources: Calculate the credibility weight for each rating source based on pairwise correlations. First, for each pairwise correlation, calculate its contribution to the weight: if the correlation is negative, the contribution is zero (negative correlation indicates that the rating sources are in opposite directions and are unreliable); if the correlation is positive, the contribution is the square of the correlation (the square operation emphasizes high correlation and suppresses low correlation). Then, for each rating source, aggregate all pairwise weight contributions associated with it: collect the weight contributions of this rating source with all other rating sources; calculate the mean of these weight contributions as the original weight for this rating source.

[0076] Finally, the original weights are normalized: the mean of all original weights is calculated; the normalized weight for each rating source is equal to its original weight divided by the mean. The mean of the normalized weights is one, which is easy to understand and compare.

[0077] Save weight information: Save the pairwise correlations between scoring sources and the confidence weights of scoring sources as files for subsequent analysis and verification.

[0078] Immediately, other relevance measures can be used instead of the Spearman correlation coefficient when calculating the confidence weights of rating sources. For example, the Pearson correlation coefficient is suitable for linear relationships, the Kendall rank correlation coefficient is more robust to outliers, and mutual information can capture non-linear relationships.

[0079] As an optional implementation of this embodiment, the step of aggregating the mixed labels of all rating sources in the scenario to generate the consensus label corresponding to the scenario includes: taking the median of the mixed labels of all rating sources in each scenario as the consensus label of the scenario; or, taking a weighted average of the mixed labels of all rating sources in each scenario according to the credibility weight of the corresponding rating source as the consensus label of the scenario, wherein the weighted average formula is: consensus label = Σ(credibility weight × mixed label) / Σcredibility weight.

[0080] In this optional implementation, for each scenario, the mixed labels from all rating sources are aggregated to generate the final consensus label. The consensus label is the comprehensive evaluation result of all rating sources in that scenario and will be used as the training target for the machine learning model.

[0081] Specifically, the implementation includes the following two methods: Method 1: Simple Median Aggregation (Default Method): When rating source credibility weights are not enabled, median aggregation is used. The median has good robustness and is not affected by extreme values. Specifically, it is implemented by grouping by the scenario identifier field and taking the median of the mixed labels for each scenario. The aggregation formula is: the consensus label equals the median of the mixed labels from all rating sources within the scenario.

[0082] The advantages of median aggregation are its robustness to outliers; extreme ratings from individual rating sources do not significantly affect the results. It is also computationally simple and efficient, requiring no additional parameters. It should be understood that mean aggregation can be used instead of median aggregation to generate consensus labels. Mean aggregation has the advantage of lower computational complexity, but its disadvantage is greater sensitivity to outliers. It is suitable for scenarios where the quality of rating sources is relatively uniform.

[0083] It should be understood that a truncated mean can also be used instead of median aggregation. The truncated mean removes a certain percentage (e.g., 10%) of the highest and lowest ratings before calculating the mean. This approach balances the robustness of the median with the stability of the mean, and is suitable for scenarios where the rating distribution is nearly symmetrical.

[0084] Method two is weighted aggregation (when rating source weights are enabled). When rating source credibility weights are enabled, weighted aggregation is used. Rating sources with higher credibility receive greater weight during aggregation. Specifically, for each scenario, the mixed labels of all rating sources within that scenario are obtained; based on the rating source identifier, the corresponding credibility weight is retrieved from the weight table; rating sources with zero or missing weights are filtered out; the weighted label is calculated as the mixed label multiplied by the credibility weight; the weighted labels are summed, and the weights are summed; the consensus label is equal to the sum of weighted labels divided by the sum of weights.

[0085] The aggregation formula is: the consensus label equals the sum of the credibility weights of all rating sources multiplied by the mixed label, divided by the sum of the credibility weights of all rating sources.

[0086] The advantage of weighted aggregation is that the contribution of high-quality rating sources is amplified, while the contribution of low-quality rating sources is suppressed; the weights are automatically adjusted based on the consistency of the rating sources without manual intervention.

[0087] It should be understood that an iterative weighted aggregation method can be used. First, aggregation is performed using equal weights. Then, the weights are updated based on the consistency between each scoring source and the initial consensus label. The new weights are then used to re-aggregate, and this process is repeated iteratively until convergence. This approach yields more accurate weights, but requires more computation.

[0088] Finally, an aggregation operation is performed on each scoring dimension (business score, traffic score, and comprehensive score) to obtain consensus labels for the three dimensions: business dimension consensus label: reflects business risk; traffic dimension consensus label: reflects traffic risk; and comprehensive dimension consensus label: reflects comprehensive risk.

[0089] The generated consensus labels are saved as files in a specified format for use in subsequent machine learning model training.

[0090] For example, the chief analyst merges multi-dimensional consensus tags: Consensus tags from business, traffic, and comprehensive dimensions are merged according to scenario identifier fields to form a complete tag table. The tag file is then saved: the tag table is saved as a compressed CSV file (e.g., gzip compression) to save storage space and transmission bandwidth. Output fields include scenario identifier fields: vehicle identification code, date, vehicle... Loss amount The consensus labels for various aspects include: third-party liability amount, driver compensation amount, and passenger compensation amount; business dimension consensus label: our_biz_weighted; traffic dimension consensus label: our_traffic; and comprehensive dimension consensus label: our_total. Finally, intermediate results are saved: If the calculation of rating source credibility weights is enabled, the following intermediate results can also be saved: Rating source pairwise correlation table: records the number of overlapping samples and Spearman correlation coefficient for each pair of rating sources; Rating source credibility weight table: records the credibility weight of each rating source. These intermediate results can be used to verify and analyze the fusion process.

[0091] The innovative aspects of this embodiment include the following: Global quantile normalization and direction unification techniques: Quantile normalization is performed on all historical samples of each rating source, and a unified rating direction is configured according to the rating semantics. The technical problem addressed is that different rating sources use different rating directions (some prefer higher scores, others prefer lower scores) and rating scales (some from 0 to 100, others from 0 to 1000), which can lead to score cancellation or scale inconsistencies if directly merged. Technical methods include: calculating the quantile ranking of each rating in all samples of each rating source, unifying ratings from different scales to the 0-1 range; pre-configuring the rating direction (LOWER_IS_BETTER or HIGHER_IS_BETTER) for each rating source; for rating sources where "higher scores are better," inverting the quantile (1 minus the quantile) to unify it to a direction where "higher values ​​indicate higher risk." The corresponding advantages are: eliminating differences in rating direction and scale, making ratings from different sources comparable and fusionable, laying the foundation for subsequent mixing and aggregation operations.

[0092] Local-Global Hybrid Label Calculation Techniques: An adaptive local-global hybrid strategy is designed to achieve a balance between global consistency and local discriminative power. The technical problem addressed is that pure local quantile methods are unstable in small-sample scenarios, while pure global quantile methods lack local discriminative power, failing to simultaneously meet the requirements of stability and discriminative power. The techniques include: calculating local quantiles within each scenario using an improved smoothing quantile formula; calculating hybrid weights based on scenario features (number of rating sources, rating distribution dispersion); and fusing global and local quantiles according to the hybrid weights: the hybrid label equals (one minus the weight) multiplied by the global quantile plus the weight multiplied by the local quantile. The corresponding advantages are: maintaining stability in small-sample scenarios (more reliant on global quantiles), maintaining discriminative power in large-sample and discretely distributed scenarios (more reliant on local quantiles), and automatically adjusting the fusion strategy according to scenario features.

[0093] Adaptive Hybrid Weight Calculation Techniques: An adaptive hybrid weight calculation method based on scene features is designed. The technical problem addressed is that fixed weights cannot adapt to the characteristics of different scenes; in some scenes, the global quantile is better, while in others, the local quantile is better, requiring automatic adjustment based on scene features. Technical methods include: Sample size factor: equal to (number of rating sources minus one) divided by (number of rating sources minus one plus sample size sensitivity coefficient). Dispersion factor: equal to the interquartile range divided by (interquartile range plus dispersion sensitivity coefficient). Hybrid weight equals the sample size factor multiplied by the dispersion factor. Corresponding advantages: Automatically adjusts the fusion strategy according to the number of rating sources and the dispersion of the rating distribution within the scene, without requiring manual setting of scene-specific parameters, applicable to scenes with various features.

[0094] Adaptive smoothing parameter technique: A method to dynamically adjust smoothing parameters based on the sample size within a scenario. The technical problem it solves is that a fixed smoothing parameter is not optimal for scenarios with large variations in sample size; a larger smoothing parameter is needed for small samples to reduce noise, while a smaller smoothing parameter can be used for large samples to improve accuracy. Techniques include: using a sigmoid curve to achieve adaptive adjustment of the smoothing parameter; the adaptive smoothing parameter being equal to the base smoothing parameter multiplied by (one plus an exponential function value); and the exponential function being adjusted based on the difference between the sample size and the reference sample size. The corresponding advantages are: using a larger smoothing parameter for small samples improves stability, while using a smaller smoothing parameter for large samples improves accuracy, automatically adapting to scenarios with different sample sizes.

[0095] The key technical point of the pairwise correlation-based rating source credibility weighting method is to calculate the credibility weight of each rating source based on the pairwise consistency among rating sources. The technical problem it addresses is that different rating sources have varying data quality and algorithm levels; simple aggregation can be affected by low-quality rating sources, leading to a decrease in the quality of consensus labels. The technical methods include: calculating the Spearman rank correlation coefficient for each pair of rating sources in a common scenario; using the square of positive correlation as the weight contribution, and zero contribution for negative correlation; aggregating all weight contributions for each rating source to obtain the credibility weight; and using the credibility weight for weighted aggregation. The corresponding advantages are: rating sources consistent with the majority of rating sources receive higher weights, automatically identifying and reducing the influence of abnormal rating sources, and improving the overall quality of consensus labels.

[0096] Multi-dimensional consensus label generation technology: Consensus labels are generated for multiple scoring dimensions. The technical problem addressed is that different business scenarios focus on different dimensions of risk, and single-dimensional labels cannot meet diverse business needs. Technical methods include: performing steps such as unifying direction, normalizing, mixing, and aggregating the commercial score, compulsory risk score, and comprehensive score respectively. This generates independent consensus labels for the three dimensions. Corresponding advantages: Meets the needs of different business scenarios, provides a more comprehensive risk assessment, and allows for flexible selection of which dimension's label to use.

[0097] Compared with existing technologies, it has the following beneficial effects: Direction Independence: Through direction unification processing, this invention supports the fusion of rating sources with different semantics ("the higher the score, the better" or "the lower the score, the better") without the need for manual adjustment of the rating direction. The system can automatically unify the direction simply by pre-configuring the rating direction of each rating source, greatly reducing the complexity of multi-source rating fusion.

[0098] Scale independence: Through global quantile normalization, this invention eliminates the differences between different rating scales (such as 0 to 100 points, 0 to 1000 meters, 0 to 1.0), making ratings from different sources comparable. Regardless of the scale used in the original rating, after normalization, they are all unified to a risk quantile of 0 to 1.

[0099] Stability: The local-global hybrid strategy remains stable in small sample scenarios. When only a few scoring sources provide data for a certain scenario, the hybrid weights approach zero, and the hybrid labels are mainly determined by the global quantiles, avoiding the instability of the local quantile method in small sample scenarios.

[0100] Adaptability: The system automatically adjusts the blending weights based on the number and distribution characteristics of rating sources within the scene, eliminating the need for manual setting of scene-specific parameters. In scenes with different characteristics, the system can automatically select the optimal fusion strategy.

[0101] Robustness: By using median aggregation and confidence weighting, this invention reduces the impact of anomalous rating sources. Extreme ratings or systematic biases from individual rating sources do not significantly affect the quality of the consensus labels.

[0102] Interpretability: The credibility weights of the rating sources reflect the degree of consistency between each rating source and the group, making it easier for business personnel to understand and verify the fusion process. Business personnel can see which rating sources received higher weights and which received lower weights, thereby building trust in the fusion results.

[0103] Scalability: This invention supports dynamically adding or removing scoring sources. Adding a scoring source only requires configuring its direction information to participate in fusion; removing a scoring source does not require modifying the system, only that the scoring source's score is no longer included in the data.

[0104] Efficiency: By pre-computing global quantiles and batch processing, this invention supports efficient processing of large-scale data. For datasets containing millions or even tens of millions of records, this invention can complete the generation of consensus labels within a reasonable timeframe.

[0105] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0106] According to an embodiment of the present invention, a consensus label generation device for multi-source heterogeneous risk scoring is also provided, comprising: a data acquisition unit for acquiring multi-source heterogeneous scoring data and performing preprocessing to obtain effective scoring data, wherein the multi-source heterogeneous scoring data consists of risk scoring data from multiple scoring sources for the same batch of target objects; a data processing unit for performing global quantile normalization and scoring direction unification processing on each scoring source based on the effective scoring data to obtain global risk quantiles with unified scale and unified risk direction for each scoring source; wherein the global quantile normalization is used to eliminate differences in scoring scales among different scoring sources, and the scoring direction unification is used to unify the risk semantics of all scoring sources to the case that the larger the value, the higher the risk; for each risk assessment scenario, a hybrid weight is calculated based on the features of the scenario; and the global risk quantile is fused with the local quantiles within the scenario according to the hybrid weight to generate a hybrid label for each scoring source in the scenario, wherein the scenario features include at least the number of effective scoring sources and the dispersion of the scoring distribution within the scenario; and a consensus label generation unit for aggregating the hybrid labels of all scoring sources within the scenario for each risk assessment scenario to generate a consensus label corresponding to the scenario.

[0107] According to embodiments of the present invention, the present invention also provides an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to implement the methods described in any of the above embodiments.

[0108] According to embodiments of the present invention, the present invention also provides a readable storage medium storing computer instructions that enable a computer to perform the methods described in any of the above embodiments when executed.

[0109] According to embodiments of the present invention, the present invention also provides a computer program product that, when executed by a processor, can implement the methods described in any of the above embodiments.

[0110] Figure 3 A schematic block diagram of an example electronic device 300 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices.

[0111] like Figure 3 As shown, the electronic device 300 includes a computing unit 301, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 302 or a computer program loaded from a storage unit 308 into a random access memory (RAM) 303. The RAM 303 may also store various programs and data required for the operation of the electronic device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0112] Multiple components in electronic device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of displays, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows electronic device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0113] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as the object matching method. For example, in some embodiments, the object matching method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by the computing unit 301, one or more steps of the methods described above may be performed.

[0114] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0115] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0116] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

Claims

1. A consensus label generation method for multi-source heterogeneous risk scoring, characterized in that, include: Acquire multi-source heterogeneous scoring data and perform preprocessing to obtain effective scoring data; Based on the effective scoring data, global quantile normalization and scoring direction unification are performed on each scoring source to obtain global risk quantiles with a unified scale and risk direction for each scoring source; wherein, the multi-source heterogeneous scoring data refers to the risk scoring data of multiple scoring sources for the same batch of target objects; the global quantile normalization is used to eliminate the differences in scoring scales between different scoring sources, and the scoring direction unification is used to unify the risk semantics of all scoring sources so that the larger the value, the higher the risk. For each risk assessment scenario, a hybrid weight is calculated based on the characteristics of that scenario; and the global risk quantile is fused with the local quantile within that scenario according to the hybrid weight to generate a hybrid label for each scoring source in that scenario, wherein the scenario characteristics include at least the number of valid scoring sources and the dispersion of the scoring distribution within that scenario; For each risk assessment scenario, the hybrid labels from all scoring sources within that scenario are aggregated to generate a consensus label corresponding to that scenario.

2. The consensus label generation method for multi-source heterogeneous risk scoring according to claim 1, characterized in that, The process of performing global quantile normalization and direction unification for each scoring source includes: Pre-configure the scoring direction for each scoring source; For each rating source, calculate the global quantile ranking of each rating value in all its historical rating samples, wherein the global quantile ranking is the proportion of samples less than or equal to that rating value to all samples of that rating source; Risk quantiles are determined based on the scoring direction and global quantile ranking.

3. The consensus label generation method for multi-source heterogeneous risk scoring according to claim 1, characterized in that, For each risk assessment scenario, the hybrid weights are calculated based on the characteristics of that scenario, including: Calculate the sample size factor based on the number of valid scoring sources in the scenario; The dispersion factor is calculated based on the interquartile range of the global risk quantile within this scenario. The mixed weights for this scenario are determined based on the sample size factor and the dispersion factor.

4. The consensus label generation method for multi-source heterogeneous risk scoring according to claim 3, characterized in that, Based on the hybrid weighting, the global risk quantile is fused with the local quantile within the scenario to generate a hybrid label for each scoring source in that scenario, including: All risk quantiles within each risk assessment scenario are ranked in ascending order, and a smoothing quantile strategy is used to calculate the local quantile corresponding to each risk quantile. Let the mixed label = (1 - mixed weight) × global risk quantile + mixed weight × local quantile.

5. The consensus label generation method for multi-source heterogeneous risk scoring according to claim 4, characterized in that, The smoothing quantile strategy is as follows: smoothing quantile = (rank - a) / (number of samples + 1 - 2a), where a is the smoothing parameter and the number of samples is the number of valid rating sources in the scenario; when the number of samples ≤ 1, the local quantile is directly taken as 0.5; The smoothing parameter a is adjusted adaptively, where the adaptive smoothing parameter a = a0 × (1 + e^(-k × (n-n0) / n0)), where a0 is the basic smoothing parameter; n0 is the reference sample size; k is the attenuation coefficient; and n is the number of effective scoring sources in the scenario.

6. The consensus label generation method for multi-source heterogeneous risk scoring according to claim 4, characterized in that, After generating the mixed labels, and before aggregating the mixed labels from all rating sources within the scenario, the method further includes calculating the credibility weight of each rating source: Construct a scenario and scoring source scoring matrix, wherein each row of the matrix corresponds to a risk assessment scenario and each column corresponds to a mixed label of a scoring source; For each pair of rating sources in the rating matrix of scene and rating source, calculate their Spearman rank correlation coefficient in the common scene, so as to determine the pairwise correlation through the Spearman rank correlation coefficient. The Spearman rank correlation coefficient is used to measure the monotonic relationship between two variables. The larger the value, the stronger the positive correlation. Based on pairwise correlation, the credibility weight of each rating source is determined.

7. The consensus label generation method for multi-source heterogeneous risk scoring according to claim 6, characterized in that, The aggregation of the mixed tags from all rating sources within the scenario to generate the consensus tag corresponding to the scenario includes: The median of the mixed labels from all rating sources within each scenario is taken as the consensus label for that scenario. Alternatively, for the mixed labels of all rating sources within each scenario, a weighted average is calculated based on the credibility weight of the corresponding rating source, which serves as the consensus label for that scenario. The weighted average formula is: Consensus label = Σ(credibility weight × mixed label) / Σcredibility weight.

8. The consensus label generation method for multi-source heterogeneous risk scoring according to claim 1, characterized in that, The process of acquiring and preprocessing multi-source scoring data includes: Based on predefined fields for uniquely identifying a risk assessment scenario and a scoring field, the corresponding multi-source scoring data is loaded from the original data source. The system performs data type conversion, invalid value handling, deduplication, and masking on the data corresponding to each field to obtain valid data. In the masking process, the scores of a specific scoring source in a specific dimension are set as missing values ​​according to the preset mask configuration.

9. A consensus label generation device for multi-source heterogeneous risk scoring, characterized in that, include: The data acquisition unit is used to acquire multi-source heterogeneous scoring data and perform preprocessing to obtain effective scoring data, wherein the multi-source heterogeneous scoring data is risk scoring data of multiple scoring sources for the same batch of target objects; The data processing unit, based on the valid scoring data, performs global quantile normalization and scoring direction unification on each scoring source to obtain global risk quantiles with a unified scale and risk direction for each scoring source. The global quantile normalization eliminates differences in scoring scales among different scoring sources, and the unified scoring direction unifies the risk semantics of all scoring sources so that a larger value indicates higher risk. For each risk assessment scenario, a hybrid weight is calculated based on the features of that scenario. The global risk quantile is then fused with the local quantiles within that scenario according to the hybrid weight to generate a hybrid label for each scoring source in that scenario. The scenario features include at least the number of valid scoring sources and the dispersion of the scoring distribution within that scenario. The consensus label generation unit is used to aggregate the mixed labels from all scoring sources within each risk assessment scenario to generate the consensus label corresponding to that scenario.

10. An electronic device, comprising: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to cause the at least one processor to perform the method described in any implementation of the first aspect.

Images