Financial risk identification system based on artificial intelligence

Abstract

The invention discloses a financial risk identification system based on artificial intelligence. The system comprises a data acquisition module, a data preprocessing module, a risk perception model construction module, a hyper-parameter optimization module and a financial risk identification module. The invention relates to the technical field of financial risk data processing, in particular to a financial risk identification system based on artificial intelligence, and the method comprises the steps: obtaining financial risk original data through data collection; a data preprocessing method of data cleaning, data coding, data normalization and data set segmentation is adopted; a composite graph neural network model is adopted to carry out financial risk identification, and through multi-modal data fusion, an interpretable risk cause path is generated, and model transparency and decision support capability are enhanced; according to the method, model hyper-parameter optimization is carried out by adopting an optimization algorithm, a global optimal solution can be efficiently searched, and the model can fully mine a potential mode in financial risk data, so that more accurate identification is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of financial risk data processing, and specifically refers to a financial risk identification system based on artificial intelligence. Background Art

[0002] Intelligent financial risk identification refers to the deep mining and analysis of enterprise financial data by using technologies such as artificial intelligence, big data analysis, and machine learning, so as to automatically identify potential financial risks; it can help enterprises timely discover problems such as financial loopholes, fraud, and abnormal capital flows, improve the accuracy and efficiency of financial management, and ensure the stability and health of the enterprise's financial condition.

[0003] However, traditional financial risk identification methods usually rely on a single data source and static rules, and cannot effectively capture multi-dimensional risk factors such as enterprise relationships, market dynamics, and text information. Moreover, they lack interpretability and cannot provide detailed paths for risk causes, resulting in the technical problem that decision-makers are difficult to understand the model output and take targeted measures; traditional financial risk identification methods have technical problems when optimizing model parameters. It is difficult to adapt to high-dimensional and complex financial data, and in a complex parameter space, it is impossible to fully explore the global optimal solution, resulting in insufficient model performance. Summary of the Invention

[0004] In view of the above situation, to overcome the defects of the prior art, the present invention provides a financial risk identification system based on artificial intelligence. Aiming at the technical problems that traditional financial risk identification methods usually rely on a single data source and static rules, cannot effectively capture multi-dimensional risk factors such as enterprise relationships, market dynamics, and text information, and lack interpretability, cannot provide detailed paths for risk causes, resulting in decision-makers being difficult to understand the model output and take targeted measures, this solution creatively uses a composite graph neural network model for financial risk identification. Through multi-modal data fusion, it comprehensively captures internal and external risk factors of the enterprise and generates an interpretable risk cause path, enhancing the model transparency and decision support ability; aiming at the technical problems that traditional financial risk identification methods have when optimizing model parameters, it is difficult to adapt to high-dimensional and complex financial data, and in a complex parameter space, it is impossible to fully explore the global optimal solution, resulting in insufficient model performance, this solution creatively uses an optimization algorithm to optimize the model hyperparameters, which can efficiently find the global optimal solution in a multi-dimensional parameter space and is not easily trapped in a local optimal solution, enabling the model to fully mine the potential patterns in financial risk data, thereby achieving more accurate identification.

[0005] The technical solution adopted by the present invention is as follows: The financial risk identification system based on artificial intelligence provided by the present invention includes a data collection module, a data preprocessing module, a risk perception model construction module, a hyperparameter optimization module, and a financial risk identification module;

[0006] The data collection module obtains the original data set of financial risk by collecting data from the enterprise financial management system;

[0007] The data preprocessing module performs preprocessing by data cleaning, data encoding, data normalization and data set segmentation to obtain a prediction training set, a prediction test set and a current financial data set;

[0008] The risk perception model construction module obtains the risk perception model by constructing a composite graph neural network model and using it as the risk perception model;

[0009] The hyperparameter optimization module optimizes the hyperparameters of the risk perception model by using an optimization algorithm to obtain an optimized risk perception model;

[0010] The financial risk identification module uses the optimized risk perception model to perform financial risk identification and obtain risk perception reference data.

[0011] Furthermore, in the data acquisition module, the financial risk original data set specifically includes a historical financial original data set and a current financial original data set, and the historical financial original data set and the current financial original data set specifically include financial indicator data, enterprise relationship data and macro-industry data, and the historical financial original data set also includes financial risk label data, and the financial indicator data specifically include asset and liability data, profit data and financial flow data, and the enterprise relationship data specifically include supply chain relationship data, equity relationship data, peer competition data, cooperation years data and cooperation share data, and the macro-industry data specifically include industry index data, industry growth rate data and industry interest rate change data, and the financial risk labels specifically include low risk and high risk.

[0012] Furthermore, in the data preprocessing module, the data cleaning is specifically to remove missing values ​​and duplicate values, the data encoding is specifically to encode the original data using the one-hot encoding method, the data normalization is specifically to normalize the original data using the minimum-maximum normalization method, and the data set segmentation is specifically to segment the historical financial original data set after the data cleaning, data encoding and data normalization into a prediction training set and a prediction test set. Through data preprocessing, the prediction training set, the prediction test set and the current financial data set are obtained.

[0013] Further, in the risk perception model construction module, it is used to construct the model required for financial risk identification. Specifically, by constructing a composite graph neural network model and using it as the risk perception model, the risk perception model is obtained. The composite graph neural network model specifically includes a causal feature selection sub-module, a temporal convolution sub-module, a dual-branch graph neural network sub-module, a risk rule generation sub-module, and an output sub-module;

[0014] The specific content of the risk perception model construction module includes the construction of the causal feature selection sub-module, the construction of the temporal convolution sub-module, the construction of the dual-branch graph neural network sub-module, the construction of the risk rule generation sub-module, the design of the loss function, and the construction and training of the model;

[0015] The construction of the causal feature selection sub-module is used to automatically determine the causally related features through a data-driven method. Specifically, the structure equation model optimization algorithm of the acyclic directed graph is used to select features. The content includes:

[0016] Obtain the causal weight matrix, and the formula used is as follows:

[0017] ;

[0018] In the formula, represents the function value to be optimized, represents the algorithm input data matrix, represents the causal weight matrix, represents the L1 regularization coefficient used to control the sparsity of the causal graph, represents the calculation of the Frobenius norm, represents the calculation of the L1 norm;

[0019] Dynamic feature selection, and the formula used is as follows:

[0020] ;

[0021] In the formula, represents the masked weight matrix, represents the indicator function. If the value of the element of the causal weight matrix is greater than , the value is 1. If the value of the element of the causal weight matrix is less than , the value is 0, represents the output features of the causal feature selection sub-module, represents the input data of the causal feature selection sub-module, represents the matrix diagonal element extraction function, represents the element-wise multiplication;

[0022] The construction of the temporal convolution sub-module is used to capture multi-scale temporal patterns of financial data, including:

[0023] Parallel multi-scale dilated convolution, with the formula as follows:

[0024] ;

[0025] In the formula, represents the output feature of the k-th convolutional layer, represents the ReLU activation function, represents a one-dimensional convolutional function with a kernel size of 3 and a dilation rate of , and k represents the convolutional layer index;

[0026] Multi-scale feature fusion, with the formula as follows:

[0027] ;

[0028] In the formula, Hs represents the multi-scale fusion feature, represents the concatenation operation function, represents the output feature of the 0-th convolutional layer, represents the output feature of the 1-st convolutional layer, represents the output feature of the 2-nd convolutional layer;

[0029] Adaptive feature fusion, with the formula as follows:

[0030] ;

[0031] In the formula, represents the attention feature, represents the softmax function, represents the temporal convolution learnable attention matrix, represents the output feature of the temporal convolution sub-module, and ci represents the control coefficient;

[0032] The construction of the dual-branch graph neural network sub-module is used to analyze the impacts of micro-enterprise relationships and macro-industry conditions, including:

[0033] The construction of the micro-branch, including:

[0034] Define the set of micro-graph relationships, with the formula as follows:

[0035] ;

[0036] In the formula, MiG represents the set of micro-graph relationships, Su represents the supply chain relationship, Ec represents the equity association relationship, Cr represents the competition relationship, and Pa represents the cooperation relationship;

[0037] Specific relational graph convolution, specifically performing graph convolution independently for each relationship, with the formula as follows:

[0038] ;

[0039] In the formula, represents the output feature of relationship r at the (l + 1)-th layer, represents the degree matrix of relationship r, represents the adjacency matrix of relationship r, represents the output feature of relationship r at the l-th layer, represents the learnable weight of relationship r at the l-th layer;

[0040] Cross-relationship aggregation, with the formula as follows:

[0041] ;

[0042] In the formula, H_mi represents the output feature of the micro-branch, represents the learnable query vector of the micro-branch, represents the output feature of relationship r at the L-th layer, where L represents the number of graph convolution layers of the micro-branch;

[0043] Macro-branch construction, including:

[0044] Dynamic adjacency matrix construction, specifically calculating industry correlation based on a time sliding window, with the formula as follows:

[0045] ;

[0046] In the formula, represents the element at the i-th row and j-th column of the dynamic adjacency matrix of the macro-branch, represents the Pearson correlation coefficient calculation function, represents the feature of node i within the time window , represents the feature of node j within the time window , represents the size of the time window;

[0047] Dynamic attention calculation, with the formula as follows:

[0048] ;

[0049] In the formula, represents the attention coefficient of node i to node j, represents the LeakyReLU activation function, La represents the learnable attention vector of the macro-branch, represents the learnable attention matrix of the macro-branch, represents the feature of node i, Represents the features of node j;

[0050] Feature update, using the following formula:

[0051] ;

[0052] In the formula, Represents the features of node i at the (l + 1)-th layer, Represents the neighbor set of node i, Represents the learnable weight matrix for feature update, Represents the features of node i at the l-th layer;

[0053] Dual-branch feature aggregation, using the following formula:

[0054] ;

[0055] In the formula, Represents the output features of the dual-branch graph neural network sub-module, Represents the output features of the macro branch;

[0056] The risk rule generation sub-module is constructed to automatically generate interpretable financial risk rules. Specifically, a neural rule engine is used to generate interpretable rules, including:

[0057] Rule condition learning. Specifically, a thresholded fully connected layer is used to generate rule conditions, using the following formula:

[0058] ;

[0059] In the formula, Represents the num-th rule, Represents the weight of the num-th rule for the features, Represents the threshold of the num-th rule;

[0060] Calculate the rule condition score, using the following formula:

[0061] ;

[0062] In the formula, Represents the rule condition score of the num-th rule, Represents the sigmoid function;

[0063] Calculate the rule importance score, using the following formula:

[0064] ;

[0065] In the formula, Represents the rule importance score of the num-th rule, Indicates the learnable weight vector of the num-th rule;

[0066] Obtain the rule-based risk prediction value, and the formula used is as follows:

[0067] ;

[0068] In the formula, Indicates the rule-based risk prediction value, Num indicates the total number of rules;

[0069] Generate rule explanations, and the formula used is as follows:

[0070] ;

[0071] In the formula, Indicates the feature importance score of the q-th feature, Indicates the element in the p-th row and q-th column of the causal weight matrix, Ex indicates the generated explanation, Indicates the rule explanation generation function, Indicates the feature explanation generation function, Indicates sorting in descending order and taking the top K values, Indicates the rule importance score of the first rule, Indicates the rule importance score of the second rule, Indicates the rule importance score of the Num-th rule, Indicates the feature importance score of the first feature, Indicates the feature importance score of the second feature, Indicates the feature importance score of the Q-th feature, Q indicates the total number of features;

[0072] The designed loss function is specifically to design the model loss function by combining cross-entropy loss and rule consistency loss, and the formula used is as follows:

[0073] ;

[0074] In the formula, Indicates the model risk level prediction value, Indicates the model output weight, Indicates the model output bias term, Indicates the cross-entropy loss value, Sa indicates the total number of samples, Indicates the true value of the sa-th sample, Indicates the model risk level prediction value of the sa-th sample, Indicates the rule consistency loss value, Indicates calculating the L2 norm, Loss indicates the model loss value;

[0075] The construction and training of the model specifically involve constructing a composite graph neural network model through the construction of the causal feature selection sub-module, the construction of the temporal convolution sub-module, the construction of the dual-branch graph neural network sub-module, the construction of the risk rule generation sub-module, and the design of the loss function, training the model based on the prediction training set, validating the model performance based on the prediction test set, obtaining the composite graph neural network model, and using it as the risk perception model.

[0076] Further, in the hyperparameter optimization module, it is used to optimize the model hyperparameters. Specifically, it uses an optimization algorithm to optimize the model hyperparameters of the risk perception model, obtaining an optimized risk perception model. The specific content includes algorithm initialization, global search, local search, determining the global optimal solution, and optimizing the model hyperparameters.

[0077] The algorithm initialization specifically involves initializing the search space and constructing an individual unit set. The individual unit is used to represent the hyperparameter combination of the risk perception model, and the fitness function of the individual unit is the loss function of the risk perception model.

[0078] The formula for the global search is as follows:

[0079] ;

[0080] In the formula, represents the velocity of the a-th individual unit at the (dt + 1)-th iteration, represents the inertia weight, represents the velocity of the a-th individual unit at the dt-th iteration, represents the weight used to control learning its own optimal solution, represents the own optimal solution of the a-th individual unit, represents the position of the a-th individual unit at the dt-th iteration, represents the weight used to control learning the global optimal solution, represents the global optimal solution at the dt-th iteration, represents the position of the a-th individual unit at the (dt + 1)-th iteration;

[0081] The local search includes:

[0082] Calculate the probability of selecting a neighboring solution. The formula is as follows:

[0083] ;

[0084] In the formula, represents the probability that the a-th individual unit selects the b-th neighboring individual unit, represents the fitness calculation function, and N represents the number of neighboring individual units. Denote the position of the b-th neighbor individual unit at the (dt + 1)-th iteration;

[0085] Design the neighbor solution learning strategy. Specifically, the individual unit selects a neighbor individual unit based on the neighbor solution selection probability and optimizes its own position based on the position of the selected neighbor individual unit. The formula used is as follows:

[0086] ;

[0087] In the formula, Denote the position of the a-th individual unit at the (dt + 1)-th iteration in the local search, de represents a decay factor that linearly decreases from 2 to 1, and rand represents a random number within the range [0, 1];

[0088] Design the neighbor solution comparison strategy. Specifically, after the individual unit optimizes its own position based on the position of the neighbor individual unit, it selects a neighbor individual unit again based on the neighbor solution selection probability and compares the fitness with the selected neighbor individual unit to determine whether to update the position. The formula used is as follows:

[0089] ;

[0090] In the formula, Denote the fitness comparison function, Denote the final position of the a-th individual unit at the (dt + 1)-th iteration, Denote the position of the c-th neighbor individual unit at the (dt + 1)-th iteration;

[0091] The determination of the global optimal solution is specifically to continuously iterate and update until the iteration termination condition of the optimization algorithm is reached, and take the position of the individual unit with the lowest fitness as the global optimal solution. The global optimal solution is specifically the optimal risk perception model hyperparameter combination, and the iteration termination condition of the optimization algorithm specifically includes that the fitness of the individual unit is less than the set threshold and the maximum number of iterations is reached;

[0092] The optimization of the model hyperparameters is specifically to optimize the hyperparameters of the risk perception model through the algorithm initialization, the global search, the local search, and the determination of the global optimal solution to obtain an optimized risk perception model.

[0093] Furthermore, in the financial risk identification module, specifically, use the current financial data set as the input of the optimized risk perception model to perform financial risk identification to obtain risk perception reference data, and based on the risk perception reference data, comprehensively evaluate the occurrence of financial risks.

[0094] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0095] (1) In view of the technical problems that traditional financial risk identification methods usually rely on a single data source and static rules, cannot effectively capture multi-dimensional risk factors such as inter-enterprise relationships, market dynamics and text information, and lack interpretability, cannot provide a detailed path of risk causes, and make it difficult for decision makers to understand the model output and take targeted measures, this solution creatively adopts a composite graph neural network model for financial risk identification. Through multimodal data fusion, it comprehensively captures internal and external risk factors of the enterprise, generates an interpretable risk cause path, and enhances the model transparency and decision support capabilities.

[0096] (2) Traditional financial risk identification methods have the problem that they are difficult to adapt to high-dimensional and complex financial data when optimizing model parameters. In complex parameter spaces, they are unable to fully explore the global optimal solution, resulting in insufficient model performance. This solution creatively uses an optimization algorithm to optimize model hyperparameters. It can efficiently find the global optimal solution in a multi-dimensional parameter space and is not easily trapped in a local optimal solution. This enables the model to fully explore the potential patterns in financial risk data, thereby achieving more accurate identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 A schematic diagram of the modules of the artificial intelligence-based financial risk identification system provided by the present invention;

[0098] Figure 2 It is a flowchart of the data preprocessing module;

[0099] Figure 3 Schematic diagram of the process of building modules for the risk perception model;

[0100] Figure 4 This is a flowchart of the hyperparameter optimization module;

[0101] Figure 5 This is the loss change curve of the training process;

[0102] Figure 6 It is the accuracy change curve of the training process;

[0103] Figure 7 is the confusion matrix of classification results;

[0104] Figure 8 ROC curve for model performance.

[0105] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0106] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0107] In the description of the present invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc. indicating the orientation or positional relationship are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention.

[0108] Embodiment 1, refer to Figure 1 , the financial risk identification system based on artificial intelligence provided by the present invention includes a data acquisition module, a data preprocessing module, a risk perception model construction module, a hyperparameter optimization module, and a financial risk identification module;

[0109] The data acquisition module obtains the original financial risk data set by collecting data from the enterprise financial management system;

[0110] The data preprocessing module performs preprocessing by data cleaning, data encoding, data normalization, and data set segmentation to obtain a prediction training set, a prediction test set, and the current financial data set;

[0111] The risk perception model construction module obtains the risk perception model by constructing a composite graph neural network model and using it as the risk perception model;

[0112] The hyperparameter optimization module performs hyperparameter optimization on the risk perception model by using [specific method not provided in the original text] to obtain the optimized risk perception model;

[0113] The financial risk identification module performs financial risk identification by using the optimized risk perception model to obtain the risk perception reference data.

[0114] Embodiment 2, refer to Figure 1In the data acquisition module, the financial risk original data set specifically includes a historical financial original data set and a current financial original data set. The historical financial original data set and the current financial original data set specifically include financial indicator data, enterprise relationship data and macro-industry data. The historical financial original data set also includes financial risk label data. The financial indicator data specifically includes asset and liability data, profit data and financial flow data. The enterprise relationship data specifically includes supply chain relationship data, equity-related data, peer competition data, cooperation years data and cooperation share data. The macro-industry data specifically includes industry index data, industry growth rate data and industry interest rate change data. Financial risk labels specifically include low risk and high risk.

[0115] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the data preprocessing module, the data cleaning is specifically to remove missing values ​​and duplicate values, the data encoding is specifically to encode the original data using the one-hot encoding method, the data normalization is specifically to normalize the original data using the minimum-maximum normalization method, and the data set segmentation is specifically to segment the historical financial original data set after the data cleaning, data encoding and data normalization into a prediction training set and a prediction test set. Through data preprocessing, the prediction training set, the prediction test set and the current financial data set are obtained.

[0116] Example 4, see Figure 1 , Figure 3 , Figure 5 , Figure 6 , Figure 7 and Figure 8 This embodiment is based on the above embodiment. In the risk perception model construction module, it is used to construct a model required for financial risk identification. Specifically, a composite graph neural network model is constructed and used as a risk perception model to obtain a risk perception model. The composite graph neural network model specifically includes a causal feature selection submodule, a time convolution submodule, a dual-branch graph neural network submodule, a risk rule generation submodule and an output submodule.

[0117] The risk perception model construction module specifically includes the construction of a causal feature selection submodule, a time convolution submodule, a dual-branch graph neural network submodule, a risk rule generation submodule, a loss function design, and the construction and training of a model;

[0118] The causal feature selection submodule is constructed to automatically determine causal related features through a data-driven method, specifically using a structural equation model optimization algorithm of an acyclic directed graph to select features, including:

[0119] Obtain the causal weight matrix, and the formula used is as follows:

[0120] ;

[0121] In the formula, represents the function value to be optimized, represents the algorithm input data matrix, represents the causal weight matrix, represents the L1 regularization coefficient used to control the sparsity of the causal graph, represents the calculation of the Frobenius norm, represents the calculation of the L1 norm;

[0122] For dynamic feature selection, the formula used is as follows:

[0123] ;

[0124] In the formula, represents the mask weight matrix, represents the indicator function. If the value of an element of the causal weight matrix is greater than , the value is 1. If the value of an element of the causal weight matrix is less than , the value is 0, represents the output feature of the causal feature selection sub-module, represents the input data of the causal feature selection sub-module, represents the matrix diagonal element extraction function, represents element-wise multiplication;

[0125] The time convolution sub-module is constructed to capture the multi-scale time series patterns of financial data, and the content includes:

[0126] For parallel multi-scale dilated convolution, the formula used is as follows:

[0127] ;

[0128] In the formula, represents the output feature of the k-th convolutional layer, represents the ReLU activation function, represents a one-dimensional convolutional function with a convolutional kernel size of 3 and a dilation rate of , and k represents the convolutional layer index;

[0129] For multi-scale feature fusion, the formula used is as follows:

[0130] ;

[0131] In the formula, Hs represents the multi-scale fusion feature, Denote the splicing operation function, Denote the output features of the 0th convolutional layer, Denote the output features of the 1st convolutional layer, Denote the output features of the 2nd convolutional layer;

[0132] Adaptive feature fusion, and the formula used is as follows:

[0133] ;

[0134] In the formula, Denote the attention features, Denote the softmax function, Denote the time convolutional learnable attention matrix, Denote the output features of the time convolutional sub-module, and ci denotes the control coefficient;

[0135] The construction of the double-branch graph neural network sub-module is used to analyze the impacts of micro-enterprise relationships and macro-industry conditions, and the content includes:

[0136] The construction of the micro-branch, and the content includes:

[0137] Define the set of micro-graph relationships, and the formula used is as follows:

[0138] ;

[0139] In the formula, MiG denotes the set of micro-graph relationships, Su denotes the supply chain relationship, Ec denotes the equity association relationship, Cr denotes the competition relationship, and Pa denotes the cooperation relationship;

[0140] Specific relationship graph convolution, specifically, perform graph convolution on each relationship independently, and the formula used is as follows:

[0141] ;

[0142] In the formula, Denote the output features of relationship r at the l+1 layer, Denote the degree matrix of relationship r, Denote the adjacency matrix of relationship r, Denote the output features of relationship r at the l layer, Denote the learnable weight of relationship r at the l layer;

[0143] Cross-relationship aggregation, and the formula used is as follows:

[0144] ;

[0145] In the formula, H_mi denotes the output features of the micro-branch, Denote the learnable query vector of the micro-branch, Denote the output feature of the relationship r at the L-th layer, where L represents the number of layers of the micro-branch graph convolution;

[0146] Macro-branch construction, including:

[0147] Construction of the dynamic adjacency matrix, specifically calculating the industry correlation based on a time sliding window, and the formula used is as follows:

[0148] ;

[0149] In the formula, Denote the element in the i-th row and j-th column of the dynamic adjacency matrix of the macro-branch, Denote the Pearson correlation coefficient calculation function, Denote the feature of node i within the time window ; Denote the feature of node j within the time window ; Denote the size of the time window;

[0150] Dynamic attention calculation, and the formula used is as follows:

[0151] ;

[0152] In the formula, Denote the attention coefficient of node i to node j, Denote the LeakyReLU activation function, La denote the learnable attention vector of the macro-branch, Denote the learnable attention matrix of the macro-branch, Denote the feature of node i, Denote the feature of node j;

[0153] Feature update, and the formula used is as follows:

[0154] ;

[0155] In the formula, Denote the feature of node i at the (l + 1)-th layer, Denote the neighbor set of node i, Denote the learnable weight matrix for feature update, Denote the feature of node i at the l-th layer;

[0156] Dual-branch feature aggregation, and the formula used is as follows:

[0157] ;

[0158] In the formula, Denote the output feature of the dual-branch graph neural network sub-module, Denote the output feature of the macro-branch;

[0159] The risk rule generation sub-module is constructed to automatically generate interpretable financial risk rules. Specifically, a neural rule engine is used to generate interpretable rules, and the content includes:

[0160] Rule condition learning, specifically using a thresholded fully connected layer to generate rule conditions. The formula used is as follows:

[0161] ;

[0162] In the formula, represents the num-th rule, represents the weight of the num-th rule for the feature, represents the threshold of the num-th rule;

[0163] Calculate the rule condition score. The formula used is as follows:

[0164] ;

[0165] In the formula, represents the rule condition score of the num-th rule, represents the sigmoid function;

[0166] Calculate the rule importance score. The formula used is as follows:

[0167] ;

[0168] In the formula, represents the rule importance score of the num-th rule, represents the learnable weight vector of the num-th rule;

[0169] Obtain the rule-based risk prediction value. The formula used is as follows:

[0170] ;

[0171] In the formula, represents the rule-based risk prediction value, Num represents the total number of rules;

[0172] Generate rule explanations. The formula used is as follows:

[0173] ;

[0174] In the formula, represents the feature importance score of the q-th feature, represents the element in the p-th row and q-th column of the causal weight matrix, Ex represents the generated explanation, represents the rule explanation generation function, represents the feature explanation generation function, Indicates descending order and taking the top K values, Indicates the rule importance score of the first rule, Indicates the rule importance score of the second rule, Indicates the rule importance score of the Num-th rule, Indicates the feature importance score of the first feature, Indicates the feature importance score of the second feature, Indicates the feature importance score of the Q-th feature, where Q represents the total number of features;

[0175] The described design loss function is specifically a model loss function designed by combining cross-entropy loss and rule consistency loss. The formula used is as follows:

[0176] ;

[0177] In the formula, Indicates the predicted value of the model risk level, Indicates the output weight of the model, Indicates the output bias term of the model, Indicates the cross-entropy loss value, Sa represents the total number of samples, Indicates the true value of the sa-th sample, Indicates the predicted value of the model risk level of the sa-th sample, Indicates the rule consistency loss value, Indicates calculating the L2 norm, Loss represents the model loss value;

[0178] The described construction and training of the model is specifically to construct a composite graph neural network model through the construction of the causal feature selection sub-module, the construction of the temporal convolution sub-module, the construction of the dual-branch graph neural network sub-module, the construction of the risk rule generation sub-module, and the design loss function, and train the model based on the prediction training set, verify the model performance based on the prediction test set, obtain the composite graph neural network model, and use it as the risk perception model.

[0179] By performing the above operations, for the technical problem that traditional financial risk identification methods usually rely on a single data source and static rules, cannot effectively capture multi-dimensional risk factors such as inter-enterprise relationships, market dynamics, and text information, and lack interpretability and cannot provide detailed paths for risk causes, resulting in decision-makers being difficult to understand the model output and take targeted measures, this solution creatively uses a composite graph neural network model for financial risk identification. Through multi-modal data fusion, it comprehensively captures internal and external risk factors of enterprises and generates interpretable risk cause paths, enhancing model transparency and decision support capabilities.

[0180] Example Five, refer toFigure 1 and Figure 4 This embodiment is based on the above embodiment. In the hyperparameter optimization module, it is used to optimize the model hyperparameters. Specifically, an optimization algorithm is adopted to optimize the model hyperparameters of the risk perception model, and an optimized risk perception model is obtained. The specific content includes algorithm initialization, global search, local search, determining the global optimal solution, and optimizing the model hyperparameters;

[0181] The algorithm initialization is specifically to initialize the search space and construct an individual unit set. The individual unit is used to represent the hyperparameter combination of the risk perception model, and the fitness function of the individual unit is the loss function of the risk perception model;

[0182] The global search uses the following formula:

[0183] ;

[0184] In the formula, represents the velocity of the a-th individual unit at the (dt + 1)-th iteration, represents the inertia weight, represents the velocity of the a-th individual unit at the dt-th iteration, represents the weight for controlling learning its own optimal solution, represents the own optimal solution of the a-th individual unit, represents the position of the a-th individual unit at the dt-th iteration, represents the weight for controlling learning the global optimal solution, represents the global optimal solution at the dt-th iteration, represents the position of the a-th individual unit at the (dt + 1)-th iteration;

[0185] The local search includes:

[0186] Calculate the probability of selecting a neighboring solution. The formula used is as follows:

[0187] ;

[0188] In the formula, represents the probability that the a-th individual unit selects the b-th neighboring individual unit, represents the fitness calculation function, N represents the number of neighboring individual units, represents the position of the b-th neighboring individual unit at the (dt + 1)-th iteration;

[0189] Design a neighboring solution learning strategy. Specifically, the individual unit selects a neighboring individual unit based on the probability of selecting a neighboring solution, and optimizes its own position based on the position of the selected neighboring individual unit. The formula used is as follows:

[0190] ;

[0191] In the formula, represents the position of the ath individual unit at the dt+1th iteration in the local search, de represents the decay factor whose size decreases linearly from 2 to 1, and rand represents a random number in the range [0,1];

[0192] Design a neighbor solution comparison strategy. Specifically, after the individual unit optimizes its own position based on the position of the neighbor individual unit, it selects the neighbor individual unit again based on the neighbor solution selection probability, and compares the fitness with the selected neighbor individual unit to determine whether to update the position. The formula used is as follows:

[0193] ;

[0194] In the formula, represents the fitness comparison function, represents the final position of the ath individual unit at the dt+1th iteration, represents the position of the cth neighbor individual unit at the dt+1th iteration;

[0195] The determination of the global optimal solution is specifically to continuously iterate and update until the optimization algorithm iteration termination condition is reached, and the position of the individual unit with the lowest fitness is taken as the global optimal solution. The global optimal solution is specifically the optimal risk perception model hyperparameter combination. The optimization algorithm iteration termination condition specifically includes that the individual unit fitness is less than a set threshold and reaches the maximum number of iterations;

[0196] The optimization model hyperparameters are specifically performed by optimizing the hyperparameters of the risk perception model through the algorithm initialization, the global search, the local search and the determination of the global optimal solution to obtain an optimized risk perception model.

[0197] By performing the above operations, the traditional financial risk identification method has the technical problem of being difficult to adapt to high-dimensional and complex financial data when optimizing model parameters, and being unable to fully explore the global optimal solution in the complex parameter space, resulting in insufficient model performance. This solution creatively uses an optimization algorithm to optimize the model hyperparameters. It can efficiently find the global optimal solution in the multi-dimensional parameter space and is not easily trapped in the local optimal solution, so that the model can fully explore the potential patterns in the financial risk data, thereby achieving more accurate identification.

[0198] Example 6, see Figure 1 This embodiment is based on the above embodiment. In the financial risk identification module, specifically, the current financial data set is used as the input of the optimized risk perception model to perform financial risk identification, obtain risk perception reference data, and comprehensively judge the occurrence of financial risks based on the risk perception reference data.

[0199] It should be noted that, in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device.

[0200] Although the embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention.

[0201] The above description of the present invention and its embodiments is not restrictive. What is shown in the drawings is only one of the embodiments of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and, without departing from the purpose of the present invention, design similar structural modes and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. Financial risk identification system based on artificial intelligence, characterized by: The system includes a data acquisition module, a data preprocessing module, a risk perception model building module, a hyperparameter optimization module and a financial risk identification module; The data collection module obtains a financial risk original data set through data collection, and the financial risk original data set specifically includes a historical financial original data set and a current financial original data set; The data preprocessing module obtains a prediction training set, a prediction test set and a current financial data set through data preprocessing; The risk perception model construction module obtains the risk perception model by constructing a composite graph neural network model and using it as the risk perception model. The composite graph neural network model specifically includes a causal feature selection submodule, a time convolution submodule, a dual-branch graph neural network submodule, a risk rule generation submodule and an output submodule; The hyperparameter optimization module optimizes the hyperparameters of the risk perception model by using an optimization algorithm to obtain an optimized risk perception model; The financial risk identification module uses the optimized risk perception model to perform financial risk identification and obtain risk perception reference data.

2. The artificial intelligence-based financial risk identification system according to claim 1, characterized in that: The risk perception model construction module specifically includes the construction of a causal feature selection submodule, a time convolution submodule, a dual-branch graph neural network submodule, a risk rule generation submodule, a loss function design, and the construction and training of a model; The causal feature selection submodule is constructed to automatically determine causal related features through a data-driven method, specifically using a structural equation model optimization algorithm of an acyclic directed graph to select features, including: To obtain the causal weight matrix, the formula used is as follows: ； In the formula, represents the value of the function to be optimized, represents the algorithm input data matrix, represents the causal weight matrix, represents the L1 regularization coefficient used to control the sparsity of the causal graph, Indicates the calculation of the Frobenius norm, Indicates calculation of L1 norm; Dynamic feature selection, the formula used is as follows: ； In the formula, represents the mask weight matrix, represents the indicator function, if the causal weight matrix The value of the element is greater than , then the value is 1, if the causal weight matrix The value of the element is less than , then the value is 0, represents the output feature of the causal feature selection submodule, represents the input data of the causal feature selection submodule, represents the matrix diagonal element extraction function, represents element-wise multiplication; The temporal convolution submodule is constructed to capture the multi-scale temporal patterns of financial data, including: Parallel multi-scale dilated convolution, the formula used is as follows: ； In the formula, represents the output features of the kth convolutional layer, represents the ReLU activation function, Indicates that the convolution kernel size is 3 and the expansion rate is One-dimensional convolution function, k represents the convolution layer index; Multi-scale feature fusion, the formula used is as follows: ； In the formula, Hs represents the multi-scale fusion feature, Represents the splicing operation function, represents the output features of the 0th convolutional layer, represents the output features of the first convolutional layer, Represents the output features of the second convolutional layer; Adaptive feature fusion, the formula used is as follows: ； In the formula, Represents the attention feature, represents the softmax function, represents the temporal convolutional learnable attention matrix, represents the output feature of the temporal convolution submodule, and ci represents the control coefficient; The dual-branch graph neural network submodule is constructed to analyze the impact of micro-enterprise relationships and macro-industry conditions, including: Micro-branch construction, including: Define the micrograph relationship set, the formula used is as follows: ； In the formula, MiG represents the micrograph relationship set, Su represents the supply chain relationship, Ec represents the equity relationship, Cr represents the competitive relationship, and Pa represents the cooperative relationship; Specific relationship graph convolution, specifically, performing graph convolution on each relationship independently, the formula used is as follows: ； In the formula, represents the output features of the l+1th layer relation r, represents the degree matrix of relation r, represents the adjacency matrix of relation r, represents the output features of the l-th layer relation r, represents the learnable weight of the l-th layer relation r; Cross-relationship aggregation, the formula used is as follows: ； Where H_mi represents the micro-branch output characteristics, represents the query vector that can be learned by micro-branch, represents the output features of the Lth layer relation r, where L represents the number of micro-branch graph convolution layers; Macro branch construction, including: Dynamic adjacency matrix construction, specifically calculating industry relevance based on a time sliding window, uses the following formula: ； In the formula, represents the element in the i-th row and j-th column of the macro-branch dynamic adjacency matrix, represents the Pearson correlation coefficient calculation function, Indicates that node i is in the time window The characteristics within Indicates that node j is in the time window The characteristics within Indicates the time window size; Dynamic attention calculation, the formula used is as follows: ； In the formula, represents the attention coefficient of node i to node j, represents the LeakyReLU activation function, La represents the macro branch learnable attention vector, denotes that the macro branch can learn the attention matrix, represents the characteristics of node i, represents the characteristics of node j; Feature update, the formula used is as follows: ； In the formula, represents the feature of node i at layer l+1, represents the neighbor set of node i, represents the feature update learnable weight matrix, Represents the characteristics of node i in layer l; The formula used for dual-branch feature aggregation is as follows: ； In the formula, represents the output features of the dual-branch graph neural network submodule, Represents the output characteristics of the macro branch; The risk rule generation submodule is constructed to automatically generate interpretable financial risk rules, specifically using a neural rule engine to generate interpretable rules, including: Rule condition learning, specifically using a thresholded fully connected layer to generate rule conditions, the formula used is as follows: ； In the formula, Indicates the numth rule, represents the weight of the num-th rule on the feature, Indicates the threshold of the numth rule; To calculate the rule condition score, the formula used is as follows: ； In the formula, Indicates the rule condition score of the numth rule, Represents the sigmoid function; The rule importance score is calculated using the following formula: ； In the formula, represents the rule importance score of the numth rule, Represents the learnable weight vector of the numth rule; To obtain the risk prediction value based on the rule, the formula used is as follows: ； In the formula, represents the risk prediction value based on the rule, and Num represents the total number of rules; Generate rule explanation, the formula used is as follows: ； In the formula, represents the feature importance score of the qth feature, represents the p-th row and q-th column element of the causal weight matrix, Ex represents the generated explanation, represents the rule interpretation generating function, represents the feature explanation generating function, It means to sort in descending order and take the first K values. represents the rule importance score of the first rule, represents the rule importance score of the second rule, represents the rule importance score of the Numth rule, represents the feature importance score of the first feature, represents the feature importance score of the second feature, represents the feature importance score of the Qth feature, where Q represents the total number of features; The designed loss function specifically combines the cross entropy loss and the rule consistency loss to design the model loss function, and the formula used is as follows: ； In the formula, represents the predicted value of the model risk level, represents the model output weight, represents the model output bias term, represents the cross entropy loss value, Sa represents the total number of samples, represents the true value of the sa-th sample, represents the model risk level prediction value of the sa-th sample, represents the rule consistency loss value, Indicates the calculation of L2 norm, Loss indicates the model loss value; The constructing and training model specifically involves constructing a composite graph neural network model through the causal feature selection submodule, the time convolution submodule, the dual-branch graph neural network submodule, the risk rule generation submodule and the design loss function, and training the model based on the prediction training set, verifying the model performance based on the prediction test set, and obtaining the composite graph neural network model as a risk perception model.

3. The financial risk identification system based on artificial intelligence according to claim 1 is characterized in that: In the hyperparameter optimization module, it is used to optimize the model hyperparameters, specifically, to optimize the model hyperparameters of the risk perception model using an optimization algorithm to obtain an optimized risk perception model, and the specific contents include algorithm initialization, global search, local search, determination of the global optimal solution and optimization of the model hyperparameters; The algorithm initialization specifically initializes the search space and constructs an individual unit set, wherein the individual unit is used to represent a hyperparameter combination of the risk perception model, and the fitness function of the individual unit is a loss function of the risk perception model; The global search uses the following formula: ； In the formula, represents the velocity of the ath individual unit at the dt+1th iteration, represents the inertia weight, represents the velocity of the ath individual unit at the dtth iteration, represents the weight used to control the learning of its own optimal solution, represents the optimal solution of the ath individual unit, represents the position of the ath individual unit at the dtth iteration, represents the weight used to control the learning of the global optimal solution, represents the global optimal solution at the dtth iteration, represents the position of the ath individual unit at the dt+1th iteration; The local search includes: Calculate the probability of neighbor solution selection using the following formula: ； In the formula, represents the probability that the ath individual unit selects the bth neighbor individual unit, represents the fitness calculation function, N represents the number of neighboring individual units, represents the position of the bth neighbor individual unit at the dt+1th iteration; Design a neighbor solution learning strategy, specifically, individual units select neighbor individual units based on the neighbor solution selection probability, and optimize their own positions based on the positions of the selected neighbor individual units. The formula used is as follows: ； In the formula, represents the position of the ath individual unit at the dt+1th iteration in the local search, de represents the decay factor whose size decreases linearly from 2 to 1, and rand represents a random number in the range [0,1]; Design a neighbor solution comparison strategy. Specifically, after the individual unit optimizes its own position based on the position of the neighbor individual unit, it selects the neighbor individual unit again based on the neighbor solution selection probability, and compares the fitness with the selected neighbor individual unit to determine whether to update the position. The formula used is as follows: ； In the formula, represents the fitness comparison function, represents the final position of the ath individual unit at the dt+1th iteration, represents the position of the cth neighbor individual unit at the dt+1th iteration; The determination of the global optimal solution is specifically to continuously iterate and update until the optimization algorithm iteration termination condition is reached, and the position of the individual unit with the lowest fitness is taken as the global optimal solution. The global optimal solution is specifically the optimal risk perception model hyperparameter combination. The optimization algorithm iteration termination condition specifically includes that the individual unit fitness is less than a set threshold and reaches the maximum number of iterations; The optimization model hyperparameters are specifically performed by optimizing the hyperparameters of the risk perception model through the algorithm initialization, the global search, the local search and the determination of the global optimal solution to obtain an optimized risk perception model.

4. The financial risk identification system based on artificial intelligence according to claim 1 is characterized in that: In the data acquisition module, the financial risk original data set specifically includes a historical financial original data set and a current financial original data set. The historical financial original data set and the current financial original data set specifically include financial indicator data, enterprise relationship data and macro-industry data. The historical financial original data set also includes financial risk label data. The financial indicator data specifically includes asset and liability data, profit data and financial flow data. The enterprise relationship data specifically includes supply chain relationship data, equity relationship data, peer competition data, cooperation years data and cooperation share data. The macro-industry data specifically includes industry index data, industry growth rate data and industry interest rate change data. Financial risk labels specifically include low risk and high risk.

5. The financial risk identification system based on artificial intelligence according to claim 1 is characterized in that: In the data preprocessing module, the data cleaning is specifically to remove missing values ​​and duplicate values, the data encoding is specifically to encode the original data using the one-hot encoding method, the data normalization is specifically to normalize the original data using the minimum-maximum normalization method, and the data set segmentation is specifically to segment the historical financial original data set after the data cleaning, data encoding and data normalization into a prediction training set and a prediction test set. Through data preprocessing, the prediction training set, the prediction test set and the current financial data set are obtained.

6. The artificial intelligence-based financial risk identification system according to claim 1, characterized in that: In the financial risk identification module, specifically, the current financial data set is used as the input of the optimized risk perception model to perform financial risk identification, obtain risk perception reference data, and based on the risk perception reference data, comprehensively judge the occurrence of financial risks.

Images