Financial risk analysis method, device, storage medium and product

Abstract

This application relates to the field of financial risk analysis technology, providing financial risk analysis methods, equipment, storage media, and products. The method includes: obtaining a candidate expert module set based on multimodal data to be analyzed, and obtaining initial weight information based on a dynamic strategy; obtaining semantic feature information based on the multimodal data to be analyzed; obtaining an activated expert module set and target weight information based on the semantic feature information, the candidate expert module set, and the initial weight information; processing the semantic feature information based on the activated expert module set to obtain expert processing result set information, and combining it with the target weight information to obtain risk analysis information. Through a two-level expert module selection process, the accuracy of expert selection is improved, and the accuracy of risk analysis is enhanced. Determining initial weight information based on a dynamic strategy allows for rapid adaptation to the ever-changing risk strategies in the financial field, improving the adaptability of risk analysis and making the risk analysis information more reliable.

Description

Technical Field

[0001] This application relates to the field of financial risk analysis technology, and in particular to financial risk analysis methods, equipment, storage media and products. Background Technology

[0002] With the development of artificial intelligence (AI) technology, AI is widely used in the field of financial risk analysis, especially risk assessment models based on machine learning and deep learning. These models play a crucial role in various financial risk analysis scenarios and can effectively improve analysis efficiency. Mixture of Experts (MoE) models, as an effective AI architecture for improving model capacity and performance, distribute the input analysis task to different expert modules for processing.

[0003] However, in existing technologies, the routing strategy for selecting expert modules in hybrid expert models is too simple. When dealing with complex financial multimodal risk analysis tasks, there is a problem of insufficient accuracy in expert selection. It is unable to select the optimal expert module based on inputs containing both structured and unstructured data, thereby reducing the accuracy of risk analysis. At the same time, it cannot adapt quickly to the ever-changing risk strategies in the financial field, resulting in poor adaptability of risk analysis.

[0004] Therefore, how to accurately select expert modules for processing and quickly adapt to risk strategies in order to improve the accuracy and adaptability of risk analysis is an urgent problem to be solved. Summary of the Invention

[0005] This application aims to address at least one of the technical problems existing in related technologies. To this end, this application proposes a financial risk analysis method that can accurately select expert modules and quickly adapt to risk strategies, thereby improving the accuracy and adaptability of risk analysis.

[0006] This application also proposes devices, storage media, and products.

[0007] The financial risk analysis method according to the first aspect of this application includes: Acquire multimodal data to be analyzed; Based on the multimodal data to be analyzed, a set of candidate expert modules is obtained based on domain classification processing, and initial weight information corresponding to the set of candidate expert modules is obtained based on a dynamic strategy. The multimodal data to be analyzed is subjected to feature extraction and cross-modal fusion processing to obtain semantic feature information; Based on the semantic feature information, the candidate expert module set, and the initial weight information, obtain the activated expert module set and the target weight information corresponding to the activated expert module set; Based on the activated expert module set, the semantic feature information is processed to obtain expert processing result set information; Risk analysis information is obtained based on the expert processing result set information and the target weight information.

[0008] The financial risk analysis method according to the embodiments of this application has at least the following beneficial effects: Based on the acquired multimodal data to be analyzed, expert modules are screened according to their domain analysis and intent mapping to obtain a candidate expert module set. Furthermore, based on a dynamic strategy, the initial weight information of the candidate expert module set is determined. This achieves rapid screening of expert modules in the first stage, quickly narrowing the scope from a large expert database to obtain a candidate expert module set highly relevant to the current input. Semantic feature information is obtained by performing feature extraction and cross-modal fusion processing on the multimodal data to be analyzed. This semantic feature information is then combined with further screening of the candidate expert module set to select expert modules more suitable for the input semantics, forming an activated expert module set. Simultaneously, adjustments are made based on the initial weight information combined with the semantic feature information to obtain target weight information. This achieves precise selection of expert modules in the second stage, enabling the selection and activation of the optimal expert module based on the semantic features of different input multimodal data to be analyzed. The semantic feature information is processed according to the activated expert module set to obtain expert processing result set information. This is further processed in conjunction with the target weight information to obtain risk analysis information. Therefore, by employing a two-stage expert module selection process—namely, a rapid first-stage screening based on domain classification and intent mapping, and a precise second-stage selection process incorporating input semantics—the accuracy of expert selection and risk analysis can be improved. Furthermore, determining initial weight information based on dynamic strategies allows for rapid adaptation to the ever-changing risk strategies in the financial field, enhancing the adaptability of risk analysis and ultimately making the obtained risk analysis information more reliable.

[0009] According to one embodiment of this application, the step of obtaining a candidate expert module set based on domain classification processing according to the multimodal data to be analyzed, and obtaining initial weight information corresponding to the candidate expert module set based on a dynamic strategy, includes: The multimodal data to be analyzed is subjected to domain identification processing to obtain the domain probability distribution vector; Based on the multimodal data to be analyzed and the domain probability distribution vector, the candidate expert module set is obtained through intent mapping processing. The initial weight information is obtained based on the candidate expert module set and the dynamic strategy.

[0010] According to one embodiment of this application, obtaining the initial weight information based on the candidate expert module set and the dynamic strategy includes: Based on the multimodal data to be analyzed, the semantic weights of each expert module in the candidate expert module set are determined through semantic analysis. Based on the dynamic strategy, the strategy weight of each expert module in the candidate expert module set is determined; Based on the multimodal data to be analyzed, and based on historical behavior analysis, the historical behavior weights of each expert module in the candidate expert module set are determined. Based on the semantic weight, strategy weight, and historical behavior weight of each expert module in the candidate expert module set, the corresponding initial weight is determined, and the initial weight information is obtained.

[0011] According to one embodiment of this application, obtaining an activated expert module set and target weight information corresponding to the activated expert module set based on the semantic feature information, the candidate expert module set, and the initial weight information includes: Based on the semantic feature information and the candidate expert module set, expert modules are selected based on similarity to obtain a preliminary expert module set; Based on the initial expert module set and the initial weight information, weight adjustment processing is performed to obtain the adjusted weight information corresponding to the initial expert module set; Based on the adjusted weight information, a preset number of expert modules are selected from the initial expert module set in descending order of weight, and the activated expert module set and the corresponding target weight information are obtained.

[0012] According to one embodiment of this application, the step of selecting expert modules based on similarity according to the semantic feature information and the candidate expert module set to obtain a preliminary expert module set includes: Based on the expert database, obtain the expert feature vectors of each expert module in the candidate expert module set; Calculate the similarity between the semantic feature information and each of the expert feature vectors to obtain similarity information; Based on the similarity information, and on the condition that the similarity is greater than a preset similarity threshold, expert modules are selected from the candidate expert module set to obtain the initial expert module set.

[0013] According to one embodiment of this application, the step of performing weight adjustment processing based on the initial expert module set and the initial weight information to obtain adjusted weight information corresponding to the initial expert module set includes: Based on the initial weight information, determine the initial weights of each expert module in the initial expert module set; Based on the initial weight information, an entropy constraint term is determined, which is used to prevent excessive concentration of weights. Based on the similarity information and the entropy constraint, the initial weights of each expert module in the initial expert module set are adjusted to obtain the adjustment weight information.

[0014] According to one embodiment of this application, obtaining risk analysis information based on the expert processing result set information and the target weight information includes: Based on the expert processing result set information and the target weight information, the risk scores predicted by each expert module are weighted and calculated to obtain a comprehensive risk score. Based on the semantic feature information and the comprehensive risk score, feature risk contribution information is obtained through feature contribution analysis and processing. Based on the semantic feature information, the feature risk contribution information, and the preset explanation template, risk explanation text information and confidence information are obtained through large language model processing. The risk analysis information is obtained based on the comprehensive risk score, the characteristic risk contribution information, the risk explanation text information, and the confidence level information.

[0015] According to one embodiment of this application, the semantic feature information includes multiple semantic feature vectors; the step of obtaining feature risk contribution information based on the semantic feature information and the comprehensive risk score, according to feature contribution analysis processing, includes: Calculate the interaction strength between each semantic feature vector in the semantic feature information to obtain feature interaction strength information; Based on the feature interaction strength information, the semantic feature vectors in the semantic feature information are grouped to obtain multiple semantic feature vector groups, and each semantic feature vector group includes at least one of the semantic feature vectors. Based on the comprehensive risk score, the Shapley value corresponding to each semantic feature vector group is calculated, and the Shapley value represents the risk contribution. Based on the semantic feature vector grouping and the corresponding Shapley value, and based on the feature interaction strength information, interaction strength correction processing is performed on each semantic feature vector to obtain the feature risk contribution degree corresponding to the semantic feature vector; The feature risk contribution information is obtained based on the feature risk contribution of each of the semantic feature vectors.

[0016] According to one embodiment of this application, the step of obtaining risk explanation text information and confidence information based on the semantic feature information, the feature risk contribution information, and the preset explanation template, using large language model processing, includes: Based on the semantic feature information, the feature risk contribution information, and the preset explanation template, the preset number of large language models are called for processing to obtain a preset number of first explanation text information; Based on each of the first explanatory text information, output divergence information is obtained, and the output divergence information characterizes the output stability; Based on the output divergence information, the first explanatory text information with the smallest output divergence is taken as the risk explanatory text information; The confidence information is obtained based on the output divergence information.

[0017] According to one embodiment of this application, the multimodal data to be analyzed includes structured data and unstructured data; before obtaining a candidate expert module set based on domain classification processing according to the multimodal data to be analyzed, and obtaining initial weight information corresponding to the candidate expert module set based on a dynamic strategy, the method further includes: Based on the multimodal data to be analyzed, the structured data is dynamically encrypted, and the unstructured data is differentially private.

[0018] An electronic device according to a second aspect of this application includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described financial risk analysis method.

[0019] According to a third aspect of this application, a non-transitory computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the above-described financial risk analysis method.

[0020] A computer program product according to a fourth aspect of this application includes a computer program that, when executed by a processor, implements the above-described financial risk analysis method.

[0021] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies 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.

[0023] Figure 1 This is a flowchart illustrating the financial risk analysis method provided in this application.

[0024] Figure 2 This is a schematic diagram of the processing procedure of the financial risk analysis method provided in this application.

[0025] Figure 3 This is a schematic diagram illustrating the initial weight information obtained by the financial risk analysis method provided in this application.

[0026] Figure 4 This is a schematic diagram of the financial risk analysis method provided in this application for obtaining and activating the expert module set.

[0027] Figure 5 This is a schematic diagram illustrating the financial risk analysis method provided in this application for obtaining the contribution of characteristic risks.

[0028] Figure 6 This is a schematic diagram illustrating the financial risk analysis method provided in this application for obtaining risk interpretation text information.

[0029] Figure 7 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0030] The embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but should not be used to limit the scope of this application.

[0031] In the description of the embodiments of this application, it should be noted that the terms "first", "second" and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0032] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the embodiments of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0033] refer to Figure 1 and Figure 2 This application provides financial risk analysis methods, including: S100: Acquire multimodal data to be analyzed; S200: Based on the multimodal data to be analyzed, obtain a set of candidate expert modules based on domain classification processing, and obtain the initial weight information corresponding to the set of candidate expert modules based on a dynamic strategy; S300: Perform feature extraction and cross-modal fusion processing on the multimodal data to be analyzed to obtain semantic feature information; S400: Based on the semantic feature information, the candidate expert module set, and the initial weight information, obtain the activated expert module set and the target weight information corresponding to the activated expert module set; S500: Based on the activated expert module set, the semantic feature information is processed to obtain expert processing result set information; S600: Obtain risk analysis information based on the expert processing result set information and the target weight information.

[0034] Based on the acquired multimodal data to be analyzed, expert modules are screened according to their domain and intent mapping to obtain a candidate expert module set. An initial weight information for this candidate expert module set is determined based on a dynamic strategy. This first stage achieves rapid screening of expert modules, quickly narrowing down the scope from a large expert database to obtain a candidate expert module set highly relevant to the current input. Semantic feature information is obtained through feature extraction and cross-modal fusion processing of the multimodal data to be analyzed. This semantic feature information is then used to further screen the candidate expert module set, selecting expert modules more suitable for the input semantics to form an activated expert module set. Simultaneously, the target weight information is obtained by adjusting the initial weight information in conjunction with the semantic feature information. This second stage achieves precise selection of expert modules, enabling the selection of the optimal expert module for activation based on the semantic features of different input multimodal data to be analyzed. The semantic feature information of the activated expert module set is processed to obtain an expert processing result set. This result set is further processed in conjunction with the target weight information to obtain risk analysis information.

[0035] Therefore, by employing a two-stage expert module selection process—namely, a rapid first-stage screening based on domain classification and intent mapping, and a precise second-stage selection process incorporating input semantics—the accuracy of expert selection and risk analysis can be improved. Furthermore, determining initial weight information based on dynamic strategies allows for rapid adaptation to the ever-changing risk strategies in the financial field, enhancing the adaptability of risk analysis and ultimately making the obtained risk analysis information more reliable.

[0036] It is understood that multimodal data to be analyzed can include structured and unstructured data. Structured data includes credit data, transaction records, etc., while unstructured data includes text records, loan application texts, customer service conversation records, etc. The multimodal data to be analyzed in this application was obtained with the user's authorization, in compliance with regulations.

[0037] It should be noted that risk strategies in the financial sector, including regulatory policies and risk control standards, are constantly being updated. This application, by determining initial weight information based on a dynamic strategy, can flexibly adjust the impact of different expert modules on risk analysis, thereby quickly adapting to dynamically updated risk strategies. The dynamic strategy can be set according to the risk strategy or determined after processing by the risk strategy.

[0038] refer to Figure 3 In some embodiments of the financial risk analysis method of this application, step S200 includes: The multimodal data to be analyzed is subjected to domain identification processing to obtain the domain probability distribution vector; Based on the multimodal data to be analyzed and the domain probability distribution vector, the candidate expert module set is obtained through intent mapping processing. The initial weight information is obtained based on the candidate expert module set and the dynamic strategy.

[0039] First, coarse-grained domain identification processing is performed on the multimodal data to be analyzed to determine the domain classification of financial risk analysis, such as loan risk analysis, transaction risk analysis, and anti-fraud risk analysis, and to obtain domain probability distribution vectors reflecting the probability of belonging to different domains. Then, based on the domain probability distribution vectors, fine-grained intent mapping processing is used to perform more precise intent matching, obtain a set of candidate expert modules, and then combine a dynamic strategy to determine the initial weights of each expert module in the candidate expert module set, obtaining initial weight information. In this way, by using coarse-grained domain classification and fine-grained intent mapping, and employing a hierarchical expert routing strategy, the goal of quickly and accurately selecting expert modules in the first stage can be achieved.

[0040] In some embodiments of this application, the domain identification process can use a lightweight text classifier, such as DistilBERT, to quickly classify the input multimodal data to be analyzed and obtain the probability distribution of belonging to different domains.

[0041] In some embodiments of this application, the intent mapping process can use an encoder, such as a context-encoder, to extract the semantics of the input multimodal data to be analyzed, and combine it with a domain probability distribution vector to perform mapping in the intent space of the highest probability domain. The intent space includes multiple intent tags, and intent tags can be matched through similarity calculation. Based on the intent tag matching results, the corresponding expert modules are obtained. It should be noted that the expert database includes expert modules associated with different intents in different domains. After determining the most matching domain and intent, the relevant expert modules can be used as a candidate expert module set.

[0042] In some embodiments of this application, in addition to obtaining a set of candidate expert modules by combining domain recognition processing with intent mapping processing, a hierarchical preliminary screening can also be performed based on multimodal data to be analyzed, using methods such as decision trees or cascade classifiers, to obtain a set of candidate expert modules.

[0043] In some embodiments of this application, the process of obtaining a candidate expert module set based on domain classification processing according to the multimodal data to be analyzed, and obtaining the initial weight information corresponding to the candidate expert module set based on a dynamic strategy, can be implemented using a perceptual gating network model.

[0044] In some embodiments of the financial risk analysis method of this application, obtaining the initial weight information based on the candidate expert module set and the dynamic strategy includes: Based on the multimodal data to be analyzed, the semantic weights of each expert module in the candidate expert module set are determined through semantic analysis. Based on the dynamic strategy, the strategy weight of each expert module in the candidate expert module set is determined; Based on the multimodal data to be analyzed, and based on historical behavior analysis, the historical behavior weights of each expert module in the candidate expert module set are determined. Based on the semantic weight, strategy weight, and historical behavior weight of each expert module in the candidate expert module set, the corresponding initial weight is determined, and the initial weight information is obtained.

[0045] The initial weights of expert modules are determined from three dimensions: semantic weights derived from semantic analysis of multimodal data, historical behavior weights derived from historical behavior analysis, and policy weights determined by dynamic strategies. The dynamic strategy adjusts the policy weights to quickly adapt to continuously updated risk strategies. This ensures that the allocation of initial weights for the candidate expert module set considers both the semantics of the input data and the patterns of user historical behavior, while also responding in real-time to changes in risk strategies. This facilitates the accurate provision of benchmarks for the importance of different expert modules in risk analysis.

[0046] It should be noted that, typically, risk strategies, such as updates to regulatory policies and risk control standards, require long update cycles for risk analysis models or rely on manual adjustments. However, this application, based on a dynamic strategy, can quickly respond to updates to risk strategies and adjust strategy weights accordingly to adapt to changes in risk strategies.

[0047] In some embodiments of this application, the latest risk strategy text can be parsed using sequence models such as the Transformer encoder to generate a risk sensitivity vector. The policy sensitivity vector can quantify the numerical representation of the degree of response of each expert module to the current risk strategy change, and then adjust the dynamic strategy according to the risk sensitivity vector.

[0048] In some embodiments of this application, the initial weights can be determined by the following expression: in, These are the initial weights; For semantic weights, For semantic gating functions, This is multimodal data to be analyzed; For strategy weights, For policy gating functions, Information processed and encoded for risk strategies; Weighting based on historical behavior, For historical behavior gating functions, This is multimodal data to be analyzed; , , The coefficients can be determined through training and learning, representing the relative importance of the three gate functions; It is a normalized exponential function.

[0049] refer to Figure 4 In some embodiments of the financial risk analysis method of this application, step S400 includes: Based on the semantic feature information and the candidate expert module set, expert modules are selected based on similarity to obtain a preliminary expert module set; Based on the initial expert module set and the initial weight information, weight adjustment processing is performed to obtain the adjusted weight information corresponding to the initial expert module set; Based on the adjusted weight information, a preset number of expert modules are selected from the initial expert module set in descending order of weight, and the activated expert module set and the corresponding target weight information are obtained.

[0050] First, based on semantic feature information and a candidate expert module set, a preliminary selection is made based on similarity, resulting in an initial expert module set. Then, combined with the initial weight information, weight adjustment processing is performed to obtain adjusted weight information. Finally, a predetermined number of expert modules are selected for activation from largest to smallest based on the adjusted weight information, and target weight information is determined. In this way, by calculating the similarity between semantics and expert modules, candidate expert modules are further filtered. Combined with weight adjustment processing, a predetermined number of expert modules are finally selected for activation based on the weights, forming an activated expert module set. This completes the second stage of precise expert module selection, achieving deep fusion of prior initial weight information and semantic features, which helps to make the selection of expert modules more accurate.

[0051] It should be noted that this application introduces a two-step screening mechanism of similarity calculation and weight adjustment to ensure that the activated expert module has the highest matching degree with the risk analysis task corresponding to the current input multimodal data to be processed.

[0052] Weight adjustment is a process of modifying the initial weights by incorporating new constraints or metrics, such as the similarity between expert modules and semantic features.

[0053] In some embodiments of the financial risk analysis method of this application, the step of selecting expert modules based on similarity according to the semantic feature information and the candidate expert module set to obtain a preliminary expert module set includes: Based on the expert database, obtain the expert feature vectors of each expert module in the candidate expert module set; Calculate the similarity between the semantic feature information and each of the expert feature vectors to obtain similarity information; Based on the similarity information, and on the condition that the similarity is greater than a preset similarity threshold, expert modules are selected from the candidate expert module set to obtain the initial expert module set.

[0054] Based on the expert database, expert feature vectors for each candidate expert module are obtained. Then, the similarity between semantic feature information and each expert feature vector is calculated. A preliminary set of expert modules is selected based on similarity scores exceeding a preset threshold. In this way, a quantitative indicator measures the degree of fit between the semantic features of the input data and the professional capabilities of the expert modules, effectively filtering out expert modules with low similarity.

[0055] In some embodiments of this application, the acquisition of the initial expert module set can be represented by the following expression: in, This is a preliminary selection of expert modules; A set of candidate expert modules; This is a similarity calculation function; This is the i-th expert module; The feature vector is a semantic feature information; Let be the expert feature vector of the i-th expert module; This is a preset similarity threshold.

[0056] It should be noted that the expert feature vector is an encoded representation of the data type or task type that the expert module is good at handling. During the pre-training phase, the parameters of a specific network layer of the expert module or the average activation state when processing typical samples can be extracted as the expert feature vector. In some embodiments, the expert module can be constructed using lightweight fine-tuning techniques such as Low-Rank Adaptation (LoRA) to reduce storage and computational overhead.

[0057] In some embodiments of this application, similarity information can be obtained by calculating cosine similarity or Euclidean distance, etc.

[0058] In some embodiments of the financial risk analysis method of this application, the step of performing weight adjustment processing based on the initial expert module set and the initial weight information to obtain adjusted weight information corresponding to the initial expert module set includes: Based on the initial weight information, determine the initial weights of each expert module in the initial expert module set; Based on the initial weight information, an entropy constraint term is determined, which is used to prevent excessive concentration of weights. Based on the similarity information and the entropy constraint, the initial weights of each expert module in the initial expert module set are adjusted to obtain the adjustment weight information.

[0059] Initial weights and entropy constraints are determined based on the initial weight information. These are then combined with similarity information and entropy constraints to prevent excessive weight concentration, adjusting the initial weights of the initially selected expert modules. This adaptive adjustment of expert module weights based on semantic feature similarity avoids weight concentration in a few modules, thus improving the accuracy of risk analysis.

[0060] In some embodiments of this application, the acquisition of adjustment weight information can be represented by the following expression: in, This refers to the adjusted weight distribution, i.e., the adjusted weight information; This represents the initial weight distribution; The similarity distribution between expert feature vectors and semantic features, i.e., similarity information; This is the element-wise multiplication operator; For entropy constraint terms; This is the adjustment coefficient.

[0061] The entropy constraint term refers to the introduction of information entropy theory as a regulating factor to measure and control the uniformity of the weight distribution. Specifically, the Shannon entropy of the initial weight distribution can be calculated and added as a regularization term to the weight update calculation formula.

[0062] In some embodiments of the financial risk analysis method of this application, S600 includes: Based on the expert processing result set information and the target weight information, the risk scores predicted by each expert module are weighted and calculated to obtain a comprehensive risk score. Based on the semantic feature information and the comprehensive risk score, feature risk contribution information is obtained through feature contribution analysis and processing. Based on the semantic feature information, the feature risk contribution information, and the preset explanation template, risk explanation text information and confidence information are obtained through large language model processing. The risk analysis information is obtained based on the comprehensive risk score, the characteristic risk contribution information, the risk explanation text information, and the confidence level information.

[0063] The risk scores predicted by each activated expert module are weighted based on target weights to obtain a comprehensive risk score. Then, based on the comprehensive risk score and semantic feature information, feature contribution analysis is used to obtain feature risk contribution information. Using a pre-defined explanation template, a large language model is used to process the risk explanation text and confidence level information. Finally, based on the comprehensive risk score, feature risk contribution information, risk explanation text, and confidence level information, risk analysis information is obtained. This not only outputs a final quantified comprehensive risk score but also reveals the contribution of different features to risk, while generating detailed, structured explanation text with confidence assessments. This enables a detailed analysis and assessment of financial risk, enhancing the interpretability of financial risk analysis.

[0064] refer to Figure 5 In some embodiments of the financial risk analysis method of this application, the semantic feature information includes multiple semantic feature vectors; the step of obtaining feature risk contribution information based on the semantic feature information and the comprehensive risk score, according to feature contribution analysis processing, includes: Calculate the interaction strength between each semantic feature vector in the semantic feature information to obtain feature interaction strength information; Based on the feature interaction strength information, the semantic feature vectors in the semantic feature information are grouped to obtain multiple semantic feature vector groups, and each semantic feature vector group includes at least one of the semantic feature vectors. Based on the comprehensive risk score, the Shapley value corresponding to each semantic feature vector group is calculated, and the Shapley value represents the risk contribution. Based on the semantic feature vector grouping and the corresponding Shapley value, and based on the feature interaction strength information, interaction strength correction processing is performed on each semantic feature vector to obtain the feature risk contribution degree corresponding to the semantic feature vector; The feature risk contribution information is obtained based on the feature risk contribution of each of the semantic feature vectors.

[0065] By calculating the interaction strength between semantic feature vectors and then grouping them, the Shapley value of each semantic feature vector group is calculated based on the comprehensive risk score. Finally, interaction strength correction processing is performed on the semantic feature vectors based on the interaction strength information to obtain the feature risk contribution. In this way, grouping reduces the complexity of calculating the Shapley value and effectively improves computational efficiency. Combined with interaction strength correction processing, the individual feature risk contribution of each semantic feature vector can be obtained, improving computational efficiency while maintaining accuracy.

[0066] It should be noted that the Shapley value is determined by the Shapley algorithm. Since the computational complexity of the Shapley algorithm increases exponentially with the number of features, the number of features cannot exceed 20 in practical applications. In this application, by grouping semantic feature vectors and calculating the Shapley value for each group, the computational efficiency of the Shapley algorithm can be effectively improved. For example, if there are 5 pairs of strongly interacting semantic feature vectors, dividing the 10 semantic feature vectors into 5 groups reduces the computational complexity of the Shapley algorithm from 2^10 = 1024 to 2^5 = 32, improving computational efficiency by 32 times. It should be further noted that the feature risk contribution is calculated on a grouped basis using semantic feature vectors. The feature risk contribution may correspond to two or more semantic feature vectors. Through interaction strength correction processing, the feature risk contribution generated by the combined effect of the semantic feature vector groups is allocated according to the interaction strength to determine the individual feature risk contribution of each semantic feature component.

[0067] In some embodiments of this application, the feature risk contribution of the semantic feature vector can be determined by the following expression: in, Contribution to characteristic risk; The Shapley value; This is the interaction strength correction coefficient; semantic feature vector semantic feature vectors that have grouping relationships The intensity of interaction between them; Represents the semantic feature vector Semantic feature vectors with grouping relationships Sum of the interaction strengths between them; Let be a probability function.

[0068] In some embodiments of this application, different semantic feature vectors can characterize items such as load rate, revenue, and abnormal IP address. Based on the corresponding feature risk contribution, the contribution of the corresponding item to the risk can be determined, such as the risk contribution of the debt ratio being 0.28.

[0069] refer to Figure 6 In some embodiments of the financial risk analysis method of this application, the step of obtaining risk explanation text information and confidence information based on the semantic feature information, the feature risk contribution information, and the preset explanation template, using large language model processing, includes: Based on the semantic feature information, the feature risk contribution information, and the preset explanation template, the preset number of large language models are called for processing to obtain a preset number of first explanation text information; Based on each of the first explanatory text information, output divergence information is obtained, and the output divergence information characterizes the output stability; Based on the output divergence information, the first explanatory text information with the smallest output divergence is taken as the risk explanatory text information; The confidence information is obtained based on the output divergence information.

[0070] Because large language models suffer from output randomness, they struggle to provide stable risk interpretations, thus failing to meet financial regulatory requirements. Furthermore, the randomness of large language models can cause fluctuations in the risk contribution of the same input feature across multiple inferences, invalidating the mathematical guarantee of the Shapley value. To address this, this application first utilizes a pre-defined interpretation template to improve the stability of the large language model's output. Simultaneously, by repeatedly calling the large language model, multiple first interpretation texts are obtained, and output divergence information is calculated to characterize the output stability of the large language model when semantic feature information and feature risk contribution information are used as inputs. The first interpretation text with the smallest output divergence is selected as the final risk interpretation text, and confidence information is obtained based on the output divergence information. In this way, through multiple outputs, the first interpretation text with the smallest output divergence, i.e., the most stable, is selected as the risk interpretation text, mitigating the randomness problem of the text generated by the large language model. Simultaneously, the confidence information reflects the uncertainty of the output, thereby indirectly reflecting the reliability of the feature risk contribution information.

[0071] It is understandable that the smaller the output divergence, the higher the stability and the greater the confidence, and vice versa.

[0072] In some embodiments of this application, the preset number of iterations can be 10, and the preset number of outputs is also 10. That is, the large language model is called 10 times, and semantic feature information, feature risk contribution information, and a preset explanation template are input for processing to obtain 10 output first explanation text information. The output divergence information can be obtained by calculating KL divergence, or it can be determined by calculating cosine distance, etc.

[0073] In some embodiments of this application, during the model training phase, the large language model may employ a comprehensive loss function to optimize the accuracy of attribution explanation, output stability, and output normalization. The comprehensive loss function can be expressed by the following expression: in, Loss function; Calculate the loss for the Shapley value to optimize the accuracy of the explanation of attribution; This is used to optimize output stability by reducing output stability. The loss due to policy compliance is used to optimize the standardization of outputs; , , The balancing coefficient is adjusted through model training.

[0074] refer to Figure 2 In some embodiments of the financial risk analysis method of this application, the multimodal data to be analyzed includes structured data and unstructured data; before S200, the method further includes: Based on the multimodal data to be analyzed, the structured data is dynamically encrypted, and the unstructured data is differentially private.

[0075] After acquiring multimodal data for analysis, structured data is first dynamically encrypted, while unstructured data undergoes differential privacy processing. This approach maximizes the protection of sensitive information in the original data while ensuring the multimodal data can be used for feature extraction and analysis. This enhances security, prevents data leaks, meets privacy requirements in cross-institutional, multi-party collaborative computing scenarios, and breaks down data silos.

[0076] In some embodiments of this application, structured data can be encrypted using CKKS homomorphic encryption, and unstructured text can be processed using differential privacy. Furthermore, the text data can be cleaned, segmented, and pre-encoded using the FinBERT-base model.

[0077] In some embodiments of the financial risk analysis method of this application, the step of performing feature extraction processing on the multimodal data to be analyzed to obtain semantic feature information includes: The structured data and unstructured data in the multimodal data to be analyzed are jointly encoded to obtain the semantic feature information; The joint encoding process performs cross-modal data consistency verification on the structured data and the unstructured data based on a cross-attention mechanism to generate a fused multi-scale semantic representation as the semantic feature information; The multi-scale semantic representation includes local representations for characterizing fine-grained semantics and global representations for characterizing the overall context.

[0078] To more intuitively understand the financial risk analysis method of this application, an illustrative example is provided: Example 1: In a credit risk assessment scenario, analyze and evaluate the user's credit risk: The multimodal data to be analyzed includes: Structured data includes: basic customer information (age, gender, occupation, education level, marital status, etc.), credit records (credit score, number of overdue payments, debt ratio, credit card usage, etc.), and transaction history (income, expenses, transfer frequency, number of large transactions, etc. in the past 12 months). Unstructured data: Loan application text (average 500 words), income certificate text (average 800 words), other text assets (average 300 words); Processing flow: Structured data is encrypted using the CKKS homomorphic encryption scheme, while unstructured data is processed using differential privacy; text data is cleaned, segmented, and initially encoded using the FinBERT-base model. Coarse-grained domain classification: Features were extracted using DistilBERT, and the probability distribution of the five domains was output. The "credit assessment" domain had the highest probability (0.72). Fine-grained intent mapping: The IRE encoder extracts intent representations, matches the intent "small consumer loan default risk assessment" (confidence level 0.89), and obtains a set of candidate expert modules (8 expert modules). Weight adjustment based on dynamic strategy: RTE analyzes the latest risk strategy and determines the initial weight distribution W_prior; Feature extraction and fusion: The context encoder (FinBERT deep layer) jointly encodes unstructured and structured data, outputting 512-dimensional semantic feature information; Similarity calculation: Calculate the similarity between semantic features and expert feature vectors to obtain similarity information; Adjusting weights: Based on the initial weight information and similarity information, weights are adjusted to obtain adjusted weight information; Activate expert modules: Based on the adjustment weight information, activate 6 expert modules and determine the target weight information. The 6 expert modules perform parallel inference, output their respective processing results, and obtain the expert processing result set information. Calculate the comprehensive risk score: By integrating the processing results of 6 experts and combining the target weight information, a weighted average is obtained to obtain the comprehensive risk score (0.73, high risk). Determine the contribution of characteristic risks: Calculate the contribution of characteristic risks. The top-3 characteristics are: debt ratio (contribution 0.28), credit card usage rate (0.22), and the statement of fund usage in the application document (0.18). Generate risk explanation text information: The large language model is called multiple times to output the risk explanation text information and execution degree information based on the output divergence; Risk analysis information is obtained based on comprehensive risk scores, characteristic risk contribution information, risk explanation text information, and confidence level information.

[0079] Example 2: In an anti-fraud detection scenario, real-time transaction anti-fraud detection is performed to quickly identify suspicious transactions and intercept fraudulent activities. The multimodal data to be analyzed includes: Structured data: Transaction information (15 fields), Device information (12 fields), Account information (18 fields); Unstructured data: transaction notes (average 50 words), customer service dialogue text (average 2000 words), public opinion information text (average 300 words); Processing flow: Real-time transaction data is collected from the trading system (latency <50ms), and structured data is encrypted using CKKS; unstructured data undergoes differential privacy processing. Coarse-grained domain classification: Using DistilBERT to extract features, the "fraud detection" domain has the highest probability (0.89). Fine-grained intent mapping: The IRE encoder extracts intent representations, matches the intent of "large-scale cross-border transaction fraud" (confidence level 0.94), and obtains a set of candidate expert modules (6 anti-fraud expert modules). Weight adjustment based on dynamic strategy: RTE analyzes the latest fraud strategy and determines the initial weight distribution W_prior; Feature extraction and fusion: The context encoder (FinBERT deep layer) jointly encodes unstructured and structured data to output semantic feature information; Similarity calculation: Calculate the similarity between semantic features and expert feature vectors to obtain similarity information; Adjusting weights: Based on the initial weight information and similarity information, weights are adjusted to obtain adjusted weight information; Activate expert modules: Based on the adjustment weight information, activate 4 anti-fraud expert modules and determine the target weight information. The 4 anti-fraud expert modules perform parallel inference, output their respective processing results, and obtain the expert processing result set information. Calculate the comprehensive risk score: By integrating the processing results of 6 experts and combining the target weight information, a weighted average is used to obtain the comprehensive risk score (fraud probability 0.87, high risk). Determine the contribution of feature risks: Calculate the contribution of feature risks. The top-5 features are: abnormal transaction amount (contribution 0.31), abnormal IP address (0.24), abnormal device fingerprint (0.18), ambiguous transaction remarks (0.15), and sudden increase in account balance (0.12). Generate risk explanation text information: The large language model is called multiple times to output the risk explanation text information and execution degree information based on the output divergence; Risk analysis information is obtained based on comprehensive risk scores, characteristic risk contribution information, risk explanation text information, and confidence level information.

[0080] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840. The processor 810, communications interface 820, and memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions stored in the memory 830 to execute the aforementioned financial risk analysis method.

[0081] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to related technologies, or a portion 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 methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0082] On the other hand, this application discloses a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can execute the financial risk analysis method provided by the above-described method embodiments.

[0083] In another aspect, embodiments of this application also provide a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the financial risk analysis methods provided in the above embodiments.

[0084] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate this application and are not intended to limit this application. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that various combinations, modifications, or equivalent substitutions of the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application and should be covered within the scope of the claims of this application.

Claims

1. A financial risk analysis method, characterized in that, include: Acquire multimodal data to be analyzed; Based on the multimodal data to be analyzed, a set of candidate expert modules is obtained based on domain classification processing, and initial weight information corresponding to the set of candidate expert modules is obtained based on a dynamic strategy. The multimodal data to be analyzed is subjected to feature extraction and cross-modal fusion processing to obtain semantic feature information; Based on the semantic feature information, the candidate expert module set, and the initial weight information, obtain the activated expert module set and the target weight information corresponding to the activated expert module set; Based on the activated expert module set, the semantic feature information is processed to obtain expert processing result set information; Risk analysis information is obtained based on the expert processing result set information and the target weight information.

2. The financial risk analysis method according to claim 1, characterized in that, The step of obtaining a candidate expert module set based on domain classification processing according to the multimodal data to be analyzed, and obtaining initial weight information corresponding to the candidate expert module set based on a dynamic strategy, includes: The multimodal data to be analyzed is subjected to domain identification processing to obtain the domain probability distribution vector; Based on the multimodal data to be analyzed and the domain probability distribution vector, the candidate expert module set is obtained through intent mapping processing. The initial weight information is obtained based on the candidate expert module set and the dynamic strategy.

3. The financial risk analysis method according to claim 2, characterized in that, The step of obtaining the initial weight information based on the candidate expert module set and the dynamic strategy includes: Based on the multimodal data to be analyzed, the semantic weights of each expert module in the candidate expert module set are determined through semantic analysis. Based on the dynamic strategy, the strategy weight of each expert module in the candidate expert module set is determined; Based on the multimodal data to be analyzed, and based on historical behavior analysis, the historical behavior weights of each expert module in the candidate expert module set are determined. Based on the semantic weight, strategy weight, and historical behavior weight of each expert module in the candidate expert module set, the corresponding initial weight is determined, and the initial weight information is obtained.

4. The financial risk analysis method according to claim 1, characterized in that, The step of obtaining the activated expert module set and the target weight information corresponding to the activated expert module set based on the semantic feature information, the candidate expert module set, and the initial weight information includes: Based on the semantic feature information and the candidate expert module set, expert modules are selected based on similarity to obtain a preliminary expert module set; Based on the initial expert module set and the initial weight information, weight adjustment processing is performed to obtain the adjusted weight information corresponding to the initial expert module set; Based on the adjusted weight information, a preset number of expert modules are selected from the initial expert module set in descending order of weight, and the activated expert module set and the corresponding target weight information are obtained.

5. The financial risk analysis method according to claim 4, characterized in that, The step of selecting expert modules based on similarity according to the semantic feature information and the candidate expert module set to obtain a preliminary expert module set includes: Based on the expert database, obtain the expert feature vectors of each expert module in the candidate expert module set; Calculate the similarity between the semantic feature information and each of the expert feature vectors to obtain similarity information; Based on the similarity information, and on the condition that the similarity is greater than a preset similarity threshold, expert modules are selected from the candidate expert module set to obtain the initial expert module set.

6. The financial risk analysis method according to claim 5, characterized in that, The step of performing weight adjustment processing based on the initial expert module set and the initial weight information to obtain adjusted weight information corresponding to the initial expert module set includes: Based on the initial weight information, determine the initial weights of each expert module in the initial expert module set; Based on the initial weight information, an entropy constraint term is determined, which is used to prevent excessive concentration of weights. Based on the similarity information and the entropy constraint, the initial weights of each expert module in the initial expert module set are adjusted to obtain the adjustment weight information.

7. The financial risk analysis method according to claim 1, characterized in that, The step of obtaining risk analysis information based on the expert processing result set information and the target weight information includes: Based on the expert processing result set information and the target weight information, the risk scores predicted by each expert module are weighted and calculated to obtain a comprehensive risk score. Based on the semantic feature information and the comprehensive risk score, feature risk contribution information is obtained through feature contribution analysis and processing. Based on the semantic feature information, the feature risk contribution information, and the preset explanation template, risk explanation text information and confidence information are obtained through large language model processing. The risk analysis information is obtained based on the comprehensive risk score, the characteristic risk contribution information, the risk explanation text information, and the confidence level information.

8. The financial risk analysis method according to claim 7, characterized in that, The semantic feature information includes multiple semantic feature vectors; The step of obtaining feature risk contribution information based on the semantic feature information and the comprehensive risk score, and through feature contribution analysis, includes: Calculate the interaction strength between each semantic feature vector in the semantic feature information to obtain feature interaction strength information; Based on the feature interaction strength information, the semantic feature vectors in the semantic feature information are grouped to obtain multiple semantic feature vector groups, and each semantic feature vector group includes at least one of the semantic feature vectors. Based on the comprehensive risk score, the Shapley value corresponding to each semantic feature vector group is calculated, and the Shapley value represents the risk contribution. Based on the semantic feature vector grouping and the corresponding Shapley value, and based on the feature interaction strength information, interaction strength correction processing is performed on each semantic feature vector to obtain the feature risk contribution degree corresponding to the semantic feature vector; The feature risk contribution information is obtained based on the feature risk contribution of each of the semantic feature vectors.

9. The financial risk analysis method according to claim 7, characterized in that, The step of obtaining risk explanation text information and confidence information based on the semantic feature information, the feature risk contribution information, and the preset explanation template, using a large language model, includes: Based on the semantic feature information, the feature risk contribution information, and the preset explanation template, the preset number of large language models are called for processing to obtain a preset number of first explanation text information; Based on each of the first explanatory text information, output divergence information is obtained, and the output divergence information characterizes the output stability; Based on the output divergence information, the first explanatory text information with the smallest output divergence is taken as the risk explanatory text information; The confidence information is obtained based on the output divergence information.

10. The financial risk analysis method according to any one of claims 1 to 9, characterized in that, The multimodal data to be analyzed includes structured data and unstructured data; before obtaining a candidate expert module set based on domain classification processing according to the multimodal data to be analyzed, and obtaining the initial weight information corresponding to the candidate expert module set based on a dynamic strategy, the method further includes: Based on the multimodal data to be analyzed, the structured data is dynamically encrypted, and the unstructured data is differentially private.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the financial risk analysis method as described in any one of claims 1 to 10.

12. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the financial risk analysis method as described in any one of claims 1 to 10.

13. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the financial risk analysis method as described in any one of claims 1 to 10.

Images