A financial risk identification method, system and device based on multi-source heterogeneous data

Abstract

This invention discloses a method, system, and device for identifying financial risks based on multi-source heterogeneous data, belonging to the field of data processing technology. The method includes: collecting multi-source heterogeneous data and preprocessing it to generate a standardized feature matrix; calculating information entropy for each dimension of features and converting it into a difference degree as initial objective weights; calculating dynamically adjusted feature weights based on historical audit accuracy to generate multi-dimensional behavioral feature vectors; inputting the multi-dimensional behavioral feature vectors into a pre-trained ensemble learning model to output the probability distribution value of corporate financial risk; comparing the probability distribution value with an adaptive risk threshold determined based on historical labeled samples and constrained by a preset false alarm rate; if the threshold is exceeded, it is judged as high risk and a warning instruction is generated and sent to the audit terminal; responding to the verification feedback data returned by the audit terminal, updating the base classifier parameters and the adaptive risk threshold. This improves the accuracy, interpretability, and adaptability of financial risk identification.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a method, system, and device for identifying financial risks based on multi-source heterogeneous data. Background Technology

[0002] Financial risk identification is a crucial topic in the fields of auditing and financial risk prevention. Traditional financial risk identification primarily relies on manual auditing and analysis of single financial indicators, such as abnormal gross profit margins and changes in accounts receivable turnover. However, this method suffers from low efficiency, narrow coverage, and strong subjectivity, making it difficult to address increasingly complex and covert financial fraud methods.

[0003] With the development of data analysis technology, methods for automated risk identification based on financial data have emerged. These methods typically involve extracting financial indicator information, employing ensemble learning models such as random forests, adaptive reinforcement learning, and guided clustering to predict the risk of financial fraud, outputting a high-risk dataset of financial fraud, and generating a risk analysis report.

[0004] The existing technology still has the following shortcomings: First, the data source is singular, mainly relying on the financial statement data disclosed by the company itself, lacking cross-validation data from independent sources such as the supply chain, making it difficult to effectively identify risks of fictitious transactions; second, the determination of feature weights mostly relies on automatic calculation by the model, lacking an interpretable weight adjustment mechanism; third, the risk threshold is a fixed preset value and cannot be dynamically adjusted. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method, system, and device for identifying financial risks based on multi-source heterogeneous data.

[0006] In a first aspect, the present invention provides a financial risk identification method based on multi-source heterogeneous data, applied to a distributed data processing system, comprising the following steps: S1. Collect multi-source heterogeneous data, including at least time-series financial statement data, supply chain flow data, and unstructured text data of public opinion; and preprocess the collected data to generate a standardized feature matrix; S2. For each dimension of the standardized feature matrix, calculate the information entropy and convert it to obtain the difference degree, and use the difference degree as the initial objective weight; use the historical audit accuracy rate obtained from historical audit feedback statistics as the confidence factor, weight the initial objective weight and then re-normalize it to obtain the dynamically adjusted feature weight, and then generate a multi-dimensional behavioral feature vector. S3. Input the multi-dimensional behavioral feature vectors into the pre-trained ensemble learning model and output the probability distribution value of corporate financial risk. S4. Compare the probability distribution value with the adaptive risk threshold determined based on historical labeled samples and constrained by a preset false alarm rate. If the value exceeds the adaptive risk threshold, it is determined to be high risk and an early warning instruction is generated and sent to the audit terminal. S5. In response to the verification feedback data returned by the audit terminal, extract the true positive samples and false positive samples, and use an incremental learning algorithm with an experience replay buffer to update the parameters of the base classifier in the ensemble learning model and the adaptive risk threshold.

[0007] As a further limitation of the technical solution of the present invention, in S1, the step of preprocessing the collected data to generate a standardized feature matrix includes: S11. Process the time-series data of financial statements to extract numerical financial features, process the supply chain flow data to extract supply chain features, and perform word vector encoding on the unstructured text data of public opinion to obtain text features. S12. A cross-modal alignment method using a scaled dot product attention mechanism projects text features, numerical financial features, and supply chain features onto the same dimensional space, and performs missing value imputation and minification. Max normalization generates a standardized feature matrix; Among them, supply chain features are cross-verified features from third parties with independent sources of enterprise transactions, used to identify fictitious transactions or related fraudulent activities; text features are used to capture external regulatory and market perception signals of enterprise risks.

[0008] By collecting three types of multi-source heterogeneous data—time-series financial statement data, supply chain flow data, and unstructured text data from public opinion—the supply chain features, as cross-validation features from independent sources of enterprise transactions, can effectively identify fictitious transactions or related-party fraud. Text features are used to capture external regulatory and market perceptions of enterprise risk. This approach overcomes the limitations of information silos and improves the ability to identify hidden financial risks through multi-source data cross-validation.

[0009] As a further limitation of the technical solution of the present invention, step S12 specifically includes: S121. Perform linear projections on the text features, numerical financial features, and supply chain features respectively, mapping them to the embedding space of the same dimension d to obtain the corresponding feature vectors. S122. Using the numerical financial feature vector as the query vector Q, and the text feature vector as the key vector K and value vector V, calculate the first attention output of the financial features and text features through scaling dot product attention. S123. Using the numerical financial feature vector as the query vector Q, and the supply chain feature vector as the key vector K and value vector V, calculate the second attention output of the financial feature and the supply chain feature through scaling dot product attention. S124. Element-wise add the first attention output and the second attention output, and perform a residual concatenation with the numerical financial feature vector. Then, perform layer normalization to obtain the fused feature matrix; wherein, the attention calculation formula is:

[0010] in, The dimension of the key vector; S125. Perform missing value imputation on the fused feature matrix and use Min-Max normalization to linearly map the feature values ​​of each dimension to the [0,1] interval to generate a standardized feature matrix.

[0011] Using numerical financial feature vectors as query vectors Q, and textual feature vectors and supply chain feature vectors as key vectors K and value vectors V respectively, residual connections and layer normalization are performed after calculating attention outputs twice. This mechanism effectively solves the technical problem of the difficulty in fusing three types of heterogeneous data, enabling features from different modalities to be jointly analyzed in a unified feature space.

[0012] As a further limitation of the technical solution of the present invention, step S2 includes: S21. For the j-th dimension feature in the standardized feature matrix, calculate its information entropy. ;

[0013] In the formula, n is the sample size. For the first The sample at the th The probability of taking a value on each feature; S22. Calculate the degree of difference based on information entropy: The difference is taken as the first The initial objective weight of the first feature; based on historical audit feedback statistics, obtain the first feature's initial objective weight. Historical audit accuracy corresponding to each feature , and use it as a confidence factor; S23, regarding the first The weighted difference is obtained by weighting the initial objective weights of each feature: ; S24. Renormalize the weighted variance of all features to obtain the dynamically adjusted feature weights:

[0014] Where m is the total number of dimensions of the features; S25. Multiply the eigenvalues ​​of each dimension in the standardized feature matrix by the corresponding feature weights to generate a multi-dimensional behavioral feature vector.

[0015] For each dimension of the standardized feature matrix, information entropy is calculated and converted into a degree of difference as the initial objective weight. Simultaneously, historical audit accuracy, obtained from historical audit feedback statistics, is introduced as a confidence factor. The initial objective weights are then weighted and re-normalized to obtain dynamically adjusted feature weights. This mechanism incorporates professional credibility in the auditing field into the objective calculation of feature weights, making the weight determination process more interpretable and auditable.

[0016] As a further limitation of the technical solution of the present invention, S3 specifically includes: S31. Pre-build an ensemble learning model, wherein the ensemble learning model includes at least two base classifiers: gradient boosting tree and support vector machine; S32. Input the multi-dimensional behavioral feature vector into each base classifier to obtain the risk probability prediction value output by each base classifier. S33. Weighted fusion of the risk probability predictions output by each base classifier:

[0017] Where T is the number of base classifiers. Let the risk probability be the output of the t-th base classifier. The fusion weights of the t-th base classifier; S34. Output the fused risk probability distribution value As a probability distribution value of corporate financial risk.

[0018] The ensemble learning model includes two base classifiers: gradient boosting trees and support vector machines. These are weighted and fused to output a risk probability distribution. The two base classifiers are complementary in their decision-making principles; by combining their advantages through weighted fusion, the accuracy and stability of risk probability prediction are improved.

[0019] As a further limitation of the technical solution of the present invention, step S4 includes: S41. Obtain the historical labeled sample set, which includes positive samples that have been identified as risk and negative samples that have been identified as normal. S42. Input the historical labeled samples into the trained ensemble learning model to obtain the risk probability prediction value of each historical labeled sample; S43. Based on the predicted risk probability values ​​and real labels of each historical labeled sample, determine the adaptive risk threshold under the preset false alarm rate constraint. S44. Obtain the probability distribution value of the enterprise financial risk output by S3, and compare it with the adaptive risk threshold: If the probability distribution value is greater than or equal to the adaptive risk threshold, it is determined to be high risk; If the probability distribution value is less than the adaptive risk threshold, it is determined to be low risk; S45. When a high-risk condition is identified, an early warning instruction is generated and sent to the audit terminal.

[0020] As a further limitation of the technical solution of the present invention, S43 specifically includes: S431. Sort the historical labeled samples from high to low according to their predicted risk probability values; S432. Using the predicted risk probability value of each sample as a candidate threshold, for each candidate threshold, samples with a predicted risk probability value greater than or equal to the threshold are classified as risky, and samples with a predicted probability value less than the threshold are classified as normal; then, based on the true label of the sample, calculate the false positive rate and true positive rate corresponding to each candidate threshold; where: False positive rate = Number of samples judged as risky but with a true label of normal / Total number of samples with a true label of normal; True positive rate = Number of samples that were judged as risk and were actually labeled as risk / Total number of samples that were actually labeled as risk.

[0021] S433, Set the upper limit of false alarm rate constraint ; S434, In all cases where the false positive rate is ≤ Among the candidate thresholds, the candidate threshold that maximizes the true positive rate is selected as the adaptive risk threshold.

[0022] An adaptive risk threshold is determined based on historical labeled samples and a preset false positive rate constraint. The probability distribution value output by the model is compared with the adaptive threshold to determine the risk level. Compared with the existing technology that uses a fixed preset threshold, this method can dynamically adjust the threshold according to the different tolerance levels of the false positive rate in the audit scenario. It maximizes the true positive rate while ensuring that the false positive rate does not exceed the preset upper limit, effectively balancing the accuracy and coverage of the warning and reducing the waste of audit resources.

[0023] As a further limitation of the technical solution of the present invention, in S5, in response to the verification feedback data returned by the audit terminal, true positive samples and false positive samples are extracted, and the parameters of the base classifier in the ensemble learning model and the adaptive risk threshold are updated using an incremental learning algorithm with an experience replay buffer, specifically including: S51. Establish an experience playback buffer and use a first-in-first-out queue with a set maximum capacity to store historical training samples and verification feedback samples returned by the audit terminal. S52. In response to the verification feedback data returned by the audit terminal, extract true positive samples and false positive samples from them and add them to the experience playback buffer; wherein, the true positive samples are samples that the model warned of and that have been verified by the audit to have risks, and the false positive samples are samples that the model warned of but that have been verified by the audit to have no risks; S53. Randomly sample from the experience replay buffer to construct incremental training batches; S54. Update the parameters according to the type of the base classifier in the ensemble learning model: If it is a gradient boosting tree model, then the new weak classifier is fitted starting from the residual of the previous model. If it is a support vector machine model, the hyperplane is updated using an incremental support vector machine algorithm based on support vector caching; S55. Based on the prediction results of the updated base classifiers on the validation set or newly added verification feedback samples, recalculate the fusion weights of each base classifier. The updated model output is obtained through weighted fusion. S56. Based on the updated model output and existing labeled samples, reselect the adaptive risk threshold according to the preset false alarm rate constraint to complete the update.

[0024] By storing historical samples and newly added feedback samples in an experience replay buffer, and simultaneously sampling both historical and new samples during incremental updates, the catastrophic forgetting problem is effectively overcome. For the gradient boosting tree model, new weak classifiers are fitted using the residuals from the previous iteration. For the support vector machine model, an incremental algorithm based on support vector caching is used to update the hyperplane. This achieves the model's continuous self-evolution capability, enabling it to adapt to constantly evolving financial risk patterns.

[0025] Secondly, the technical solution of the present invention also provides a financial risk identification system based on multi-source heterogeneous data, applied to a distributed data processing system, comprising: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data and preprocess it to generate a standardized feature matrix. The feature extraction and weight calculation module is used to calculate the information entropy from the standardized feature matrix and convert it to obtain the difference degree as the initial objective weight. Based on the historical audit accuracy rate as the confidence factor, the initial objective weight is weighted and normalized to generate a multi-dimensional behavioral feature vector. The risk reasoning module contains a pre-trained ensemble learning model, which is used to input multi-dimensional behavioral feature vectors into the ensemble learning model and output the probability distribution value of corporate financial risk. The risk assessment and early warning module is used to determine an adaptive risk threshold based on historical labeled samples and a preset false alarm rate constraint, compare the probability distribution value with the adaptive risk threshold, generate an early warning instruction when a high risk is determined, and send it to the audit terminal. The incremental learning and feedback module is used to receive verification feedback data returned by the audit terminal, extract true positive samples and false positive samples, and update the parameters of the base classifier and the adaptive risk threshold in the ensemble learning model using an incremental learning algorithm with an experience replay buffer.

[0026] Thirdly, the present invention also provides an electronic device, the electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; the memory storing computer program instructions executable by the at least one processor, the computer program instructions being executed by the at least one processor to enable the at least one processor to perform the financial risk identification method based on multi-source heterogeneous data as described in the first aspect.

[0027] As can be seen from the above technical solutions, this application has the following advantages: This invention improves the accuracy and adaptability of financial risk identification and enhances the timeliness of risk response through a variety of technical means such as multi-source heterogeneous data fusion, interpretable dynamic feature weight adjustment, adaptive risk threshold, closed-loop incremental learning, and cross-modal attention alignment. Attached Figure Description

[0028] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description 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.

[0029] Figure 1 This is a flowchart illustrating the method provided in an embodiment of the present invention.

[0030] Figure 2 A block diagram of a system provided in an embodiment of the present invention. Detailed Implementation

[0031] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0032] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0033] like Figure 1 As shown, this embodiment of the invention provides a financial risk identification method based on multi-source heterogeneous data, applied to a distributed data processing system, including the following steps: S1. Collect multi-source heterogeneous data, including at least time-series financial statement data, supply chain flow data, and unstructured text data of public opinion; and preprocess the collected data to generate a standardized feature matrix; Multi-source heterogeneous data is collected through distributed crawler nodes. This data includes at least: time-series financial statement data (including balance sheets, income statements, and cash flow statements) obtained from publicly disclosed corporate channels; supply chain flow data (including purchase orders, sales records, and logistics information) obtained from supply chain management systems or third-party trading platforms; and unstructured text data on public opinion (including news reports, regulatory inquiries, and market commentary) obtained from news media, social media platforms, and regulatory agency websites. The collected data is preprocessed to generate a standardized feature matrix. Specifically, this includes: S11. Process the time-series data of financial statements to extract numerical financial features, such as accounts receivable turnover, gross profit margin, inventory turnover, debt-to-equity ratio, and the ratio of operating cash flow to net profit; process the supply chain flow data to extract supply chain features, such as supplier concentration, customer concentration, transaction amount anomaly rate, and procurement or sales volatility coefficient; perform word vector encoding on the unstructured text data of public opinion to obtain text features; in this embodiment of the invention, a pre-trained language model (such as BERT) is used to extract text features.

[0034] The aforementioned supply chain features are cross-verified features from third parties with independent sources of enterprise transactions, used to identify fictitious transactions or related fraudulent activities; text features are used to capture external regulatory and market perception signals of enterprise risk.

[0035] S12. A cross-modal alignment method using a scaled dot product attention mechanism projects text features, numerical financial features, and supply chain features onto the same dimensional space, and performs missing value imputation and minification. Max normalization generates a standardized feature matrix; specifically, it includes: S121. Perform linear projections on the text features, numerical financial features, and supply chain features respectively, mapping them to the embedding space of the same dimension d to obtain the corresponding feature vectors. S122. Using the numerical financial feature vector as the query vector Q, and the text feature vector as the key vector K and value vector V, calculate the first attention output of the financial features and text features through scaling dot product attention. S123. Using the numerical financial feature vector as the query vector Q, and the supply chain feature vector as the key vector K and value vector V, calculate the second attention output of the financial feature and the supply chain feature through scaling dot product attention. S124. Element-wise add the first attention output and the second attention output, and perform a residual concatenation with the numerical financial feature vector. Then, perform layer normalization to obtain the fused feature matrix; wherein, the attention calculation formula is:

[0036] in, The dimension of the key vector; S125. Perform missing value imputation on the fused feature matrix and use Min-Max normalization to linearly map the feature values ​​of each dimension to the [0,1] interval to generate a standardized feature matrix.

[0037] S2. For each dimension of the standardized feature matrix, calculate the information entropy and convert it to obtain the difference degree, and use the difference degree as the initial objective weight; based on the historical audit accuracy rate obtained from historical audit feedback statistics as the confidence factor, re-normalize the initial objective weights after weighting, and obtain the dynamically adjusted feature weights, thereby generating a multi-dimensional behavioral feature vector; specifically including: S21. For the j-th dimension feature in the standardized feature matrix, calculate its information entropy. ;

[0038] In the formula, n is the sample size. For the first The sample at the th The probability of taking a value on each feature; S22. Calculate the degree of difference based on information entropy: The difference is taken as the first The initial objective weight of the first feature; based on historical audit feedback statistics, obtain the first feature's initial objective weight. Historical audit accuracy corresponding to each feature This is used as a confidence factor; the accuracy rate is obtained by statistically analyzing the confirmation rate of warnings related to this feature after they are verified by the audit, with an initial value of 0.8, which is dynamically updated as audit feedback accumulates.

[0039] S23, regarding the first The weighted difference is obtained by weighting the initial objective weights of each feature: ; S24. Renormalize the weighted variance of all features to obtain the dynamically adjusted feature weights:

[0040] Where m is the total number of dimensions of the features; S25. Multiply the eigenvalues ​​of each dimension in the standardized feature matrix by the corresponding feature weights to generate a multi-dimensional behavioral feature vector.

[0041] S3. Input the multi-dimensional behavioral feature vectors into the pre-trained ensemble learning model to output the probability distribution value of corporate financial risk. In this embodiment, the ensemble learning model includes two base classifiers: Gradient Boosting Tree (GBDT) and Support Vector Machine (SVM). Collect time-series financial statement data, supply chain flow data, and unstructured text data of public opinion from certain companies during historical periods, perform feature extraction and cross-modal alignment, and generate a standardized feature matrix. Each sample corresponds to a multi-dimensional feature vector of a company in a specific reporting period.

[0042] Each sample is labeled with a true label. y ∈{0,1}, where: positive samples ( y =1): Sample of enterprises found to have engaged in abnormal financial activities by regulatory agencies or identified as high-risk by audits; negative sample ( y =0): A sample of companies whose audit results were normal and which did not experience any financial risk events within the following three years. The positive-to-negative sample ratio was set to 1:10 to address the sample imbalance issue. The dataset was then... Divided into training sets for model parameter learning Test set used for final performance evaluation and validation set.

[0043] S31. Pre-build an ensemble learning model, wherein the ensemble learning model includes at least two base classifiers: gradient boosting tree and support vector machine; Base classifier A (Gradient Boosting Tree GBDT): Employs the XGBoost algorithm. The objective function is set to a logarithmic loss function, and a sequence of decision trees is iteratively generated using gradient descent. This model focuses on capturing nonlinear interactions between features and residual information.

[0044] Base classifier B (Support Vector Machine SVM): An SVM classifier using a Gaussian kernel function. The penalty coefficient C and kernel function parameter γ are set, leveraging its ability to find the optimal hyperplane in high-dimensional space, focusing on handling boundary partitioning problems with small sample sizes.

[0045] Gradient boosting tree training: training set ,in Indicates the first Multidimensional feature vectors of each sample ∈{0,1} represents the corresponding financial risk label (0 for normal, 1 for fraud). Positive sample proportion Initialize the predicted values:

[0046] for t =1,2,…, T Calculate the pseudo residuals: Fitting with regression trees ,get Update the model Output probability: An early stopping mechanism is used during training; training stops if the AUC on the validation set does not improve after 10 consecutive iterations.

[0047] Support Vector Machine Training: Using RBF core: Solve the dual problem: ,constraint: ; Decision function: SV is the set of support vectors (i.e., Samples corresponding to >0). Output probability: A and B are obtained on the validation set through maximum likelihood estimation.

[0048] S32. Input the multi-dimensional behavioral feature vector into each base classifier to obtain the risk probability prediction value output by each base classifier; in this embodiment, the risk probability prediction value... and .

[0049] S33. Weighted fusion of the risk probability predictions output by each base classifier:

[0050] Where T is the number of base classifiers. Let the risk probability be the output of the t-th base classifier. Let be the fusion weights of the t-th base classifier; let Given the F1 scores of the two base classifiers on the validation set, the fusion weights are:

[0051] S34. Output the fused risk probability distribution value: This serves as the probability distribution value for corporate financial risk.

[0052] S4. Compare the probability distribution value with an adaptive risk threshold determined based on historical labeled samples and constrained by a preset false alarm rate. If the value exceeds the adaptive risk threshold, it is determined to be high-risk, and an early warning instruction is generated and sent to the audit terminal; specifically including: S41. Obtain the historical labeled sample set, which includes positive samples that have been identified as risk and negative samples that have been identified as normal. S42. Input the historical labeled samples into the trained ensemble learning model to obtain the risk probability prediction value of each historical labeled sample; S43. Based on the predicted risk probability values ​​and true labels of each historical labeled sample, determine the adaptive risk threshold under the preset false alarm rate constraint; the method for determining the adaptive risk threshold is as follows: S431. Sort the historical labeled samples from high to low according to their predicted risk probability values; S432. Using the predicted risk probability value of each sample as a candidate threshold, for each candidate threshold, samples with a predicted risk probability value greater than or equal to the threshold are classified as risky, and samples with a predicted probability value less than the threshold are classified as normal; then, based on the true label of the sample, calculate the false positive rate and true positive rate corresponding to each candidate threshold; where: False positive rate = Number of samples judged as risky but with a true label of normal / Total number of samples with a true label of normal; True positive rate = Number of samples that were judged as risk and were actually labeled as risk / Total number of samples that were actually labeled as risk.

[0053] S433, Set the upper limit of false alarm rate constraint ; S434, In all cases where the false positive rate is ≤ Among the candidate thresholds, the candidate threshold that maximizes the true positive rate is selected as the adaptive risk threshold.

[0054] S44. Obtain the probability distribution value of the enterprise financial risk output by S3, and compare it with the adaptive risk threshold: If the probability distribution value is greater than or equal to the adaptive risk threshold, it is determined to be high risk; If the probability distribution value is less than the adaptive risk threshold, it is determined to be low risk; S45. When a high-risk situation is identified, an early warning instruction is generated and sent to the audit terminal. The early warning instruction includes the enterprise identifier, risk probability value, and risk level.

[0055] S5. In response to the verification feedback data returned by the audit terminal, extract the true positive samples and false positive samples, and update the parameters of the base classifier and the adaptive risk threshold in the ensemble learning model using an incremental learning algorithm with an experience replay buffer. Specifically, this includes: S51. Establish an experience playback buffer and use a first-in-first-out queue with a set maximum capacity to store historical training samples and verification feedback samples returned by the audit terminal. S52. In response to the verification feedback data returned by the audit terminal, extract true positive samples and false positive samples from them and add them to the experience playback buffer; wherein, the true positive samples are samples that the model warned of and that have been verified by the audit to have risks, and the false positive samples are samples that the model warned of but that have been verified by the audit to have no risks; S53. Randomly sample from the experience replay buffer to construct an incremental training batch; the sampling ratio is set to 50% for new samples and 50% for historical samples, and the batch size is 256.

[0056] S54. Update the parameters according to the type of the base classifier in the ensemble learning model: If it is a gradient boosting tree model, then the new weak classifier is fitted starting from the residual of the previous model. If it is a support vector machine model, the hyperplane is updated using an incremental support vector machine algorithm based on support vector caching; S55. Based on the prediction results of the updated base classifiers on the validation set or newly added verification feedback samples, recalculate the fusion weights of each base classifier. The fusion weights are recalculated according to the method in step S3, and the updated model output is obtained through weighted fusion. S56. Based on the updated model output and existing labeled samples, reselect the adaptive risk threshold according to the preset false alarm rate constraint to complete the update. The updated model and threshold are used for risk identification in the next cycle.

[0057] The ensemble learning model in this application undergoes two phases of optimization: Phase 1: Initial Training (corresponding to S3) Before model deployment, full training is performed using the historical labeled sample set to obtain initial base classifier parameters, fusion weights, and adaptive risk thresholds. The purpose of initial training is to provide a good starting point for the model. After model deployment, incremental updates are performed according to the methods described in steps S51-S56: Establish an experience replay buffer to store historical samples and true / false positive samples from audit feedback; randomly sample from the buffer to construct incremental training batches; gradient boosting tree: continue fitting new Δ values ​​from the current residual. T =10 rounds; Support Vector Machine: Retain the original support vectors, identify new support vectors, and then recalculate; recalculate F 1. Score and update the fusion weights; redetermine the adaptive risk threshold according to steps S431-S434.

[0058] Phase Two: Incremental Update (corresponding to S5) After model deployment, in response to verification feedback data returned from the audit terminal, an incremental learning algorithm with an experience replay buffer is used to continuously update the model. The purpose of incremental updates is to enable the model to adapt to evolving financial risk patterns while maintaining learning from historical knowledge.

[0059] Both phases share the same model architecture and evaluation metrics, but differ in their training strategies: initial training uses full-scale learning, while incremental updates use mini-batch learning. This design allows the model to maintain stability while possessing self-evolution capabilities.

[0060] like Figure 2 As shown, this embodiment of the invention also provides a financial risk identification system based on multi-source heterogeneous data, applied to a distributed data processing system, including: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data and preprocess it to generate a standardized feature matrix. The feature extraction and weight calculation module is used to calculate the information entropy from the standardized feature matrix and convert it to obtain the difference degree as the initial objective weight. Based on the historical audit accuracy rate as the confidence factor, the initial objective weight is weighted and normalized to generate a multi-dimensional behavioral feature vector. The risk reasoning module contains a pre-trained ensemble learning model, which is used to input multi-dimensional behavioral feature vectors into the ensemble learning model and output the probability distribution value of corporate financial risk. The risk assessment and early warning module is used to determine an adaptive risk threshold based on historical labeled samples and a preset false alarm rate constraint, compare the probability distribution value with the adaptive risk threshold, generate an early warning instruction when a high risk is determined, and send it to the audit terminal. The incremental learning and feedback module is used to receive verification feedback data returned by the audit terminal, extract true positive samples and false positive samples, and update the parameters of the base classifier and the adaptive risk threshold in the ensemble learning model using an incremental learning algorithm with an experience replay buffer.

[0061] This invention also provides an electronic device, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, communication interface, and memory communicate with each other via the communication bus. The communication bus can be used for information transmission between the electronic device and sensors. The processor can call logical instructions in the memory to execute the following method: S1, collecting multi-source heterogeneous data, the data including at least financial statement time-series data, supply chain flow data, and unstructured text data of public opinion; and preprocessing the collected data to generate a standardized feature matrix; S2, calculating the information entropy and converting it to obtain the difference degree for each dimension feature in the standardized feature matrix, and using the difference degree as the initial objective weight; using the historical audit accuracy rate obtained from historical audit feedback statistics as a confidence factor, weighting the initial objective weights and re-normalizing them to obtain dynamically adjusted feature weights, thereby generating multi-dimensional... S3. Input the multi-dimensional behavioral feature vector into the pre-trained ensemble learning model and output the probability distribution value of corporate financial risk; S4. Compare the probability distribution value with the adaptive risk threshold determined according to the historical labeled samples and the preset false positive rate constraint. If it exceeds the adaptive risk threshold, it is judged as high risk and an early warning instruction is generated and sent to the audit terminal; S5. In response to the verification feedback data returned by the audit terminal, extract the true positive samples and false positive samples, and use an incremental learning algorithm with an experience replay buffer to update the parameters of the base classifier in the ensemble learning model and the adaptive risk threshold.

[0062] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a 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 methods described in the various embodiments of the present invention. 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.

[0063] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying financial risks based on multi-source heterogeneous data, characterized in that, Applied to distributed data processing systems, it includes the following steps: S1. Collect multi-source heterogeneous data, including at least time-series financial statement data, supply chain flow data, and unstructured text data of public opinion; and preprocess the collected data to generate a standardized feature matrix; S2. For each dimension of the standardized feature matrix, calculate the information entropy and convert it to obtain the difference degree, and use the difference degree as the initial objective weight; use the historical audit accuracy rate obtained from historical audit feedback statistics as the confidence factor, weight the initial objective weight and then re-normalize it to obtain the dynamically adjusted feature weight, and then generate a multi-dimensional behavioral feature vector. S3. Input the multi-dimensional behavioral feature vectors into the pre-trained ensemble learning model and output the probability distribution value of corporate financial risk. S4. Compare the probability distribution value with the adaptive risk threshold determined based on historical labeled samples and constrained by a preset false alarm rate. If the value exceeds the adaptive risk threshold, it is determined to be high risk and an early warning instruction is generated and sent to the audit terminal. S5. In response to the verification feedback data returned by the audit terminal, extract the true positive samples and false positive samples, and use an incremental learning algorithm with an experience replay buffer to update the parameters of the base classifier in the ensemble learning model and the adaptive risk threshold.

2. The financial risk identification method based on multi-source heterogeneous data according to claim 1, characterized in that, In S1, the steps for preprocessing the collected data to generate a standardized feature matrix include: S11. Process the time-series data of financial statements to extract numerical financial features, process the supply chain flow data to extract supply chain features, and perform word vector encoding on the unstructured text data of public opinion to obtain text features. S12. A cross-modal alignment method using a scaled dot product attention mechanism projects text features, numerical financial features, and supply chain features onto the same dimensional space, and performs missing value imputation and minification. Max normalization generates a standardized feature matrix; Among them, supply chain features are cross-verified features from third parties with independent sources of enterprise transactions, used to identify fictitious transactions or related fraudulent activities; text features are used to capture external regulatory and market perception signals of enterprise risks.

3. The financial risk identification method based on multi-source heterogeneous data according to claim 2, characterized in that, The steps in S12 specifically include: S121. Perform linear projections on the text features, numerical financial features, and supply chain features respectively, mapping them to the embedding space of the same dimension d to obtain the corresponding feature vectors. S122. Using the numerical financial feature vector as the query vector Q, and the text feature vector as the key vector K and value vector V, calculate the first attention output of the financial features and text features through scaling dot product attention. S123. Using the numerical financial feature vector as the query vector Q, and the supply chain feature vector as the key vector K and value vector V, calculate the second attention output of the financial feature and the supply chain feature through scaling dot product attention. S124. Element-wise add the first attention output and the second attention output, and perform a residual concatenation with the numerical financial feature vector. Then, perform layer normalization to obtain the fused feature matrix; wherein, the attention calculation formula is: in, The dimension of the key vector; S125. Perform missing value imputation on the fused feature matrix and use Min-Max normalization to linearly map the feature values ​​of each dimension to the [0,1] interval to generate a standardized feature matrix.

4. The financial risk identification method based on multi-source heterogeneous data according to claim 1, characterized in that, The steps in S2 include: S21. For the j-th dimension feature in the standardized feature matrix, calculate its information entropy. ; In the formula, n is the sample size. For the first The sample at the th The probability of taking a value on each feature; S22. Calculate the degree of difference based on information entropy: The difference is taken as the first The initial objective weight of the first feature; based on historical audit feedback statistics, obtain the first feature's initial objective weight. Historical audit accuracy corresponding to each feature , and use it as a confidence factor; S23, regarding the first The weighted difference is obtained by weighting the initial objective weights of each feature: ; S24. Renormalize the weighted variance of all features to obtain the dynamically adjusted feature weights: Where m is the total number of dimensions of the features; S25. Multiply the eigenvalues ​​of each dimension in the standardized feature matrix by the corresponding feature weights to generate a multi-dimensional behavioral feature vector.

5. The financial risk identification method based on multi-source heterogeneous data according to claim 1, characterized in that, S3 specifically includes: S31. Pre-build an ensemble learning model, wherein the ensemble learning model includes at least two base classifiers: gradient boosting tree and support vector machine; S32. Input the multi-dimensional behavioral feature vector into each base classifier to obtain the risk probability prediction value output by each base classifier. S33. Weighted fusion of the risk probability predictions output by each base classifier: Where T is the number of base classifiers. Let the risk probability be the output of the t-th base classifier. The fusion weights of the t-th base classifier; S34. Output the fused risk probability distribution value As a probability distribution value of corporate financial risk.

6. The financial risk identification method based on multi-source heterogeneous data according to claim 1, characterized in that, The steps in S4 include: S41. Obtain the historical labeled sample set, which includes positive samples that have been identified as risk and negative samples that have been identified as normal. S42. Input the historical labeled samples into the trained ensemble learning model to obtain the risk probability prediction value of each historical labeled sample; S43. Based on the predicted risk probability values ​​and real labels of each historical labeled sample, determine the adaptive risk threshold under the preset false alarm rate constraint. S44. Obtain the probability distribution value of the enterprise financial risk output by S3, and compare it with the adaptive risk threshold: If the probability distribution value is greater than or equal to the adaptive risk threshold, it is determined to be high risk; If the probability distribution value is less than the adaptive risk threshold, it is determined to be low risk; S45. When a high-risk condition is identified, an early warning instruction is generated and sent to the audit terminal.

7. The financial risk identification method based on multi-source heterogeneous data according to claim 6, characterized in that, S43 specifically includes: S431. Sort the historical labeled samples from high to low according to their predicted risk probability values; S432. Using the predicted risk probability value of each sample as a candidate threshold, for each candidate threshold, samples with a predicted risk probability value greater than or equal to the threshold are classified as risky, and samples with a predicted probability value less than the threshold are classified as normal; then, based on the true label of the sample, calculate the false positive rate and true positive rate corresponding to each candidate threshold; where: False positive rate = Number of samples judged as risky but with a true label of normal / Total number of samples with a true label of normal; True positive rate = Number of samples that were judged as risk and whose true label was risk / Total number of samples whose true label was risk; S433, Set the upper limit of false alarm rate constraint ; S434, In all cases where the false positive rate is ≤ Among the candidate thresholds, the candidate threshold that maximizes the true positive rate is selected as the adaptive risk threshold.

8. The financial risk identification method based on multi-source heterogeneous data according to claim 1, characterized in that, In S5, in response to the verification feedback data returned by the audit terminal, true positive samples and false positive samples are extracted. An incremental learning algorithm with an experience replay buffer is used to update the parameters of the base classifier in the ensemble learning model and the adaptive risk threshold. Specifically, this includes: S51. Establish an experience playback buffer and use a first-in-first-out queue with a set maximum capacity to store historical training samples and verification feedback samples returned by the audit terminal. S52. In response to the verification feedback data returned by the audit terminal, extract true positive samples and false positive samples from them and add them to the experience playback buffer; wherein, the true positive samples are samples that the model warned of and that have been verified by the audit to have risks, and the false positive samples are samples that the model warned of but that have been verified by the audit to have no risks; S53. Randomly sample from the experience replay buffer to construct incremental training batches; S54. Update the parameters according to the type of the base classifier in the ensemble learning model: If it is a gradient boosting tree model, then the new weak classifier is fitted starting from the residuals of the previous model. If it is a support vector machine model, the hyperplane is updated using an incremental support vector machine algorithm based on support vector caching; S55. Based on the prediction results of the updated base classifiers on the validation set or newly added verification feedback samples, recalculate the fusion weights of each base classifier. And the updated model output is obtained through weighted fusion; S56. Based on the updated model output and existing labeled samples, reselect the adaptive risk threshold according to the preset false alarm rate constraint to complete the update.

9. A financial risk identification system based on multi-source heterogeneous data, characterized in that, Applications in distributed data processing systems include: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data and preprocess it to generate a standardized feature matrix. The feature extraction and weight calculation module is used to calculate the information entropy from the standardized feature matrix and convert it to obtain the difference degree as the initial objective weight. Based on the historical audit accuracy rate as the confidence factor, the initial objective weight is weighted and normalized to generate a multi-dimensional behavioral feature vector. The risk reasoning module contains a pre-trained ensemble learning model, which is used to input multi-dimensional behavioral feature vectors into the ensemble learning model and output the probability distribution value of corporate financial risk. The risk assessment and early warning module is used to determine an adaptive risk threshold based on historical labeled samples and a preset false alarm rate constraint, compare the probability distribution value with the adaptive risk threshold, generate an early warning instruction when a high risk is determined, and send it to the audit terminal. The incremental learning and feedback module is used to receive verification feedback data returned by the audit terminal, extract true positive samples and false positive samples, and update the parameters of the base classifier and the adaptive risk threshold in the ensemble learning model using an incremental learning algorithm with an experience replay buffer.

10. An electronic device, characterized in that, The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; the memory stores computer program instructions executable by the at least one processor, the computer program instructions being executed by the at least one processor to enable the at least one processor to perform the financial risk identification method based on multi-source heterogeneous data as described in any one of claims 1 to 8.

Images