Multi-modal-based unhealthy asset risk assessment method and device, equipment and medium

Abstract

The invention relates to the technical field of risk assessment, in particular to a multi-modal-based unhealthy asset risk assessment method and device, equipment and a medium, and the method comprises the steps: obtaining fusion feature data based on multi-modal data of unhealthy assets; inputting the fused feature data into a trained first risk assessment model to obtain a preliminary risk assessment value and a first assessment confidence coefficient; determining an actual confidence coefficient based on the first assessment confidence coefficient, and calibrating the preliminary risk assessment value according to the actual confidence coefficient to obtain a calibrated risk assessment value; optimizing the first risk assessment model based on the calibration risk assessment value and the preliminary risk assessment value to obtain a second risk assessment model; and performing evaluation based on the fused feature data through a second risk evaluation model to obtain a target risk evaluation value and a second evaluation confidence coefficient, the second evaluation confidence coefficient satisfying a preset confidence coefficient condition. According to the method, the limitation of a single information source can be overcome, adaptive optimization of the evaluation model can be realized, and the misjudgment risk is effectively reduced.

Description

Technical Field

[0001] This application relates to the field of risk assessment technology, specifically to a method, apparatus, equipment, and medium for assessing non-performing asset risk based on multimodality. Background Technology

[0002] Accurate risk assessment of non-performing assets is the core decision-making basis for subsequent asset disposal, risk pricing, and capital allocation. Currently, the industry widely adopts machine learning methods for non-performing asset risk assessment. However, existing methods for assessing the risk of non-performing assets have two major flaws: First, the assessment process often relies on modeling and analysis of a single type of data, resulting in incomplete assessment dimensions and difficulty in fully and accurately reflecting the actual risk status of non-performing assets, leading to low assessment accuracy; Second, the risk assessment results output by the models do not quantify their own reliability, lacking support for decision-making credibility, which in turn leads to low credibility of the assessment results. Summary of the Invention

[0003] To overcome the shortcomings of the prior art, this application provides a multimodal risk assessment method, apparatus, equipment and medium for non-performing assets, aiming to solve the problems of low assessment accuracy due to reliance on single data and incomplete dimensions in existing risk assessment methods, as well as the lack of reliable quantitative indicators and insufficient credibility of assessment results.

[0004] The first aspect of this application provides a multimodal approach to assessing the risk of non-performing assets, the approach comprising: Fusion feature data is obtained based on multimodal data of non-performing assets; The fused feature data is input into the trained first risk assessment model to obtain a preliminary risk assessment value and a first assessment confidence level. The actual confidence level is determined based on the first assessment confidence level, and the preliminary risk assessment value is calibrated according to the actual confidence level to obtain the calibrated risk assessment value; The first risk assessment model is optimized based on the calibration risk assessment value and the preliminary risk assessment value to obtain the second risk assessment model; The second risk assessment model evaluates the risk based on the fused feature data to obtain a target risk assessment value and a second assessment confidence level, wherein the second assessment confidence level meets the preset confidence level conditions.

[0005] Optionally, calibrating the preliminary risk assessment value based on the actual confidence level to obtain a calibrated risk assessment value includes: The actual confidence level is compared with a preset first confidence threshold and a preset second confidence threshold, wherein the first confidence threshold is greater than the second confidence threshold; If the actual confidence level is greater than the first confidence level threshold, then the preliminary risk assessment value will be output as the calibration risk assessment value. If the actual confidence level is less than the first confidence threshold but greater than the second confidence threshold, the preliminary risk assessment value is calibrated based on a preset standard risk benchmark value to obtain a calibrated risk assessment value.

[0006] Optionally, comparing the actual confidence level with the first confidence threshold and the second confidence threshold includes: Determine the industry risk level and collateral type of non-performing assets; Obtain the first adjustment coefficient corresponding to the industry risk level; Obtain the second adjustment coefficient corresponding to the type of collateral; The first confidence threshold and the second confidence threshold are adjusted based on the first adjustment coefficient and the second adjustment coefficient to obtain a first target confidence threshold and a second target confidence threshold, wherein the first target confidence threshold is greater than the second target confidence threshold; The actual confidence level is compared with the first target confidence threshold and the second target confidence threshold.

[0007] Optionally, calibrating the preliminary risk assessment value based on a preset standard risk benchmark value to obtain a calibrated risk assessment value includes: Based on the actual confidence level, the first confidence threshold, and the second confidence threshold, the confidence matching weight is obtained; The preliminary risk assessment value is calibrated based on the confidence matching weight and the preset standard risk benchmark value to obtain the calibrated risk assessment value.

[0008] Optionally, optimizing the first risk assessment model based on the calibration risk assessment value and the preliminary risk assessment value to obtain the second risk assessment model includes: Calculate the difference between the calibration risk assessment value and the preliminary risk assessment value; The confidence rate of change is obtained based on the confidence matching weight and the actual confidence level. A calibration residual loss function is constructed based on the difference, the confidence rate of change, and the preset confidence sensitivity regularization coefficient; The first risk assessment model is iteratively optimized using the calibration residual loss function; At the end of the iterative optimization, the second risk assessment model is obtained.

[0009] Optionally, the fused feature data obtained from the multimodal data of non-performing assets includes: Acquire structured financial data, unstructured legal text data, collateral image data, and transaction flow data from the multimodal data; A structured feature vector is generated based on the structured financial data; Generate text feature vectors based on the aforementioned unstructured legal text data; Generate image feature vectors based on the image data of the pledged collateral; Generate a time-series feature vector based on the transaction flow data; The structured feature vector, text feature vector, image feature vector, and time-series feature vector are weighted and fused using a multi-head attention mechanism to obtain the fused feature data.

[0010] Optionally, determining the actual confidence level based on the first assessed confidence level includes: The integrity, consistency, and timeliness of the multimodal data are obtained; A data quality score is obtained based on the completeness, consistency, and timeliness mentioned above. The actual confidence level is obtained based on the first assessment confidence level and the data quality score.

[0011] A second aspect of this application provides a multimodal non-performing asset risk assessment device, the device comprising: The feature generation module is used to obtain fused feature data based on multimodal data of non-performing assets; The first evaluation module is used to input the fused feature data into the trained first risk assessment model to obtain a preliminary risk assessment value and a first assessment confidence level. The risk calibration module is used to determine the actual confidence level based on the first assessment confidence level, and to calibrate the preliminary risk assessment value according to the actual confidence level to obtain a calibrated risk assessment value. The model optimization module is used to optimize the first risk assessment model based on the calibration risk assessment value and the preliminary risk assessment value to obtain a second risk assessment model. The second assessment module is used to assess the risk based on the fused feature data using the second risk assessment model to obtain a target risk assessment value and a second assessment confidence level, wherein the second assessment confidence level meets a preset confidence level condition.

[0012] A third aspect of this application provides an electronic device comprising a processor and a memory, wherein the processor is configured to implement the multimodal-based non-performing asset risk assessment method by executing a computer program stored in the memory.

[0013] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the multimodal-based non-performing asset risk assessment method described above.

[0014] This application has at least the following beneficial effects: This application's multimodal non-performing asset risk assessment method not only overcomes the limitations of a single information source through multimodal fusion, enhancing the reliability of the assessment, but also utilizes a confidence calibration mechanism to output risk values ​​with quantifiable credibility. This enables adaptive optimization and continuous performance improvement of the assessment model, allowing the system to dynamically adapt to business changes and provide highly accurate and reliable decision-making basis for risk management and asset disposal, effectively reducing the risk of misjudgment. Attached Figure Description

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

[0016] Figure 1 A flowchart illustrating a multimodal-based risk assessment method for non-performing assets provided in this application embodiment; Figure 2 This is a schematic diagram of a multimodal data storage interface provided in an embodiment of this application; Figure 3 A schematic diagram of the optimized interface provided for the embodiments of this application; Figure 4 A schematic diagram of a multimodal non-performing asset risk assessment device provided in this application embodiment; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] To better understand the above-mentioned objectives, features, and advantages of this application, the application will be described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein in the specification of this application is for the purpose of describing an embodiment in one alternative implementation and is not intended to be limiting of the application.

[0019] Example 1 Figure 1 This is a flowchart of the multimodal-based non-performing asset risk assessment method provided in Embodiment 1 of this application. The multimodal-based non-performing asset risk assessment method specifically includes the following steps.

[0020] S1, fused feature data is obtained based on multimodal data of non-performing assets.

[0021] Non-performing assets typically refer to financial claims and collateral assets held by enterprises or individuals that are subject to uncertainty regarding recovery or have already depreciated in value.

[0022] Non-performing asset risk is the result of multiple dimensions (risk of the debt itself, risk of debtor performance, risk of realization of collateral, risk of litigation and enforcement, etc.). A single data can only reflect one dimension of risk (for example, credit data alone cannot reflect the risk of collateral value) and has the limitation of information (for example, ledger data alone does not include debtor operating data, and it is impossible to predict its subsequent performance ability). Therefore, it is necessary to obtain multimodal data of non-performing assets and conduct risk assessment of non-performing assets based on multimodal data.

[0023] Multimodal data on non-performing assets refers to a collection of heterogeneous data of various types that characterize the risk attributes of non-performing assets. This can include structured financial data, unstructured legal text data, collateral image data, and transaction flow data. Structured financial data includes, but is not limited to, key indicators in balance sheets, income statements, and cash flow statements; unstructured legal text data includes, but is not limited to, contract texts, legal judgments, and asset ownership certificates; collateral image data includes, but is not limited to, photos of real estate, equipment, and on-site inspection images; and transaction flow data includes, but is not limited to, historical repayment records and time-series data of fund transfers. Multimodal data contains richer risk information and more comprehensive assessment dimensions, laying the foundation for quantifying the reliability of subsequent assessment results. (See also...) Figure 2 As shown, after acquiring multimodal data, preprocessing operations such as cleaning (removing invalid or missing data, correcting outliers) and format standardization (unifying data field definitions, converting unstructured data to readable formats) can be performed. After preprocessing, the multimodal data is stored. At this point, the electronic device records the data acquisition time, type, preprocessing operation time, and preprocessing operation status for each data point. Feature-level fusion algorithms (such as weighted fusion and attention-based fusion) are then used to fuse the cleaned and format-standardized multimodal data to obtain fused feature data.

[0024] In an optional embodiment, the fused feature data obtained based on the multimodal data of non-performing assets includes: Acquire structured financial data, unstructured legal text data, collateral image data, and transaction flow data from multimodal data; Generate structured feature vectors based on structured financial data; Generate text feature vectors based on unstructured legal text data; Generate image feature vectors based on image data of the pledged collateral; Generate time-series feature vectors based on transaction log data; The structured feature vector, text feature vector, image feature vector, and time-series feature vector are weighted and fused using a multi-head attention mechanism to obtain fused feature data.

[0025] The structured financial data is standardized to eliminate the influence of dimensions, and then quantitative features (such as overdue days and guarantee coverage ratio) are extracted to obtain a structured feature vector V_s (such as debt-to-equity ratio, current ratio, net profit, revenue growth rate, debt service coverage ratio, etc.), with a dimension of d_s.

[0026] The legal text data is preprocessed, including word segmentation and removal of stop words. Semantic features (such as the weight of risk keywords in due diligence reports and the integrity features of collateral) are extracted using text mining techniques and used as text feature vectors V_t (such as risk keywords such as no property available for execution, joint and several liability, and statute of limitations expiration), with a dimension of d_t.

[0027] Image data is preprocessed, including size normalization and data augmentation. Image features are extracted from the image data of collateral (such as real estate, vehicles, equipment, land, etc.) corresponding to non-performing assets. Visual information reflecting the value and integrity of collateral in the image is converted into a standardized vector to obtain an image feature vector V_i of dimension d_i (such as the integrity of the house layout, the degree of wear and tear of equipment, the land use status, whether the collateral is damaged / idle, etc.).

[0028] The transaction flow data is cleaned and serialized. Based on the historical transaction flow data of debtors and corresponding non-performing asset accounts, transaction features with time-series patterns are extracted and transformed into standardized vectors with time-series attributes, resulting in a time-series feature vector V_q (such as repayment frequency in the past 6 / 12 months, net inflow and outflow trends of funds, large abnormal transfer records, and time-series patterns of performance stability), with a dimension of _d_q. Using the unique identifier of non-performing assets (such as debt number) as a benchmark, the features extracted from different modalities are matched in dimension and aligned in time sequence to eliminate feature misalignment problems caused by different sources and formats of multimodal data, resulting in aligned multimodal feature vectors (structured feature vectors, text feature vectors, image feature vectors, and time-series feature vectors).

[0029] The aligned structured feature vectors, text feature vectors, image feature vectors, and time-series feature vectors are weighted and fused using a multi-head attention mechanism to obtain fused feature data. For example, firstly, the four feature vectors are mapped to a unified dimension d_h through linear projection to obtain the query vector Q, key vector K, and value vector V; then, the attention weights are calculated and fused using a multi-head attention mechanism, the process of which can be represented as: , in, It is the query, key, and value matrix of the h-th head. d is the dimension of the key vector; the outputs of all heads are concatenated and passed through a linear layer to obtain the fused feature vector F, with dimension d_f. This multi-head attention mechanism, by computing multiple attention heads in parallel, can simultaneously focus on information from different feature subspaces and adaptively weight the contributions of each modality feature, thereby achieving better feature fusion.

[0030] This application captures the complex dependencies between different modal features by parallel computing the correlations in different feature subspaces and assigns appropriate weights to each feature. The final output fused feature data is a multi-dimensional representation vector that integrates financial, legal, physical asset status, and historical behavior, providing a more comprehensive information foundation for subsequent risk assessment.

[0031] S2 inputs the fused feature data into the trained first risk assessment model to obtain the preliminary risk assessment value and the first assessment confidence level.

[0032] The first risk assessment model can be a deep neural network model pre-trained based on fused feature data of historical non-performing assets and their actual assessment value labels. The model outputs a preliminary risk assessment value indicating the level of risk, as well as an assessment confidence score. The assessment confidence score reflects the degree of certainty the first risk assessment model has regarding the risk assessment given the current input data, and is typically calculated based on the variance of the probability distribution predicted by the model, Dropout randomness, or the degree of prediction divergence among different sub-models in the ensemble model.

[0033] In one embodiment, the first risk assessment model is obtained by training a gradient boosting decision tree based on the fused characteristic data of historical non-performing assets and their actual assessment value labels.

[0034] The integrated characteristic data of the non-performing assets to be evaluated are input into the trained first risk assessment model, and the first risk assessment model outputs the preliminary risk assessment value and the first assessment confidence level.

[0035] Optionally, the first assessment confidence level is calculated based on the variance of the probability distribution predicted by the first risk assessment model. For example, the first assessment confidence level = 1 / (1 + variance).

[0036] S3. Determine the actual confidence level based on the first assessment confidence level, and calibrate the preliminary risk assessment value according to the actual confidence level to obtain the calibrated risk assessment value.

[0037] The first assessment confidence level reflects the reliability of the preliminary assessment result output by the first risk assessment model based on the fused feature data of the current input. The first assessment confidence level can be a risk level (e.g., 1-5) or a risk probability (e.g., 0%-100%). Since the preliminary risk assessment value is the raw output of the first risk assessment model, it may be susceptible to bias due to missing data and feature noise; direct use would lead to assessment distortion. It is necessary to use the first assessment confidence level as the core benchmark, quantify it to obtain the final confidence index (i.e., the actual confidence level), and calibrate the preliminary risk assessment value based on the actual confidence level to output the calibrated risk assessment value.

[0038] Actual confidence level is the core quantitative basis for measuring whether the risk assessment results have decision-making value. The value range can be [0,1], and the higher the value, the more reliable the assessment results are.

[0039] In an optional embodiment, determining the actual confidence level based on the first assessment confidence level includes: obtaining the integrity, consistency, and timeliness of the multimodal data; A data quality score is obtained based on the completeness, consistency, and timeliness mentioned above. The actual confidence level is obtained based on the first assessment confidence level and the data quality score.

[0040] The completeness of multimodal data is the ratio of the number of data items obtained to the total number of predefined essential data items for that type of data. For example, suppose structured financial data has six essential data items: debt-to-equity ratio, current ratio, net profit, revenue, total assets, and debt service coverage ratio. If only four items (debt-to-equity ratio, current ratio, net profit, and revenue) are actually obtained, and the total assets and debt service coverage ratio are missing, then the completeness of the structured financial data is 0.67. Similarly, unstructured legal text data has four essential data items: loan agreement, guarantee agreement, court judgment, and enforcement ruling. If all four items are obtained completely, then the completeness of the unstructured legal text data is 1.0. The completeness of multimodal data is obtained by weighted averaging the four types of modal data.

[0041] The consistency of multimodal data is the ratio of the number of consistent data links to the total number of verified links. For example, suppose the predefined total number of verified links is: ① principal amount of the debt ② debtor's name ③ address of the collateral ④ overdue start time ⑤ scope of guarantee liability. Link ① principal amount of the debt: consistent in the structured ledger (10 million yuan) and the loan contract text (10 million yuan); Link ② debtor's name: consistent in the structured ledger (XX Building Materials Co., Ltd.) and the judgment text (XX Building Materials Co., Ltd.); Link ③ address of the collateral: consistent in the structured mortgage ledger (88, XX Road, XX District, XX City). The number of consistent related points is consistent with the attached map of the collateral (No. 88, XX Road, XX District, XX City); the overdue start time is inconsistent between the structured repayment ledger (May 10, 2023) and the collection record (May 11, 2023); the scope of guarantee liability is inconsistent between the guarantee contract text (joint and several guarantee, fully covering principal and interest) and the judgment (joint and several guarantee, only covering principal). Therefore, the number of consistent related points is 3, the total number of verified related points is 5, and the consistency of the multimodal data of this non-performing asset is 0.6.

[0042] The timeliness of multimodal data is calculated based on a decay score derived from the time difference between the data generation or update time and the current evaluation time. The larger the time difference, the lower the score. For example, if the time difference is ≤3 months, the timeliness is 1.0; if 3 months < time difference ≤6 months, the timeliness is 0.8.

[0043] Data quality scores can be calculated based on completeness, consistency, and timeliness. Optionally, a weighted average of completeness, consistency, and timeliness can be used to calculate the data quality score, where the completeness weight ranges from 0.5 to 0.8, the consistency weight ranges from 0.2 to 0.4, and the timeliness weight ranges from 0.1 to 0.3. The sum of the completeness weight, consistency weight, and timeliness weight is 1. For example, the completeness weight is 0.5, the consistency weight is 0.3, and the timeliness weight is 0.2.

[0044] After obtaining the data quality score, the actual confidence level can be derived based on the first assessment confidence level and the data quality score. Optionally, the actual confidence level can be calculated as the product of the first assessment confidence level and the data quality score.

[0045] The above optional embodiments, by introducing a data quality scoring mechanism to correct the initial confidence level of the model output, ensure that the final actual confidence level used for calibration not only reflects the model's own judgment and grasp, but also, more importantly, incorporates the quality factors of the multimodal data source. This effectively avoids the situation where the reliability of the model is mistakenly overestimated due to defects in the input data itself, and significantly improves the authenticity and practicality of the confidence level measurement.

[0046] In an optional embodiment, the preliminary risk assessment value is calibrated based on the actual confidence level to obtain a calibrated risk assessment value, including: The actual confidence level is compared with a preset first confidence threshold and a preset second confidence threshold; If the actual confidence level is greater than the first confidence level threshold, then the preliminary risk assessment value will be output as the calibration risk assessment value. If the actual confidence level is less than the first confidence threshold but greater than the second confidence threshold, the preliminary risk assessment value is calibrated based on a preset standard risk benchmark value to obtain a calibrated risk assessment value.

[0047] The first confidence threshold is a pre-set high confidence threshold, which is a threshold standard for determining that the preliminary risk assessment value has high reference value, indicating that the reliability of the assessment result meets the standard. The second confidence threshold is a pre-set low confidence threshold, which is a threshold standard for determining that the preliminary risk assessment value has limited reference value, indicating that the reliability of the assessment result does not meet the standard. The first confidence threshold is greater than the second confidence threshold. For example, the first confidence threshold can be set to 0.8, and the second confidence threshold can be set to 0.5.

[0048] This application achieves tiered calibration by setting confidence thresholds, significantly improving the adaptability and robustness of the assessment system's output. When the actual confidence level is greater than the first confidence threshold, it indicates a high confidence result that can be directly adopted, ensuring the timeliness of reliable assessment. Conversely, when the actual confidence level is less than the first confidence threshold but greater than the second confidence threshold, it indicates a low confidence result that requires calibration using a standard risk benchmark value, effectively correcting biases caused by model uncertainty. This tiered strategy allows the system to adopt differentiated processing based on its own level of confidence, balancing efficiency and accuracy while ensuring the rationality and reliability of the final output. If the actual confidence level is less than the second confidence threshold, it is necessary to reacquire multimodal data of non-performing assets, obtain fused feature data based on the multimodal data, and then use the first risk assessment model based on the fused feature data to obtain a preliminary risk assessment value and a first assessment confidence level.

[0049] In an optional embodiment, comparing the actual confidence level with a first confidence threshold and a second confidence threshold includes: Determine the industry risk level and collateral type of non-performing assets; Take the first adjustment coefficient corresponding to the industry risk level; Obtain the second adjustment coefficient corresponding to the type of collateral; The first confidence threshold and the second confidence threshold are adjusted based on the first adjustment coefficient and the second adjustment coefficient to obtain a first target confidence threshold and a second target confidence threshold, wherein the first target confidence threshold is greater than the second target confidence threshold; The actual confidence level is compared with the first target confidence threshold and the second target confidence threshold.

[0050] Among these, the industry risk levels of non-performing assets can be divided into high risk, medium risk, and low risk. The types of collateral can be divided into real estate, movable property, and unsecured collateral.

[0051] Based on the industry risk level, query the predefined first mapping table. The first adjustment coefficient k1 corresponding to high risk ranges from 1.05 to 1.1, the first adjustment coefficient k2 corresponding to medium risk is 1.0, and the first adjustment coefficient k3 corresponding to low risk ranges from 0.9 to 0.95.

[0052] Based on the type of collateral, query the predefined second mapping table. The second adjustment coefficient m1 for real estate collateral ranges from 0.95 to 1.0, the second adjustment coefficient m2 for movable property collateral is 1.0, and the second adjustment coefficient m3 for uncollateralized property ranges from 1.05 to 1.1.

[0053] The first and second confidence thresholds are adjusted based on the first and second adjustment coefficients to obtain the first and second target confidence thresholds, respectively. The first target confidence threshold is greater than the second target confidence threshold. Specifically, the first target confidence threshold = first confidence threshold × first adjustment coefficient × second adjustment coefficient, and the first target confidence threshold is less than or equal to 0.95. The second target confidence threshold = second confidence threshold × first adjustment coefficient × second adjustment coefficient, and it must satisfy the condition that the second target confidence threshold is less than the first target confidence threshold.

[0054] The actual confidence level is compared with the first target confidence threshold and the second target confidence threshold. This comparison divides the interval into high confidence, medium confidence, and low confidence intervals. Specifically, the comparison results are as follows: an actual confidence level greater than or equal to the first target confidence threshold is a high confidence interval; an actual confidence level less than the first target confidence threshold but greater than the second target confidence threshold is a medium confidence interval; and an actual confidence level less than the second target confidence threshold is a low confidence interval.

[0055] This application dynamically adjusts the confidence threshold by introducing industry risk levels and collateral types, enabling the calibration standards to accurately adapt to the inherent risk attributes of different assets, significantly improving the targeting and business scenario fit of the calibration strategy. This design effectively overcomes the rigidity of fixed thresholds, improving the overall accuracy, rationality, and generalizability of risk assessment and calibration from the source.

[0056] In an optional embodiment, the preliminary risk assessment value is calibrated based on a preset standard risk benchmark value to obtain a calibrated risk assessment value, including: Based on the actual confidence level, the first confidence threshold, and the second confidence threshold, the confidence matching weight is obtained; The initial risk assessment value is calibrated based on confidence level matching weights and a preset standard risk benchmark value to obtain a calibrated risk assessment value.

[0057] The confidence matching weight is used to characterize the weight ratio of the preliminary risk assessment value in the final calibration process, and to provide quantitative support for differentiated calibration. It is the weight allocation between the preliminary risk assessment value and the standard risk benchmark value in the calibration calculation. The size of the confidence matching weight is positively correlated with the actual confidence level, and is used to regulate the contribution of the two types of values ​​to the calibration risk assessment value.

[0058] Using the second confidence threshold as the lower limit and the first confidence threshold as the upper limit, the actual confidence level is mapped to the corresponding weight. The closer the actual confidence level is to the first confidence threshold, the closer the confidence level matching weight is to 1, which means that the preliminary risk assessment value is more reliable. The closer the actual confidence level is to the second confidence threshold, the closer the confidence level matching weight is to 0, which means that the preliminary risk assessment value is less reliable. This achieves a precise fit where the higher the credibility, the greater the weight of the original assessment value, thus solving the problem of traditional calibration lacking quantitative basis.

[0059] For example, the confidence level matching weight = (actual confidence level - second confidence level threshold) / (first confidence level threshold - second confidence level threshold). Using the difference between the first and second confidence level thresholds as the denominator ensures the standardization and uniqueness of the weight calculation results, avoiding weight distortion. A weighted fusion calibration logic is employed to quantitatively integrate the preliminary risk assessment value and the standard risk benchmark value according to the confidence level matching weight, outputting a final risk assessment value that considers both the original assessment result and the benchmark reference. This ensures that the calibration result accurately reflects the actual risk of the asset while avoiding assessment bias under low confidence levels.

[0060] The standard risk benchmark value is a pre-set benchmark risk value for the non-performing asset industry / category. It is obtained by statistical analysis of historical risk data of non-performing assets of the same type and with the same qualifications. It serves as a risk reference value in low-reliability scenarios, ensuring that the calibration results are reasonable for the industry. The weights assigned to the standard risk benchmark value are complementary to the confidence level weights and their sum is 1, ensuring the standardization and closed-loop nature of the calibration calculation.

[0061] For example, the calibration risk assessment value = confidence matching weight × preliminary risk assessment value + (1 - confidence matching weight) × standard risk benchmark value.

[0062] First, by using a confidence-matching weight formula, the qualitative confidence judgment is transformed into a quantitative weight coefficient, solving the technical problem of how confidence can be transformed into a calibration basis. Then, by using a calibration risk assessment value formula, the weighted fusion of the two assessment values ​​is achieved based on the weight coefficient, solving the technical problem of how to accurately calibrate risk values ​​based on confidence. The combination of the two not only solves the pain point of the lack of reliability quantification in the assessment results of existing technologies, but also solves the defect that the original assessment bias cannot be corrected, forming a complete technical closed loop.

[0063] In an optional embodiment, if the actual confidence level is greater than the first target confidence level threshold, the preliminary risk assessment value is output as the calibration risk assessment value. If the actual confidence level is less than the first target confidence level threshold but greater than the second target confidence level threshold, the preliminary risk assessment value is calibrated based on a preset standard risk benchmark value to obtain the calibration risk assessment value. If the actual confidence level is less than the second confidence level threshold, an interactive correction request is generated and a manual correction value is obtained. The validity of the manual correction value is verified. When the manual correction value passes the verification, a manual correction confidence level weight representing the highest confidence level is assigned to the manual correction value. The calibration risk assessment value is obtained based on the manual correction confidence level weight, the manual correction value, and the preliminary risk assessment value. Specifically, the calibration risk assessment value = manual correction confidence level weight × preliminary risk assessment value + (1 - manual correction confidence level weight) × manual correction value.

[0064] The validity verification based on manually adjusted values ​​includes two aspects: verifying whether the manually adjusted values ​​are within a reasonable range of the corresponding asset's historical risk values, and performing a rapid logical consistency check based on the submitted multimodal key feature summary.

[0065] In other embodiments, the confidence matching weight can also be obtained based on the actual confidence level, the first target confidence level threshold, and the second target confidence level threshold.

[0066] This application dynamically modulates the calibration process of the standard risk benchmark by introducing confidence level matching weights, achieving adaptive matching between the calibration intensity and the uncertainty of the model itself. This makes the calibration operation no longer a fixed or linear adjustment, but allows for fine-grained and differentiated control of the proportion of the standard risk benchmark introduced based on the specific level of the assessment confidence. This mechanism significantly improves the accuracy and reasonableness of calibration results within the medium confidence range, ensuring a smooth and optimal trade-off between the final evaluation value and the prior benchmark, further enhancing the consistency of the system output and the robustness of decision-making.

[0067] Furthermore, this application introduces a parameter update and recalibration mechanism after human intervention, which not only makes full use of the experience and judgment of human experts to make up for the shortcomings of the model, but also seamlessly feeds the results of human intervention into the automated process in a parameterized manner, realizing the effective integration and closed-loop optimization of human-machine intelligence, and significantly improving the system's ability to handle complex and difficult cases and the quality of the final decision.

[0068] S4. Based on the calibration risk assessment value and the preliminary risk assessment value, optimize the first risk assessment model to obtain the second risk assessment model.

[0069] The model is optimized based on the calibrated risk assessment value and the preliminary risk assessment value. The calibrated risk assessment value is used as the accurate label and the preliminary risk assessment value is used as the input reference. The multimodal feature weights and fitting parameters of the first risk assessment model are adjusted in reverse iteration to correct the model bias, improve the assessment accuracy and the reliability of the results, and the optimized second risk assessment model has better performance.

[0070] In an optional embodiment, optimizing the first risk assessment model based on the calibration risk assessment value and the preliminary risk assessment value to obtain the second risk assessment model includes: Calculate the difference between the calibration risk assessment value and the preliminary risk assessment value; The confidence rate of change is obtained based on the confidence matching weight and the actual confidence level. A calibration residual loss function is constructed based on the difference, the confidence rate of change, and the preset confidence sensitivity regularization coefficient; The first risk assessment model is iteratively optimized using the calibration residual loss function; At the end of the iterative optimization, the second risk assessment model is obtained.

[0071] Calculate the difference ΔR between the calibration risk assessment value and the preliminary risk assessment value.

[0072] Based on the actual confidence level, the preset first confidence level threshold, and the second confidence level threshold, the confidence matching weight ω is obtained.

[0073] The confidence rate of change is obtained based on the confidence matching weight and the actual confidence level; where the confidence rate of change γ is the confidence matching weight ω × the actual confidence level. .

[0074] For example, a calibration residual loss function is constructed based on the difference, the rate of change of confidence level, and a preset confidence level sensitivity regularization coefficient: L=||ΔR||2 2 +λ·|( ) / ( × )|, where ΔR is the squared term of the difference between the calibrated risk assessment data and the preliminary risk assessment value, used to constrain the deviation range between the calibrated data and the preliminary assessment value. λ is the confidence sensitivity regularization coefficient; |( ) / ( × )| is the confidence sensitivity regularization term, which is used to constrain the degree to which the confidence matching weight changes with the actual confidence level.

[0075] The fused feature data is then input back into the risk assessment model, and the model parameters are iteratively updated using stochastic gradient descent to shift the actual confidence level corresponding to the initial risk assessment value output by the model towards the higher confidence interval. The risk assessment model at the end of the optimization process is the second risk assessment model. Figure 3 As shown, the risk assessment model after three iterations outputs a second assessment confidence level of 92% with an accuracy rate.

[0076] This application optimizes the model by dynamically adjusting the confidence level matching weights and the confidence level change rate, focusing the model training on samples with high uncertainty, thus significantly improving the targeting and convergence efficiency of the optimization. The constructed calibration residual loss function effectively combines the evaluation value bias with the dynamics of the model confidence, enabling the model to simultaneously improve prediction accuracy and result stability during iterative optimization, thereby enhancing the model's generalization ability and robustness. The resulting second risk assessment model exhibits better evaluation consistency, and its output target risk assessment value has higher reliability, significantly improving the system's decision-making quality and reliability in actual business operations.

[0077] S5, the target risk assessment value and the second assessment confidence level are obtained by evaluating the fused feature data through the second risk assessment model. The second assessment confidence level meets the preset confidence level conditions.

[0078] Multimodal fusion feature data is input into the optimized second risk assessment model for risk assessment. The second risk assessment model simultaneously outputs the target risk assessment value and the second assessment confidence level. Among them, the target risk assessment value characterizes the core risk level of non-performing assets, and the second assessment confidence level measures the reliability of the assessment results.

[0079] The preset confidence level conditions are that the second assessment confidence level is greater than the preset first confidence level threshold, or the second assessment confidence level is greater than the first target confidence level threshold. When the second assessment confidence level meets the preset confidence level conditions, the target risk assessment value can be directly used for non-performing asset disposal and risk pricing decisions. When the second assessment confidence level does not meet the preset confidence level conditions, the first risk assessment model needs to be optimized again until the second assessment confidence level obtained by evaluating based on fused feature data using the second risk assessment model meets the preset confidence level conditions. The risk assessment value corresponding to the second assessment confidence level that meets the preset confidence level conditions is then used as the target risk assessment value.

[0080] As can be seen, the multimodal non-performing asset risk assessment method provided in this application overcomes the limitations of a single data source by integrating multimodal data such as financial and legal documents, images, and transaction records, providing a richer and more objective information foundation for risk assessment. By designing a dedicated calibration residual loss function and using the calibrated results to iteratively optimize the model, self-improvement and performance enhancement are achieved, resulting in a final second risk assessment model with stronger generalization ability and robustness. Furthermore, through strategies such as dynamic adjustment of confidence thresholds and hierarchical calibration processing, the system can adopt differentiated and optimized processing methods for assessment results with different levels of certainty. In summary, the multimodal non-performing asset risk assessment method of this application not only overcomes the limitations of a single information source and enhances the reliability of the assessment through multimodal fusion, but also utilizes a confidence calibration mechanism to output risk values ​​with credible quantification, achieving adaptive optimization and continuous performance improvement of the assessment model. This enables the system to dynamically adapt to business changes, providing highly accurate and reliable decision-making basis for risk management and asset disposal, effectively reducing the risk of misjudgment.

[0081] Example 2 Figure 4 This is a schematic diagram of the module of the multimodal non-performing asset risk assessment device provided in Embodiment 2 of this application.

[0082] In some embodiments, the multimodal non-performing asset risk assessment device may include multiple functional modules composed of computer program segments. The computer programs for each program segment of the multimodal non-performing asset risk assessment device may be stored in the memory of an electronic device and executed by at least one processor to perform (see details). Figure 1(Description) A multimodal non-performing asset risk assessment function. Based on its functions, it can be divided into multiple functional modules. These modules may include: a feature generation module 1, a first assessment module 2, a risk calibration module 3, a model optimization module 4, and a second assessment module 5. The term "module" in this application refers to a series of computer program segments that can be executed by at least one processor and perform a fixed function, stored in memory. In this embodiment, the functions of each module will be detailed in subsequent embodiments.

[0083] Feature generation module 1 is used to obtain fused feature data based on multimodal data of non-performing assets; The first evaluation module 2 is used to input the fused feature data into the trained first risk assessment model to obtain the preliminary risk assessment value and the first assessment confidence level. Risk calibration module 3 is used to determine the actual confidence level based on the first assessment confidence level, and to calibrate the preliminary risk assessment value according to the actual confidence level to obtain the calibrated risk assessment value; Model optimization module 4 is used to optimize the first risk assessment model based on the calibration risk assessment value and the preliminary risk assessment value to obtain the second risk assessment model; The second assessment module 5 is used to conduct an assessment based on fused feature data using the second risk assessment model to obtain the target risk assessment value and the second assessment confidence level. The second assessment confidence level meets the preset confidence level conditions.

[0084] This application's multimodal non-performing asset risk assessment device overcomes the limitations of a single data source, providing a richer and more objective information foundation for risk assessment. Furthermore, it achieves self-improvement and performance enhancement of the model, resulting in a final second risk assessment model with stronger generalization ability and robustness. Therefore, this application's non-performing asset risk assessment device not only overcomes the limitations of a single information source and enhances the reliability of the assessment, but also achieves adaptive optimization and continuous performance improvement of the assessment model. This enables the system to dynamically adapt to business changes, providing highly accurate and reliable decision-making basis for risk control and asset disposal, effectively reducing the risk of misjudgment.

[0085] It should be understood that the various variations and specific embodiments of the multimodal non-performing asset risk assessment method provided in Embodiment 1 above are also applicable to the multimodal non-performing asset risk assessment device in this embodiment. Through the detailed description of the multimodal non-performing asset risk assessment method described above, those skilled in the art can clearly understand the implementation process of the multimodal non-performing asset risk assessment device in this embodiment. For the sake of brevity, it will not be described in detail here.

[0086] Example 3 A computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the multimodal non-performing asset risk assessment method as described in Example 1. This method not only overcomes the limitations of a single information source and enhances the reliability of the assessment, but also enables adaptive optimization and continuous performance improvement of the assessment model. This allows the system to dynamically adapt to business changes, providing highly accurate and reliable decision-making support for risk management and asset disposal, effectively reducing the risk of misjudgment.

[0087] Example 4 See Figure 5 The diagram shown is a structural schematic of an electronic device according to an embodiment of this application.

[0088] Electronic device 6 includes a memory 61, at least one processor 62, and at least one communication bus 63. The electronic device not only overcomes the limitations of a single information source and enhances the reliability of the assessment, but also enables adaptive optimization and continuous performance improvement of the assessment model. This allows the system to dynamically adapt to business changes, providing highly accurate and reliable decision-making support for risk management and asset disposal, effectively reducing the risk of misjudgment.

[0089] Those skilled in the art should understand that Figure 5 The structure of the electronic device shown does not constitute a limitation of the embodiments of this application. It can be a bus structure or a star structure. The electronic device may also include more or fewer other hardware or software than shown, or different component arrangements.

[0090] In some embodiments, the electronic device is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions. Its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), digital processors, and embedded devices. The electronic device may also include user equipment, which includes, but is not limited to, any electronic product capable of human-computer interaction with a user via a keyboard, mouse, remote control, touchpad, or voice control device, such as a personal computer, tablet computer, smartphone, or digital camera.

[0091] In the embodiments provided in this application, it should be understood that the disclosed methods, apparatuses, computer-readable storage media, and electronic devices can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple components or modules may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices, components, or modules may be electrical, mechanical, or other forms.

[0092] The components described as separate parts may or may not be physically separate. The components shown as components may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the components can be selected to achieve the purpose of this embodiment according to actual needs.

[0093] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each component can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0094] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the intelligent data backup method for electronic devices described in the various embodiments of the present invention. The aforementioned storage medium includes: USB flash drive, portable hard drive, read-only memory (ROM). Various media that can store program code, such as only memory, random access memory (RAM), magnetic disks, or optical disks.

[0095] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0096] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0097] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A multimodal risk assessment method for non-performing assets, characterized in that, The method includes: Fusion feature data is obtained based on multimodal data of non-performing assets; The fused feature data is input into the trained first risk assessment model to obtain a preliminary risk assessment value and a first assessment confidence level. The actual confidence level is determined based on the first assessment confidence level, and the preliminary risk assessment value is calibrated according to the actual confidence level to obtain the calibrated risk assessment value; The first risk assessment model is optimized based on the calibration risk assessment value and the preliminary risk assessment value to obtain the second risk assessment model; The second risk assessment model evaluates the risk based on the fused feature data to obtain a target risk assessment value and a second assessment confidence level, wherein the second assessment confidence level meets the preset confidence level conditions.

2. The multimodal-based non-performing asset risk assessment method according to claim 1, characterized in that, The step of calibrating the preliminary risk assessment value based on the actual confidence level to obtain the calibrated risk assessment value includes: The actual confidence level is compared with a preset first confidence threshold and a preset second confidence threshold, wherein the first confidence threshold is greater than the second confidence threshold; If the actual confidence level is greater than the first confidence level threshold, then the preliminary risk assessment value will be output as the calibration risk assessment value. If the actual confidence level is less than the first confidence threshold but greater than the second confidence threshold, the preliminary risk assessment value is calibrated based on a preset standard risk benchmark value to obtain a calibrated risk assessment value.

3. The multimodal-based non-performing asset risk assessment method according to claim 2, characterized in that, The step of comparing the actual confidence level with the first confidence threshold and the second confidence threshold includes: Determine the industry risk level and collateral type of non-performing assets; Obtain the first adjustment coefficient corresponding to the industry risk level; Obtain the second adjustment coefficient corresponding to the type of collateral; The first confidence threshold and the second confidence threshold are adjusted based on the first adjustment coefficient and the second adjustment coefficient to obtain a first target confidence threshold and a second target confidence threshold, wherein the first target confidence threshold is greater than the second target confidence threshold; The actual confidence level is compared with the first target confidence threshold and the second target confidence threshold.

4. The multimodal-based non-performing asset risk assessment method according to claim 2, characterized in that, The calibration of the preliminary risk assessment value based on a preset standard risk benchmark value to obtain the calibrated risk assessment value includes: Based on the actual confidence level, the first confidence threshold, and the second confidence threshold, the confidence matching weight is obtained; The preliminary risk assessment value is calibrated based on the confidence matching weight and the preset standard risk benchmark value to obtain the calibrated risk assessment value.

5. The multimodal-based non-performing asset risk assessment method according to claim 4, characterized in that, The step of optimizing the first risk assessment model based on the calibration risk assessment value and the preliminary risk assessment value to obtain the second risk assessment model includes: Calculate the difference between the calibration risk assessment value and the preliminary risk assessment value; The confidence rate of change is obtained based on the confidence matching weight and the actual confidence level. A calibration residual loss function is constructed based on the difference, the confidence rate of change, and the preset confidence sensitivity regularization coefficient; The first risk assessment model is iteratively optimized using the calibration residual loss function; At the end of the iterative optimization, the second risk assessment model is obtained.

6. The multimodal-based non-performing asset risk assessment method according to any one of claims 1 to 5, characterized in that, The fused feature data obtained from the multimodal data based on non-performing assets includes: Acquire structured financial data, unstructured legal text data, collateral image data, and transaction flow data from the multimodal data; A structured feature vector is generated based on the structured financial data; Generate text feature vectors based on the aforementioned unstructured legal text data; Generate image feature vectors based on the image data of the pledged collateral; Generate a time-series feature vector based on the transaction flow data; The structured feature vector, text feature vector, image feature vector, and time-series feature vector are weighted and fused using a multi-head attention mechanism to obtain the fused feature data.

7. The multimodal-based non-performing asset risk assessment method according to any one of claims 1 to 5, characterized in that, The step of determining the actual confidence level based on the first assessment confidence level includes: The integrity, consistency, and timeliness of the multimodal data are obtained; A data quality score is obtained based on the completeness, consistency, and timeliness mentioned above. The actual confidence level is obtained based on the first assessment confidence level and the data quality score.

8. A multimodal non-performing asset risk assessment device, characterized in that, The device includes: The feature generation module is used to obtain fused feature data based on multimodal data of non-performing assets; The first evaluation module is used to input the fused feature data into the trained first risk assessment model to obtain a preliminary risk assessment value and a first assessment confidence level. The risk calibration module is used to determine the actual confidence level based on the first assessment confidence level, and to calibrate the preliminary risk assessment value according to the actual confidence level to obtain a calibrated risk assessment value. The model optimization module is used to optimize the first risk assessment model based on the calibration risk assessment value and the preliminary risk assessment value to obtain a second risk assessment model. The second assessment module is used to assess the risk based on the fused feature data using the second risk assessment model to obtain a target risk assessment value and a second assessment confidence level, wherein the second assessment confidence level meets a preset confidence level condition.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the multimodal non-performing asset risk assessment method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the multimodal-based non-performing asset risk assessment method according to any one of claims 1 to 7.

Images