Financial risk identification method and system based on multi-modal data fusion

Abstract

This invention relates to a financial risk identification method and system based on multimodal data fusion, comprising: extracting multimodal data features to be integrated into a four-dimensional model; inputting the multimodal data features into a relationship graph inference model to obtain risk identification results; the model includes: an input layer for receiving and processing multimodal data features to obtain model input; an encoding layer for performing deep encoding and feature extraction on the model input, mining the intrinsic relationships between features to obtain encoded features; a graph neural network layer for combining the graph structure information of the relationship network features to enhance the encoded features, strengthening the feature representation of enterprise relationships to obtain enhanced features; and an output layer for outputting risk identification results based on the enhanced features. This method achieves accurate identification of financial risks based on multimodal data features and enterprise relationships, significantly improving the comprehensiveness and accuracy of financial risk identification.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for identifying financial risks based on multimodal data fusion. Background Technology

[0002] The authenticity of financial information is the core foundation for the healthy operation of the market and the protection of investors' rights. Financial fraud can seriously disrupt market credit order, mislead decision-making, and cause economic losses. With the increasing complexity of business operations, the diversification of data sources, and the concealment of fraud methods, traditional financial fraud verification and identification technologies are no longer sufficient to meet the needs of accurate, efficient, and real-time supervision and auditing. Currently, mainstream financial fraud verification technologies are mainly divided into five categories, each with significant technical limitations, as detailed below: 1. Single Financial Indicator Analysis: This method relies on key ratios and indicators within financial statements (such as gross profit margin, asset turnover, and current ratio) to determine fraud risk based on deviations from industry averages or fixed thresholds. A typical example is the Altman Z-score model, which uses weighted scores from multiple indicators to predict financial distress and indirectly reflect the possibility of fraud. This type of method uses only structured data from financial statements, ignoring external operating environments and non-financial information, resulting in insufficient ability to identify systemic fraud. Research shows that its accuracy rate is typically below 60%, and it is easily misled by deliberately manipulated financial data.

[0003] 2. Rule-Based Matching Method: This method constructs a rule base based on auditing standards and historical fraud cases, using preset rules to identify fraud clues. For example, it employs warning rules such as "gross profit margin abnormally higher than the industry average" and "accounts receivable growth rate significantly higher than revenue growth rate." This method relies on continuous maintenance and updates by domain experts, requiring at least 30% of the rules to be updated annually to maintain effectiveness. It has extremely low ability to identify new and unknown fraud patterns; rule conflict handling and weight allocation depend on human experience, lack objective quantitative standards, and exhibit poor adaptability and generalization capabilities.

[0004] 3. Traditional Machine Learning Methods: This method employs algorithms such as logistic regression, support vector machines, and random forests to train classification models using historical fraud samples. It typically uses limited financial ratios and simple non-financial indicators (number of employees, number of stores, etc.) as features. While this approach achieves data-driven modeling, it suffers from limited feature dimensions, making it difficult to capture complex fraud patterns; it fails to effectively utilize time-series information, making it unable to identify abnormal trends in financial data over time; and it does not integrate unstructured data such as text and relational networks, resulting in insufficient information utilization. In practical applications, its AUC is often in the range of 0.75-0.85, indicating a relatively high false positive rate.

[0005] 4. Single-modal deep learning methods: In recent years, deep learning-based fraud detection techniques have emerged. Examples include using LSTM (Long Short-Term Memory) to analyze financial time-series data and BERT (Bidirectional Encoder Representations from Transformers) to parse sentiment and risk information in annual report texts (Management Discussion and Analysis, MD&A). While these methods offer performance improvements in single-modal scenarios, they fail to effectively integrate multi-source data, exhibit modality bias, and are susceptible to targeted circumvention due to over-reliance on a single information source. Overall, their improvement in detection capability is limited, offering only a 5%-10% increase in accuracy compared to traditional machine learning.

[0006] 5. Manual Audit Verification Methods: As a traditional supplementary method, manual auditing verifies the authenticity of data through sampling inspections, on-site visits, and confirmation letters. It heavily relies on the auditor's experience and professional judgment, is highly subjective, costly, and time-consuming. A comprehensive audit of a large enterprise typically takes 3-6 months, incurring high labor costs; it also has an underreporting rate of approximately 15% and a false positive rate of 25%, making it difficult to meet the needs of large-scale, real-time fraud monitoring.

[0007] In summary, existing financial fraud verification technologies generally suffer from common defects such as insufficient utilization of multi-source heterogeneous data, limited identification accuracy, weak adaptability, and lack of unified quantitative standards. They are unable to cope with increasingly complex, covert, and collaborative fraud behaviors. The industry urgently needs a precise and efficient financial fraud verification method that can integrate multi-dimensional data, fuse multi-modal information, and adaptively learn new fraud patterns. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a financial risk identification method based on multimodal data fusion, characterized by comprising: The structured data features, temporal variation features, unstructured text features, and relational network features of the enterprise to be identified are extracted to obtain multimodal data features; Multimodal data features are input into a relationship graph inference model to obtain risk identification results. The relationship graph inference model includes: an input layer for receiving and processing multimodal data features to obtain model input; an encoding layer for performing deep encoding and feature extraction on the model input, mining the intrinsic relationships between features to obtain encoded features; a graph neural network layer for combining the graph structure information of the relationship network features to enhance the encoded features, strengthening the feature representation of enterprise relationships to obtain enhanced features; and an output layer for outputting risk identification results based on the enhanced features.

[0009] Furthermore, the input layer processing steps include: The attention weights of temporal variation features on unstructured text features are calculated to obtain attention-fused temporal-text features; The structured data features, attention-fusion time-series-text features, and association network features are concatenated to obtain the fused features; Add positional and modality type encodings to the fused features to obtain the model input.

[0010] Furthermore, the input layer processing steps also include: Different modal features in the fused features are mapped to the common semantic space through corresponding projections, and the mean square error of the features after different modal mappings is calculated to obtain the modal alignment loss; The model parameters are optimized based on the modal alignment loss.

[0011] Furthermore, the encoding layer includes: several Transformer encoder layers; each Transformer encoder layer includes a multi-head attention module, a feedforward neural network, a residual connection, and a LayerNorm layer connected in sequence; the feedforward neural network contains two linear transformations and a GELU activation function.

[0012] Furthermore, the graph neural network layer adopts a two-layer GCN structure; The first layer GCN is used to perform preliminary linear transformation and nonlinear activation processing on the encoded features output by the Transformer encoder based on the normalized adjacency matrix, thereby achieving preliminary feature enhancement. The second-layer GCN is used to further perform linear transformations and nonlinear activation processing on the preliminary enhanced features output by the first-layer GCN, based on the same normalized adjacency matrix, to complete the deep enhancement of features and output the final enhanced features.

[0013] Furthermore, the output layer adopts a dual-branch output structure, wherein: The regression branch is used to predict the risk fit score; Classification branches are used to predict the probability distribution of risk levels.

[0014] Furthermore, the steps for constructing a training dataset before training the relation graph reasoning model include: Obtain complete annual data for the sample companies to be labeled; Based on five indicators—financial indicator deviation, non-financial data consistency, time series anomaly, related-party transaction rationality, and textual semantic risk—the scores for each indicator of the sample companies to be labeled are determined. Specifically, financial indicator deviation measures the degree of deviation between a company's financial ratios and the industry's normal level; non-financial data consistency measures the degree of matching between non-financial operating data and financial data; time series anomaly measures the degree of abnormal fluctuations in the time series of financial indicators; related-party transaction rationality measures the degree of deviation between related-party transaction prices and fair market prices; and textual semantic risk measures the strength of risk signals in the company's publicly disclosed textual information. By weighting the scores of each indicator, the risk suitability score of each labeled sample enterprise is obtained.

[0015] Furthermore, in the training of the relational graph inference model, the total loss function is determined by a weighted average of the main regression loss, auxiliary classification loss, feature reconstruction loss, and association consistency loss; among which, The main regression loss is used to calculate the mean squared error between the fit score and the labeled score. Auxiliary classification loss, used to calculate the weighted cross-entropy loss for risk level classification; Feature reconstruction loss is used to reconstruct the original features of each modality from the fused features, and to calculate the mean square error between the reconstructed modal features and the corresponding original modal features; Association consistency loss is used to calculate the mean squared error between the financial features predicted by non-financial features and the original financial features.

[0016] Further, the steps for extracting structured data features include: calculating the financial ratio features of the latest quarter and constructing a multidimensional financial ratio feature vector; calculating the benchmarking indicators between the company and the industry average and constructing a multidimensional industry benchmarking feature vector; calculating the compound annual growth rate indicator over the past few years and constructing a multidimensional growth rate feature vector; calculating the financial risk warning indicator and constructing a multidimensional risk warning feature vector; and concatenating the multidimensional financial ratio feature vector, the multidimensional industry benchmarking feature vector, the multidimensional growth rate feature vector, and the multidimensional risk warning feature vector to obtain the structured data features. Or / and, the steps for extracting time-series change features include: obtaining the time-series matrix of financial indicators of the enterprise to be identified; detecting abnormal time points in the time-series matrix of financial indicators; encoding the location information of abnormal time points into binary vectors matching the length of the time series; concatenating the binary vectors to the feature dimensions of the time-series matrix of financial indicators to form a time-series matrix of financial indicators that incorporates abnormal location information; and inputting the time-series matrix of financial indicators that incorporates abnormal location information into a two-layer LSTM network to extract time-series change features. The steps for extracting unstructured text features include: obtaining a text set of the enterprise to be identified; extracting the embedding vectors of each text segment in the text set, and performing attention fusion on the embedding vectors of each text segment to obtain text fusion features; reducing the dimensionality of the text fusion features and adding text statistical features to obtain unstructured text features. The steps for extracting network features include: constructing an enterprise relationship graph; the nodes in the graph include any one or more of enterprises, legal persons, shareholders, and senior executives; the relationships include any one or more of investment, employment, competition, cooperation, supply chain, and associated risks; based on the enterprise relationship graph, extracting node centrality features, community centrality features of nodes in their respective communities, equity structure features, and senior executive association features, and concatenating them to obtain network features.

[0017] On the other hand, the present invention also provides a computer system including a memory and a processor; the memory stores program code executable by the processor; the program code is used to execute any of the above-described financial risk identification methods.

[0018] This invention provides a financial risk identification method and system based on multimodal data fusion. The core of the method lies in the synergistic cooperation of the "four-in-one" feature input and the "encoding + graph neural network" two-layer feature extraction. The final output layer outputs the risk identification result, which realizes the accurate identification of financial risks based on multimodal data features and enterprise relationships. This significantly improves the comprehensiveness and accuracy of financial risk identification and fundamentally solves the technical problem of poor identification effect caused by the single feature and the lack of full utilization of the relationship in traditional identification models. Attached Figure Description

[0019] Figure 1 This is a flowchart of one embodiment of the financial risk identification method of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that if the embodiments of the present invention involve directional indications, such as up, down, left, right, front, back, etc., these directional indications are only used to explain the relative positional relationships and movement of the components in a specific posture. If the specific posture changes, the directional indications will also change accordingly. Furthermore, if the embodiments of the present invention involve descriptions such as "first," "second," "S1," "S2," "step one," "step two," etc., these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance, or implicitly indicating the number of technical features indicated or the execution order of the method. Those skilled in the art will understand that anything that does not violate the inventive concept should be included within the scope of protection of the present invention.

[0022] like Figure 1 As shown, this invention provides a financial risk identification method based on multimodal data fusion, comprising: S1: Extract structured data features, temporal variation features, unstructured text features, and relational network features of the enterprise to be identified, resulting in multimodal data features, which can be optionally represented as follows: .

[0023] Specifically, structured data features are quantitative features extracted from financial and operational data, which can be directly represented by numbers, tables, and fields, such as financial ratios, industry benchmarks, growth rates, and enterprise size indicators; time-series change features are financial indicator change features extracted from the time series of financial indicators, such as fluctuation frequency, amplitude, and outliers; unstructured text features are text-related features extracted from publicly available text information of the enterprise to be identified, referring to features extracted from textual information without a fixed format and that cannot be directly quantified numerically, including but not limited to full annual reports, management discussions and analyses, announcements, news, interview transcripts, and other natural language texts, which are features obtained after semantic analysis, sentiment analysis, and word embedding processing, such as text fragment features and sentiment features; and network features are graph features extracted from enterprise relationship graphs, such as node centrality features and community centrality features.

[0024] S2: Input the multimodal data features into the relationship graph inference model to obtain the risk identification result; the relationship graph inference model includes: The input layer is used to receive and process multimodal data features to obtain the model input; The encoding layer is used to perform deep encoding and feature extraction on the model input, and to uncover the intrinsic relationships between features to obtain encoded features; The graph neural network layer is used to combine the graph structure information of the associated network features to enhance the encoded features, strengthen the feature representation of enterprise relationships, and obtain enhanced features. The output layer is used to output the risk identification results based on the enhanced features.

[0025] This invention provides a financial risk identification method based on multimodal data fusion, the core advantage of which is: (1) The structured data features, time-series change features, unstructured text features and relational network features of the enterprise to be identified were extracted, and a "four-in-one" multimodal data feature input system was constructed. Its core function is to realize the comprehensive and seamless collection of the enterprise's financial and related information. The structured data features reflect the basic financial status of the enterprise, the time-series change features capture the dynamic fluctuations of financial data, the unstructured text features mine the implicit information of the enterprise, and the relational network features present the relationship between enterprises. The four complement each other and are integrated, which completely breaks the limitation that a single feature can only reflect the local state of the enterprise. It avoids the identification deviation caused by the omission of a single type of feature from the root, and provides comprehensive and reliable feature support for subsequent accurate identification.

[0026] (2) The combination of the encoding layer and the graph neural network layer constructs a progressive two-layer feature extraction unit: In terms of model processing, the encoding layer first performs unified deep encoding and feature extraction on all input features, which can transform features of different types and dimensions into a unified feature space, effectively mining the intrinsic relationship between features (such as the correspondence between financial data and text information, and the linkage between temporal fluctuations and the network), realizing deep fusion and redundancy elimination of features, and avoiding information waste caused by the isolated existence of various features; then the graph neural network layer specifically combines the graph structure information of the network features to specifically enhance the encoded features, focusing on the key identification dimension of enterprise relationship, and strengthening the representation ability of inter-enterprise relationship features through graph structure information, so that the model can accurately capture the hidden financial risk clues in the enterprise relationship, and achieve the dual effect of "comprehensive feature fusion + key relationship enhancement".

[0027] The aforementioned "four-in-one" feature input and the "encoding + graph neural network" two-layer feature extraction, through the synergistic cooperation of the two core invention points, ultimately output the risk identification results through the output layer. This achieves accurate identification of financial risks based on multimodal data features and enterprise relationships, significantly improving the comprehensiveness and accuracy of financial risk identification. It fundamentally solves the technical problem that traditional identification models suffer from poor identification results due to single features and insufficient utilization of relationships.

[0028] It is worth noting that the key to this invention lies in the overall inventive concept formed by the "four-in-one" multimodal data feature input and the two-layer feature extraction of "encoding + graph neural network". As for the specific content, expression form, extraction method, etc. of each modality feature, they can be arbitrarily set according to the specific situation and focus of the current enterprise. The specific structure of each layer in the relation graph reasoning model can be arbitrarily set according to the specific selection of multimodal data features, and is not limited to the subsequent preferred embodiments.

[0029] (I) Preferred embodiment of the "four-in-one" feature: S11: Structured data features, which can include any one or more of the following: financial ratio features, industry benchmarking features, growth rate features, and risk warning features. Taking the four sub-features of financial ratio features, industry benchmarking features, growth rate features, and risk warning features as an example, the steps for extracting structured data features include: S111: Calculate the financial ratio characteristics of the latest quarter and construct a multidimensional financial ratio feature vector; Preferably, the financial ratio characteristics can include any one or more of the following: gross profit margin, net profit margin, asset turnover, inventory turnover, accounts receivable turnover, current ratio, quick ratio, debt-to-equity ratio, equity ratio, interest coverage ratio, revenue growth rate, and net profit growth rate.

[0030] More preferably, the multidimensional financial ratio feature vector can be a 12-dimensional financial ratio feature vector, which includes all the key financial ratio features in the example above.

[0031] A more preferable approach is to standardize each financial ratio across the industry: ,in For financial ratios, and These are the mean and standard deviation of the aforementioned financial ratios within the industry. These are standardized financial ratios.

[0032] S112: Calculate the benchmarking indicators between the enterprise and the industry average level, and construct a multi-dimensional industry benchmarking feature vector; Preferably, the benchmarking indicators may include the enterprise revenue benchmarking index, the enterprise asset size benchmarking index, the enterprise employee headcount benchmarking index, and any one or more of the industry ranking percentages of several key financial ratios among the above financial ratios.

[0033] More preferably, the multi-dimensional industry benchmarking feature vector can be selected as an 8-dimensional feature vector, including And industry percentile rankings for 5 key financial ratios.

[0034] S113: Calculate the compound annual growth rate (CAGR) for recent years and construct a multidimensional CAGR feature vector; Preferably, the compound annual growth rate (CAGR) indicator for the past three years can be calculated, and may include any one or more of the following: revenue CAGR, profit CAGR, asset CAGR, R&D investment CAGR, market share CAGR, and headcount CAGR.

[0035] More preferably, the multidimensional growth rate feature vector can be a 6-dimensional feature vector, including all the compound growth rate features in the examples above.

[0036] S114: Calculate financial risk early warning indicators and construct a multi-dimensional risk early warning feature vector; Preferably, the risk warning indicators calculated based on the traditional financial fraud warning model may include industry-calibrated risk scores such as Altman Z-score, M-score, and F-score. More preferably, the multidimensional risk warning feature vector can be selected as a 6-dimensional feature vector, including industry-calibrated Altman Z-score, M-score, F-score, comprehensive financial risk score, industry deviation score, and time series anomaly score.

[0037] S115: Concatenate multidimensional financial ratio feature vectors, multidimensional industry benchmark feature vectors, multidimensional growth rate feature vectors, and multidimensional risk warning feature vectors to obtain structured data features.

[0038] Preferably, by concatenating the aforementioned 12-dimensional financial ratio feature vector, 8-dimensional industry benchmark feature vector, 6-dimensional growth rate feature vector, and 6-dimensional risk warning feature vector, a 32-dimensional structured data feature is obtained. .

[0039] S12: Time-series variation features, which can be extracted from the time series of financial indicators using deep learning models such as LSTM. Preferably, the steps for extracting time-series variation features include: S121: Obtain the time series matrix of the financial indicators of the enterprise to be identified, which can be represented as follows: ; in, For the enterprise to be identified The time series matrix of financial indicators includes T elements; To select the time step, you can choose the most recent 60 quarters, which is 15 years of data. The number of financial metrics for each time step can be selected, including 12 key metrics. S122: Input the time series matrix of financial indicators into a two-layer LSTM network to extract time series variation features, which can be represented as follows: ; ; in, This is the first layer of LSTM, containing 64 hidden units; It is a second-layer LSTM containing 32 hidden units; and These are the parameters for the two LSTM layers, respectively. This is the output of the first LSTM layer; The output of the second LSTM layer at the last time step serves as the final temporal variation feature.

[0040] Preferably, in order to capture outliers in the time series and enhance the model's sensitivity to abnormal fluctuations, step S121 may further include the following, prior to the extraction of time series change features in step S122: S1211: Detect abnormal time points in the financial indicator time series matrix; encode the location information of abnormal time points into binary vectors that match the length of the time series, concatenate the binary vectors to the feature dimension of the financial indicator time series matrix to form a financial indicator time series matrix that integrates abnormal location information, and then execute step S122: input into a two-layer LSTM network for temporal feature extraction.

[0041] Preferably, the step of detecting outlier time points in the financial indicator time series matrix can use the Isolation Forest algorithm, as follows: ; in, Mark the results of outlier points. This is the Isolation Forest algorithm, where `nestimators` is the number of decision trees (ensemble trees) in the Isolation Forest, and is optional. Setting it to 100 indicates building an ensemble forest of 100 isolated trees to improve the stability and accuracy of anomaly detection; contamination is the proportion of outliers in the dataset (contamination level), which is optional. Setting it to 0.05 indicates that the percentage of outliers in the time series of the financial indicators to be detected is 5%, which is used to guide the algorithm in determining the threshold of outliers.

[0042] More preferably, before extracting the time-series variation features in step S122, step S121 may further include: S1212: Standardize the time series of each financial indicator: ; in, For the t-th time series, the first... One indicator, and The first The mean and standard deviation of each indicator over the entire time series. The standardized time series of the t-th time series One indicator.

[0043] S13: Unstructured text features, which can be extracted from publicly available corporate text information (such as annual reports, announcements, management interviews, etc.) to obtain text features such as BERT embeddings of annual report texts and emotional features of management interview speech. Preferably, the steps for extracting unstructured text features include: S131: Obtain the text set of the enterprise to be identified, which can be optionally represented as: ; in, Indicates enterprise A collection of texts, For enterprises The A text fragment, The number of text segments; preferably, each text segment may undergo text preprocessing such as word segmentation, stop word removal, and technical terminology recognition; S132: Extract the embedding vectors of each text segment in the text set, and perform attention fusion on the embedding vectors of each text segment to obtain the text fusion features; Preferably, a domain-pretrained model can be used for encoding, and the embedding vectors of each text segment can be obtained using a large model finely tuned for the financial domain (A Robustly Optimized BERT Pretraining Approach), represented as follows: ; in, For text fragments The embedding vector; This is the set of learnable parameters for the RoBERTa model.

[0044] More preferably, text fragment attention fusion is performed on multiple text fragments embedded from the same enterprise, as shown below: ; ; in, Embedded vector Attention weights The attention score for the k-th text segment. The number of text fragments. Embedded vector Attention score Here is the weight matrix of the attention layer. For the bias term of the attention layer, This is the overall textual semantic feature vector of the enterprise after attention-weighted fusion.

[0045] S133: Dimensionality reduction of text fusion features and addition of text statistical features to obtain unstructured text features; Specifically, the text fusion features are reduced in dimensionality and enhanced: the 768-dimensional text fusion features are reduced to 128 dimensions, and text statistical features are added to obtain unstructured features, represented as follows: ; in, and For dimensionality reduction parameters, For text statistical feature vectors, For unstructured text features, a more preferred text statistical feature vector may include: TF-IDF weights for risk words (32 dimensions): calculated based on a financial fraud risk dictionary; sentiment scores (8 dimensions): including positive, negative, and neutral sentiment intensity and sentiment fluctuation; text complexity indicators (8 dimensions): including sentence length, vocabulary difficulty, proportion of ambiguous words, etc.; and features from the Management Discussion and Analysis (MD&A) section (80 dimensions): including thematic features extracted specifically for the MD&A section of the annual report.

[0046] S14: The steps for extracting features from the associated network include: S141: Construct a corporate relationship graph; the nodes in the graph include: any one or more of enterprises, legal persons, shareholders, and senior executives; the relationships include: any one or more of investment, employment, competition, cooperation, supply chain, and related risks; Example corporate relationship graph. ,in For a set of nodes, For a set of relations, nodes , is one of the companies in the map ; S142: Based on the enterprise relationship graph, extract node centrality features, community centrality features of nodes in their respective communities, equity structure features, and executive association features, and concatenate them to obtain the relationship network features. ; Preferably, the node centrality features include any one or more of degree centrality, degree centrality, betweenness centrality, compact centrality, eigenvector centrality, and PageRank value; the community centrality features may include any one or more of the following: related-party transaction scale, related-party transaction complexity, related-party transaction concentration rate, related-party transaction price deviation, related-party path length, related-party guarantee scale, fund transfer frequency, and related-party transaction growth rate; the equity structure features may include any one or more of the following: equity concentration, equity checks and balances, institutional shareholding ratio, management shareholding ratio, equity pledge ratio, and separation of control rights and cash flow rights of the actual controller; and the executive network features may include any one or more of the following: number of executives holding concurrent positions, executive association strength, executive turnover rate, and stability of key positions.

[0047] More preferably, the node centrality feature can be selected as a 6-dimensional node centrality feature vector, including: degree centrality, betweenness centrality, compact centrality, eigenvector centrality, and PageRank value; Community centrality feature, which can be selected as an 8-dimensional community centrality feature vector; including: scale of related transactions: Related-party transaction complexity: the number and hierarchy of related parties involved in related-party transactions; Related-party transaction concentration ratio: the proportion of the largest related-party counterparty; Related-party transaction price deviation: the average deviation between related-party transaction prices and market prices; Related-party path length: the shortest path distance to key related parties; Related-party guarantee scale: the ratio of total external guarantees to net assets; Funding transaction frequency: the number of fund transactions with related parties; Related-party transaction growth rate: the growth rate of related-party transaction scale over the past 3 years; The equity structure characteristics can be selected as a 6-dimensional equity structure feature vector, including: equity concentration: the sum of the shareholding ratios of the top ten shareholders; equity checks and balances: the ratio of the sum of the shareholding ratios of the second to tenth largest shareholders to the shareholding ratio of the largest shareholder; institutional shareholding ratio: the proportion of total shares held by institutional investors to the total share capital; management shareholding ratio: the proportion of total shares held by the company's management to the total share capital; share pledge ratio: the proportion of pledged shares to the total share capital; and the degree of separation between the actual controller's control and cash flow rights. Executive network characteristics can be selected as a 6-dimensional executive network feature vector; number of executive concurrent positions: the number of times executives hold concurrent positions in other companies; executive association strength: the degree of association between executives and other companies; executive turnover rate: the proportion of executives leaving in the past year; key position stability: the tenure of the financial officer and the internal audit officer.

[0048] (II) Preferred embodiment of the model structure: (1) Input layer: In a preferred embodiment, the input layer processing steps include: S21: Calculate the attention weights of temporal variation features on unstructured text features to obtain the attention fusion temporal-text features; specifically, a cross-modal attention mechanism can be selected to fuse temporal-text cross-modal attention, expressed as: ; ; in, Features of temporal variation The One portion, Features of unstructured text. Features of temporal variation The One portion, The first of the time-series variation characteristics The attention weights of each component for unstructured text features. Here is the attention weight matrix. Temporal features after attention weighting For query With candidates The matching score.

[0049] S22: Feature Concatenation: Concatenate structured data features, attention-fusion temporal-text features, and association network features to obtain fused features, which can be represented as: .

[0050] More preferably, step S2 may also include: S23: Fusion Feature Standardization: Perform LayerNorm standardization on the stitched fusion features, which can be represented as follows: ; in, Indicates fusion features, and These are the mean and standard deviation of the fusion feature, respectively. To prevent small constants from being divided by zero.

[0051] S24: Modality Alignment: To enhance the consistency of different modal features in the common semantic space, a modality alignment loss is introduced; specifically: different modal features in the fused features are mapped to the common semantic space through corresponding projections, and the mean square error of the features after different modal mappings is calculated to obtain the modality alignment loss. The model parameters are optimized based on the modality alignment loss; specifically, it can be expressed as: ; ; ; in, and Modal and modality The features are represented by the modality-specific projection matrix after being mapped to the common space. For the modal loss, and For modality The projection parameters, As a public space dimension, For modality The original feature dimension and For modality The projection parameters, For modality The original feature dimensions.

[0052] S25: Add positional and modality type encodings to the fused features to obtain the model input. Specifically, receive 248-dimensional fused features. And add position encoding and modality type encoding, represented as: ; in, For position encoding vectors, It is a modality type encoding vector that indicates the modality to which each dimension of the feature belongs.

[0053] This embodiment presents a preferred embodiment of the input layer. Through targeted multimodal data feature fusion processing, the quality of the fused features of the input model can be further optimized, improving the accuracy and efficiency of financial risk identification. It proposes: first, calculating the attention weights of temporal variation features on unstructured text features to obtain attention-fused temporal-text features; then, concatenating the structured data features, attention-fused temporal-text features, and association network features to obtain the overall approach of fused features, achieving at least the following technical effects: First, by calculating the attention weights of temporal variation features on unstructured text features, attention-fusion temporal-text features are obtained, enabling precise synergistic association between temporal and text features. Temporal variation features reflect the dynamic fluctuations of corporate financial data (such as temporal anomalies in revenue and profit), while unstructured text features contain implicit information about corporate operational risks (such as risk warnings and management statements in annual reports). The attention mechanism can assign differentiated weights to temporal variation features based on the correlation between the two types of features and financial risks, accurately capturing the inherent correspondence between temporal fluctuations and implicit textual information (such as the matching degree between temporal anomalies in financial data and risk statements in the text). This avoids information misalignment and weakened correlation caused by the isolated fusion of the two types of features, strengthening the representational ability of temporal-text fusion features for financial risks.

[0054] Second, the step-by-step processing approach of "first fusing time-series and text features, then concatenating them with structured data features and relational network features" balances the targeting and completeness of feature fusion. Focusing on attention fusion of time-series and text features, which are highly correlated, allows for the priority extraction of their synergistic value. Then, these features are concatenated with structured data features reflecting the company's financial fundamentals and relational network features reflecting the company's interrelationships. This avoids the problems of feature redundancy, dimensional confusion, and dilution of key information that result from directly mixing and concatenating multiple types of features. It also achieves complementary fusion of various features, ensuring that the fused features retain the core information of individual features while possessing the synergistic value between features.

[0055] Third, this preliminary feature processing workflow can pre-screen, fuse, and optimize input features, effectively reducing the processing pressure on subsequent models (encoding layer, graph neural network layer), while providing a more targeted and high-quality feature foundation for subsequent model processing. The fused features obtained through attention fusion and step-by-step splicing have pre-strengthened the key feature combinations related to financial risk and eliminated invalid and redundant information, enabling the deep encoding of the subsequent encoding layer and the association strengthening of the graph neural network layer to focus more on core risk clues, further improving the efficiency and accuracy of financial risk identification. This compensates for the shortcomings of insufficient recognition accuracy and low processing efficiency caused by the direct splicing of traditional multimodal data features, further optimizing the recognition effect based on claim 1.

[0056] Furthermore, based on the overall approach described above, the input layer can optionally perform standardization, modality alignment, and add positional and modality type encoding on the fused features. This can further improve model stability and recognition accuracy. Specifically: Standardization plays a crucial role in further eliminating dimensional differences between different modal features (such as the numerical magnitude of structured data and the scaling differences in the dimension of text / relational network feature vectors). It prevents certain types of features from being overemphasized or ignored due to magnitude issues, ensures that all types of features participate in encoding at the same scale, and guarantees the fairness and accuracy of feature extraction in subsequent encoding layers.

[0057] Modality alignment aims to address the semantic misalignment problem of different types of features (structured, temporal-text fusion, and relational networks), mapping various modal features to a unified semantic space. This allows the coding layer to efficiently mine the intrinsic relationships between different modal features and avoids information waste caused by feature isolation.

[0058] The role of positional encoding is to endow fused features with positional information, clarify the temporal or logical order of various features (such as the temporal correlation of temporal-text features, the splicing order of different modal features), help the encoding layer capture the sequential dependencies of features, and improve the completeness of feature encoding.

[0059] The role of modality type encoding is to add identifiers to different modality dimensions in the fused features, clarify the modality type to which each dimension of features belongs (such as which dimensions are structured features and which are time-series-text fusion features), help the encoding layer distinguish different modality features, specifically mine the core information of various modalities, avoid encoding bias caused by confusion of different modality features, and further improve the accuracy of feature extraction in the encoding layer.

[0060] (2) Coding layer: In a preferred embodiment, the encoding layer may optionally include: several Transformer encoder layers; each Transformer encoder layer includes a multi-head attention module, a feedforward neural network, a residual connection, and a LayerNorm layer connected in sequence; the feedforward neural network includes two linear transformations and a GELU activation function.

[0061] More preferably, the encoding layer may include: 3 Transformer encoder layers; each Transformer encoder layer includes: Multi-head self-attention module, example includes: 8 attention heads, each head with 31 dimensions ( ), represented as: ; in, This is the output of the multi-head self-attention module. This indicates a splicing operation. to The outputs of each attention head are respectively. This is the concatenated linear transformation weight matrix. For the first The output of each attention head, This is the result of single-head attention calculation. For query vectors, For key vectors, For value vectors, , , The h-th attention head corresponds to The linear transformation weight matrix, For activation function, Key vector Dimensions.

[0062] A feedforward neural network, consisting of two linear transformations and the GELU activation function, is represented as follows: ; in, This is the output of the feedforward neural network. As the input to the feedforward neural network, This is the weight matrix from the input layer to the hidden layer of the feedforward network. Here, represents the bias term from the input layer to the hidden layer of the feedforward network, and GELU is the activation function. This is the weight matrix from the hidden layer to the output layer of the feedforward network. This is the bias term from the hidden layer to the output layer of the feedforward network.

[0063] This embodiment presents a preferred embodiment of the encoding layer. The specific structural optimization of this encoding layer is not a simple selection of conventional techniques, but a targeted design specific to the aforementioned multimodal data features. This further improves feature encoding quality and model training stability, providing higher-quality encoded features for subsequent financial risk identification. Specific technical effects are as follows: The encoding layer sets up several Transformer encoders. Through the collaborative work of each module in each encoder layer, efficient and accurate deep encoding of the input fusion features is achieved. The multi-head attention module can accurately uncover the intrinsic relationships between features, breaking the isolation of features from different modalities. The synergistic value of various features is extracted to avoid wasting feature information. The feedforward neural network performs nonlinear transformations on the encoded features through two linear transformations combined with the GELU activation function, effectively enhancing the representational ability of the features and better capturing the hidden financial risk clues in the features. Residual connections can effectively alleviate the gradient vanishing problem during deep training of the model, ensuring that the encoding layer can complete the feature extraction task stably and efficiently. The LayerNorm layer normalizes the features, eliminating differences in feature dimensions and distribution offsets, ensuring the stability of the output features of each module, and providing a standardized and high-quality encoding feature foundation for feature enhancement of subsequent graph neural network layers. Through the reasonable setting and functional synergy of each module in the encoding layer, the accuracy and stability of feature encoding are further optimized, the difficulty of model training is reduced, and the overall efficiency and accuracy of financial risk identification are indirectly improved, making up for the shortcomings of insufficient feature extraction and unstable model training caused by unreasonable structural design of traditional encoding layers.

[0064] (3) Graph Neural Network Layer: In a preferred embodiment, the graph neural network layer adopts a two-layer GCN (Graph Convolutional Network) structure. The first GCN layer is used to perform preliminary linear transformation and nonlinear activation processing on the encoded features output by the Transformer encoder based on the normalized adjacency matrix, thereby achieving preliminary feature enhancement. The second GCN layer is used to perform further linear transformation and nonlinear activation processing on the preliminary enhanced features output by the first GCN layer based on the same normalized adjacency matrix, thereby completing the deep enhancement of the features and outputting the final enhanced features. Optionally, it can be represented as follows: ; ; in, For the output of the Transformer encoder, To normalize the adjacency matrix, This is the first layer weight matrix. This is the output of the first layer GCN. This is the weight matrix for the second layer. This is the output of the second-layer GCN.

[0065] In this embodiment, a preferred embodiment of the graph neural network layer is given. The first layer GCN is used to focus on the basic graph structure information of the associated network features (such as the relationship between enterprise nodes and the strength of node connections), and to initially enhance the encoded features output by the encoding layer. It focuses on mining the feature information corresponding to the direct relationship between enterprises, and realizes the initial fusion of encoded features and the basic structure information of the associated network, providing a foundation for subsequent deep enhancement. The first layer GCN is used to further mine the deep graph structure information of the associated network features (such as the indirect relationship path between enterprises and the topological features of the associated network) based on the initial enhancement of the first graph neural network, and to perform secondary enhancement on the features after the initial enhancement. It focuses on capturing the financial risk association clues hidden by enterprises through multi-layered relationships, and strengthens the deep feature representation of enterprise relationships. By employing a two-layer graph neural network design of "preliminary enhancement + deep reinforcement," compared to a single graph neural network, it can fully utilize the basic structural information of the interconnected network while also uncovering deep-level connections. This avoids the shortcomings of a single graph neural network, which struggles to simultaneously address both "basic connection capture" and "deep connection mining." Furthermore, both layers of the graph neural network revolve around the graph structure information of the interconnected network features, progressively strengthening the representational ability of enterprise relationships within the encoded features. This results in more accurate risk identification in subsequent output layers, further enhancing the comprehensiveness and accuracy of financial risk identification and optimizing the model's recognition performance.

[0066] (4) Output layer: In a preferred embodiment, the output layer adopts a dual-branch output structure: The regression branch, used to predict the risk fit score, is represented as: ; in, , , To predict risk and fit the score; The classification branch, used to predict the probability distribution of risk level, is represented as: ; in, , , The probability distribution is for three risk levels (high, medium, and low).

[0067] (III) Preferred Implementation Method for Model Training: (1) The steps for constructing the training dataset include: P1: Obtain complete annual data for the sample companies to be labeled; P2: Determine the scores of each indicator for the sample companies to be labeled based on five indicators: deviation of financial indicators, consistency of non-financial data, anomaly of time series, reasonableness of related-party transactions, and semantic risk of text. P3: Weight the scores of each indicator to obtain the risk fit score for each labeled sample company. Specifically: Sample selection: Complete annual data of 500 companies were selected as the labeled sample, including companies of different industries, sizes and years of listing. Among them, known fraud cases accounted for 20% (100 companies), and normal companies accounted for 80% (400 companies).

[0068] Labeling Process: A panel of experts was formed, comprising certified public accountants with over 5 years of auditing experience, financial risk management experts, and financial technology analysts. Each expert independently assigned a risk fit score to each sample. The labels indicate how well the financial data fits the "normal business model," with lower scores indicating a higher risk of fraud. For example... This indicates an extremely high risk of fraud. This indicates an extremely low risk of fraud. Finally, a weighted calculation is used to obtain the overall risk fit score for each sample company.

[0069] Preferably, to overcome the subjective and biased errors in the annotation results, this invention also proposes a structured quantitative model for the sample annotation step before training the relational graph inference model, constructing a five-dimensional core indicator vector. The definitions of each dimension indicator are as follows: Financial indicator deviation It is used to measure the degree to which a company's financial ratios deviate from the industry norm. Defined as: ;in, For the number of selected key financial ratios, For enterprises The Financial ratios, This is the industry average financial ratio. and They are the first in the industry The maximum and minimum values ​​of each financial ratio. The selected financial ratios can include any one or more of the following 12 core indicators: gross profit margin, net profit margin, asset turnover, inventory turnover, accounts receivable turnover, current ratio, quick ratio, debt-to-equity ratio, equity ratio, interest coverage ratio, revenue growth rate, and net profit growth rate.

[0070] Non-financial data consistency It is used to measure the degree of matching between non-financial operating data and financial data. Defined as: ;in, For non-financial data types, For enterprises The This is a set of actual non-financial data, which can be selected as any one or more of the following: electricity consumption, logistics and transportation volume, number of employees, number of stores, etc. For enterprises The This refers to forecasting non-financial data, i.e., forecasting data based on financial data. For the first The weighting parameters for non-financial data items can be determined through calibration using historical data.

[0071] Time series anomalies It is used to measure the degree of abnormal fluctuation in the time series of financial indicators. Defined as: ;in, For enterprises The number of mutation points in the financial indicator series can be obtained by using the ADF test and the PELT algorithm. The average mutation magnitude is the average value of the change in the index at all mutation points. and To adjust the parameters, after experimental calibration, the values ​​can be set to 0.3 and 0.5 respectively.

[0072] Reasonableness of related-party transactions It is used to measure the degree of deviation between the price of related-party transactions and the fair market price. Defined as: ;in, For enterprises The number of related transactions, For enterprises The The actual price of the related-party transaction This is the fair market price of similar, unrelated transactions during the same period. If the company has no related-party transactions, then... .

[0073] Textual semantic risks It is used to measure the strength of risk signals in publicly available corporate text information. Defined as: ;in, For enterprises The embedding vector of the text information can be extracted, optionally through a pre-trained language model in the financial domain. and These are the parameters obtained through training with labeled data. Optional text information includes annual reports, semi-annual reports, quarterly reports, and announcements of major events.

[0074] The score for each indicator is determined using the calculation formula of the above five-dimensional core indicator vector, and the risk suitability score for each sample enterprise is determined by weighting these scores; specifically: The weight vectors of the five-dimensional indicators are determined using AHP (Analytic Hierarchy Process). ,satisfy Then the enterprise The overall fit score is defined as: ,in, For the first The weighting coefficients of each indicator For enterprises The Scores for each metric; optional base weight vector The weighting is as follows: financial indicator deviation has the highest weight (30%), followed by non-financial data consistency (25%) and time series anomalies (20%). The weighting of related-party transaction rationality (15%) and textual semantic risk (10%) is relatively low. These weightings can be adjusted according to industry characteristics.

[0075] (2) Training objective and loss function: In the training of the relational graph inference model, the total loss function is determined by a weighted average of the main regression loss, auxiliary classification loss, feature reconstruction loss, and association consistency loss; where: The main regression loss, used to calculate the mean squared error between the fit score and the labeled score, is expressed as: ; The auxiliary classification loss, used to calculate the weighted cross-entropy loss for risk level classification, is expressed as: ; Feature reconstruction loss is used to reconstruct the original features of each modality from the fused features, and to calculate the mean square error between the reconstructed modal features and the corresponding original modal features, expressed as: ; The association consistency loss is used to calculate the mean squared error between the financial characteristics predicted from non-financial characteristics and the original financial characteristics, to ensure the consistency between financial indicators and non-financial data. It is expressed as: ; Total loss function: .

[0076] in, Main regression loss, For the sample size, The fit score predicted by the model. The adaptation score for the annotation; For weighted cross-entropy loss, For category weights, For the sample of the first Each risk level has a real label (unique hot code). To predict probabilities, Optional settings This is to address the class imbalance problem (high-risk samples have the highest weight). For feature reconstruction loss, For each modal type, For original modal features, These are the modal features reconstructed from the fused features using a decoder; For the loss of association consistency, For financial characteristics, For financial characteristics predicted from non-financial characteristics, Non-financial characteristics; This represents the total loss.

[0077] In this embodiment, the total loss function weights and fuses the main regression loss, auxiliary classification loss, feature reconstruction loss, and association consistency loss. Combined with the dual-task training process of "main regression + auxiliary classification," it can achieve synergistic optimization of regression and classification tasks. Its core technical effects are: it ensures the accuracy of fitting score prediction through the main regression loss, improves the accuracy of risk level classification through the auxiliary classification loss, constrains the integrity of the representation of original features of each modality by the fusion features through the feature reconstruction loss, avoids the loss of modality information during the fusion process, and ensures the semantic consistency between financial and non-financial data through the association consistency loss, reducing the impact of data bias on model training. The weighted synergy of multiple losses can take into account the training effect of the model on both tasks, alleviate the model bias problem caused by single-task training, and optimize model parameters through multi-dimensional loss constraints, further improving the training stability and generalization ability of the model. This makes the final fitting score and risk level prediction results output by the model more accurate and reliable, providing stronger support for financial risk identification.

[0078] (3) Training process: Data preparation: Training set: Data from 400 companies (80%), including 16% known fraud cases; Validation set: Data from 50 companies (10%), including 2% known fraud cases; Test set: Data from 50 companies (10%), including 2% known fraud cases; Data Augmentation: Time Series Data: Adding Gaussian Noise ( Time misalignment (±1 quarter); Text data: synonym replacement, random insertion / deletion (10%); Class balance: SMOTE oversampling is used to process high-risk samples; Data Updates: The system adopts a T+1 update mechanism, automatically collecting and processing the latest data daily to update enterprise feature vectors. It collects data from multiple sources through API interfaces, web crawlers, and database synchronization. Specifically, it includes: structured financial data (daily updates from financial databases); unstructured text data (real-time updates from company announcements, news reports, and social media information); non-financial operational data (updated on demand from enterprise ERP systems and IoT device data); and related network data (real-time updates from announcements of equity changes, senior management changes, and related-party transactions).

[0079] Model inference: Based on multimodal data features; forward propagation, calculating the fit score through the trained BERT-GCN hybrid model. Output adaptation score and risk level probability distribution ; Dynamic threshold calculation: ; ; ; in, and For enterprises The mean and standard deviation of the fit score for the relevant industry are calculated based on industry data from the most recent quarter.

[0080] Threshold dynamic adjustment mechanism: Industry adjustment: Different industries use their own threshold parameters; Cycle adjustment: The threshold is adjusted according to the economic cycle stage (expansion, contraction, and stability). ;in This is an economic cycle index. The periodic sensitivity coefficient; In summary, this application provides a financial risk identification method based on multimodal data fusion, with at least the following specific effects: 1. Achieve a quantitative transformation in fraud detection, significantly improving the objectivity and accuracy of judgment: By constructing a quantitative system of "adaptive score" and a scientific weight allocation mechanism, the traditional qualitative judgment is transformed into standardized quantitative analysis, which completely solves the problems of reliance on human experience and inconsistent identification standards, and greatly improves the core accuracy of fraud detection compared with the traditional single threshold method.

[0081] 2. Breaking through the limitations of single-source data and enhancing the ability to capture complex fraudulent behaviors: The innovative multimodal data feature fusion mechanism deeply integrates heterogeneous data, constructs a "four-in-one" high-dimensional feature space, and mines cross-modal correlations to comprehensively cover the multi-dimensional traces of fraudulent behaviors. Compared with single financial data methods, it significantly improves the model's discriminative power and achieves a qualitative leap in the identification effect of complex fraud with strong concealment.

[0082] 3. Adapt to dynamic differences in multiple scenarios, significantly reducing false alarm rate and performance fluctuation: The dynamic threshold adjustment mechanism based on industry and cycle characteristics enables intelligent adaptation of risk judgment standards, effectively solving the limitations of fixed threshold scenarios, significantly narrowing the range of accuracy fluctuations between different industries, maintaining stable performance during economic cycle transitions, and significantly reducing false alarm rate.

[0083] 4. Overcoming the "black box" problem of models and meeting the traceability requirements of high regulatory standards: The BERT-GCN hybrid architecture achieves high recognition accuracy while clearly presenting the contribution of each factor to the risk identification results, solving the "black box" problem of deep learning models, significantly improving the trust of auditors and regulators in the model results, and greatly improving decision-making efficiency.

[0084] 5. Complete the upgrade from passive analysis to proactive early warning, significantly shortening the risk discovery cycle: The T+1 dynamic update and real-time identification mechanism enables continuous monitoring of corporate financial risks, breaking through the limitations of ex-post analysis, greatly shortening the average time to discover fraudulent behavior, gaining a golden window for risk disposal, significantly improving the recovery rate of potential losses, and meeting the performance requirements of large-scale real-time monitoring.

[0085] 6. Alleviate the dilemma of sample scarcity and ensure the long-term recognition performance stability of the model: By using efficient sample utilization strategies and incremental learning mechanisms, the value of limited labeled samples is fully explored. At the same time, the model can dynamically adapt to new fraud patterns, significantly reduce the model performance decay rate, effectively control long-term maintenance costs, and solve the industry pain point that static models are difficult to be continuously applied.

[0086] 7. High availability and strong scalability, lowering the threshold for industry implementation: The modular microservice architecture supports elastic deployment and seamless integration, achieving ultra-high system availability. It can support the simultaneous monitoring needs of ultra-large-scale enterprises, significantly reduce the cost of connecting with existing audit and supervision systems, and significantly enhance the engineering promotion value of the technical solution.

[0087] (iv) Experimental verification: This invention has undergone thorough experimental verification, proving that its technical solution is feasible and its effect is significantly superior to existing technologies. The experimental process and results are as follows: (1) Data sample situation: Sample Source: Annual data from A-share listed companies from 2015 to 2022, covering 10 major industries (finance, manufacturing, information technology, real estate, etc.); Sample Size: A total of 500 companies, including 100 known fraudulent companies (confirmed by the China Securities Regulatory Commission's penalty announcements) and 400 normal companies; Data Dimensions: Each company includes 32-dimensional structured features, 64-dimensional time-series variation features, 128-dimensional unstructured text features, and 24-dimensional relational network features; Data Division: Divided into a training set (400 companies), a validation set (50 companies), and a test set (50 companies) in an 8:1:1 ratio to ensure a consistent proportion of fraud cases across all sets; (2) Comparative experimental design: Traditional financial indicator method: based on the Altman Z-score model; Single machine learning methods: Random Forest, XGBoost; Single-modal deep learning methods: LSTM (temporal features only), BERT (text features only); The method of this invention; simultaneously designing ablation experimental models (removing one modal feature respectively).

[0088] Definition of performance indicators: Regression performance indicators: Mean squared error, Mean absolute error; Classification performance indicators: Accuracy, Precision, Recall, F1 score, AUC; Efficiency indicators: Single identification time: Time required to identify fraud risk in one company; Feature update time: Total time spent updating features in T+1 days. (3) Experimental results: Table 1: Comparison Results of Regression Performance

[0089] As shown in Table 1, the experimental results demonstrate that the complete model of this invention has an MSE of 0.52 and a MAE of 0.51, significantly outperforming existing technologies. Compared to the best baseline method (LSTM), the MSE is reduced by 49.5% and the MAE by 32.0%, indicating that this invention has higher prediction accuracy for the fit score and better consistency with the labeled score.

[0090] Table 2: Comparison Results of Classification Performance

[0091] As shown in Table 2, the complete model of this invention performs optimally across all classification metrics, achieving an AUC of 0.95, a 6.7% improvement over the best baseline method (LSTM). In particular, its ability to identify high-risk samples is significantly enhanced, with a recall rate of 0.89, meaning it can capture 89% of actual fraud cases, a 31 percentage point improvement over traditional methods, greatly reducing the risk of false negatives. The precision rate reaches 0.91, indicating that 91% of the high-risk cases predicted by the model actually involve fraudulent behavior, resulting in a low false positive rate.

[0092] Table 3: Effect of Dynamic Threshold Adjustment Mechanism

[0093] As shown in Table 3, the dynamic threshold adjustment mechanism makes the model's performance more stable across different industries and economic cycles. The industry accuracy fluctuation decreased from ±15.3% to ±4.8%, and the cycle accuracy fluctuation decreased from ±12.7% to ±3.5%, proving that the mechanism effectively solves the problem of recognition performance fluctuation caused by industry differences and cycle changes.

[0094] Table 4: Efficiency Performance Comparison Results

[0095] As shown in Table 4, the system efficiency indicators meet the needs of practical applications. The single recognition time is less than 3 seconds, supporting real-time risk identification. T+1 feature updates can be completed at night without affecting daytime use. Incremental training only takes 1.5 hours, which facilitates frequent model updates to adapt to new fraud patterns.

[0096] Ablation experiments show that removing any modality feature leads to a decrease in model performance. Among them, the correlation network feature has the greatest impact on model performance (AUC decreases by 0.03), followed by text features (AUC decreases by 0.02), which verifies the necessity of multimodal fusion and the contribution of each modality feature.

[0097] Experimental results fully demonstrate that the financial risk identification method based on multimodal data fusion proposed in this invention is significantly superior to existing technologies, possessing the following advantages: 1. High identification accuracy: AUC value reaches 0.95, an improvement of 6-20 percentage points compared to existing technologies; 2. Timely risk warning: Fraud risks are identified on average 5-8 quarters in advance; 3. Strong adaptability: The dynamic threshold mechanism effectively addresses industry differences and cyclical changes; 4. Efficiency meets practical needs: Single identification time is <3 seconds, supporting real-time monitoring for large-scale enterprises; Overall, this invention transforms the financial risk identification problem into a multimodal adaptation problem, integrates multi-source heterogeneous data, and adopts an advanced deep learning architecture and dynamic threshold adjustment mechanism, achieving a breakthrough in financial fraud identification from traditional manual judgment to data-driven intelligent decision-making, possessing significant theoretical and practical value.

[0098] On the other hand, the present invention also provides a computer storage medium storing executable program code; the executable program code is used to execute any of the above-mentioned financial risk identification methods.

[0099] On the other hand, the present invention also provides a computer system including a memory and a processor; the memory stores program code executable by the processor; the program code is used to execute any of the above-described financial risk identification methods.

[0100] For example, the program code can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the program code in the computer system.

[0101] The computer system may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer system may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the computer system may also include input / output devices, network access devices, buses, etc.

[0102] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0103] The memory can be an internal storage unit of a computer system, such as a hard disk or RAM. It can also be an external storage device of the computer system, such as a plug-in hard disk, SmartMediaCard (SMC), Secure Digital (SD) card, or FlashCard. Furthermore, the memory can include both internal and external storage units. The memory is used to store the program code and other programs and data required by the computer system. The memory can also be used to temporarily store data that has been output or will be output.

[0104] The computer storage medium and computer system described above are created based on the financial risk identification method described above. Their technical functions and beneficial effects will not be elaborated here. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A financial risk identification method based on multimodal data fusion, characterized in that, include: The structured data features, temporal variation features, unstructured text features, and relational network features of the enterprise to be identified are extracted to obtain multimodal data features; By inputting multimodal data features into the relationship graph inference model, risk identification results are obtained. The relational graph reasoning model includes: an input layer for receiving and processing multimodal data features to obtain model input; an encoding layer for performing deep encoding and feature extraction on the model input, mining the intrinsic relationships between features, and obtaining encoded features; and a graph neural network layer for combining the graph structure information of the relational network features to enhance the encoded features, strengthen the feature representation of enterprise relations, and obtain enhanced features. The output layer is used to output the risk identification results based on the enhanced features.

2. The financial risk identification method according to claim 1, characterized in that, The input layer processing steps include: The attention weights of temporal variation features on unstructured text features are calculated to obtain attention-fused temporal-text features; The structured data features, attention-fusion time-series-text features, and association network features are concatenated to obtain the fused features; Add positional and modality type encodings to the fused features to obtain the model input.

3. The financial risk identification method according to claim 2, characterized in that, The input layer processing steps also include: Different modal features in the fused features are mapped to the common semantic space through corresponding projections, and the mean square error of the features after different modal mappings is calculated to obtain the modal alignment loss; The model parameters are optimized based on the modal alignment loss.

4. The financial risk identification method according to claim 1, characterized in that, The encoding layer includes: several Transformer encoders; each Transformer encoder includes a multi-head attention module, a feedforward neural network, a residual connection, and a LayerNorm layer connected in sequence; the feedforward neural network contains two linear transformations and a GELU activation function.

5. The financial risk identification method according to claim 1, characterized in that, The graph neural network layer adopts a two-layer GCN structure; The first layer GCN is used to perform preliminary linear transformation and nonlinear activation processing on the encoded features output by the Transformer encoder based on the normalized adjacency matrix, thereby achieving preliminary feature enhancement. The second-layer GCN is used to further perform linear transformations and nonlinear activation processing on the preliminary enhanced features output by the first-layer GCN, based on the same normalized adjacency matrix, to complete the deep enhancement of features and output the final enhanced features.

6. The financial risk identification method according to claim 1, characterized in that, The output layer adopts a dual-branch output structure, wherein: The regression branch is used to predict the risk fit score; Classification branches are used to predict the probability distribution of risk levels.

7. The financial risk identification method according to claim 1, characterized in that, Before training the relation graph reasoning model, the steps for constructing the training dataset include: Obtain complete annual data for the sample companies to be labeled; Based on five indicators—financial indicator deviation, non-financial data consistency, time series anomaly, related-party transaction rationality, and textual semantic risk—the scores for each indicator of the sample companies to be labeled are determined. Specifically, financial indicator deviation measures the degree of deviation between a company's financial ratios and the industry's normal level; non-financial data consistency measures the degree of matching between non-financial operating data and financial data; time series anomaly measures the degree of abnormal fluctuations in the time series of financial indicators; related-party transaction rationality measures the degree of deviation between related-party transaction prices and fair market prices; and textual semantic risk measures the strength of risk signals in the company's publicly disclosed textual information. By weighting the scores of each indicator, the risk suitability score of each labeled sample enterprise is obtained.

8. The financial risk identification method according to claim 7, characterized in that, In the training of the relational graph inference model, the total loss function is determined by a weighted average of the main regression loss, auxiliary classification loss, feature reconstruction loss, and association consistency loss; among which, The main regression loss is used to calculate the mean squared error between the fit score and the labeled score. Auxiliary classification loss, used to calculate the weighted cross-entropy loss for risk level classification; Feature reconstruction loss is used to reconstruct the original features of each modality from the fused features, and to calculate the mean square error between the reconstructed modal features and the corresponding original modal features; Association consistency loss is used to calculate the mean squared error between the financial features predicted by non-financial features and the original financial features.

9. The financial risk identification method according to any one of claims 1 to 8, characterized in that, The steps for extracting structured data features include: calculating the financial ratio features of the latest quarter and constructing a multidimensional financial ratio feature vector; calculating the benchmarking indicators between the company and the industry average and constructing a multidimensional industry benchmarking feature vector; calculating the compound annual growth rate (CAGR) indicator over the past few years and constructing a multidimensional growth rate feature vector; calculating the financial risk warning indicator and constructing a multidimensional risk warning feature vector; and concatenating the multidimensional financial ratio feature vector, the multidimensional industry benchmarking feature vector, the multidimensional growth rate feature vector, and the multidimensional risk warning feature vector to obtain the structured data features. Or / and, the steps for extracting time-series change features include: obtaining the time-series matrix of financial indicators of the enterprise to be identified; detecting abnormal time points in the time-series matrix of financial indicators; encoding the location information of abnormal time points into binary vectors matching the length of the time series; concatenating the binary vectors to the feature dimensions of the time-series matrix of financial indicators to form a time-series matrix of financial indicators that incorporates abnormal location information; and inputting the time-series matrix of financial indicators that incorporates abnormal location information into a two-layer LSTM network to extract time-series change features. The steps for extracting unstructured text features include: obtaining a text set of the enterprise to be identified; extracting the embedding vectors of each text segment in the text set, and performing attention fusion on the embedding vectors of each text segment to obtain text fusion features; reducing the dimensionality of the text fusion features and adding text statistical features to obtain unstructured text features. The steps for extracting network features include: constructing an enterprise relationship graph; the nodes in the graph include any one or more of enterprises, legal persons, shareholders, and senior executives; the relationships include any one or more of investment, employment, competition, cooperation, supply chain, and associated risks; based on the enterprise relationship graph, extracting node centrality features, community centrality features of nodes in their respective communities, equity structure features, and senior executive association features, and concatenating them to obtain network features.

10. A computer system, characterized in that, It includes a memory and a processor; the memory stores program code that can be executed by the processor; the program code is used to perform the financial risk identification method according to any one of claims 1 to 9.

Images