Financing guarantee anti-fraud method and system based on artificial intelligence

Abstract

The invention provides a financing guarantee anti-fraud method and system based on artificial intelligence, and relates to the technical field of artificial intelligence, and the method comprises the steps: building a feature vector through obtaining enterprise multi-source heterogeneous data, building a bidirectional causal graph structure to represent a risk propagation relationship between enterprises, obtaining a risk propagation adjacency matrix through employing a NOTEARS algorithm, and obtaining a risk propagation adjacency matrix; and constructing a risk distribution measurement space by adopting a Wasserstein distance to carry out distribution robustness optimization, training to obtain an enterprise risk representation vector, and evaluating a fraud risk in combination with an integrated learning framework. According to the invention, fraudulent behaviors in a complex network can be effectively identified, and the accuracy and reliability of financing guarantee risk assessment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to artificial intelligence technology, and in particular to a financing guarantee anti-fraud method and system based on artificial intelligence. BACKGROUND

[0002] With the development of the financial market and the expansion of financing guarantee business, fraud is becoming increasingly complex and hidden, bringing huge risks and losses to financial institutions. The existing financing guarantee anti-fraud method mainly relies on manual review and simple rule judgment, which is difficult to cope with the current complex and changing fraud patterns. The existing financing guarantee anti-fraud technology still has the following defects and deficiencies:

[0003] The existing technology cannot effectively capture the risk transmission relationship between enterprises, especially cannot accurately model the two-way risk propagation mechanism between enterprises, resulting in the risk transmission effect between associated enterprises being ignored, affecting the accuracy of fraud risk assessment.

[0004] The existing risk assessment model is sensitive to changes in data distribution. When there is a deviation in the distribution of test data and training data, the generalization ability of the model decreases significantly, and it is difficult to cope with the situation of changing data distribution in actual business, lacking distribution robustness.

[0005] The existing technology usually regards enterprises as independent individuals for risk assessment, and fails to effectively integrate enterprise correlation network structure information and individual characteristics, lacking a unified framework to integrate enterprise characteristics and network propagation effect, resulting in the final fraud risk score lacking global consideration. SUMMARY

[0006] The embodiment of the present application provides a financing guarantee anti-fraud method and system based on artificial intelligence, which can solve the problems in the prior art.

[0007] The first aspect of the embodiment of the present application provides a financing guarantee anti-fraud method based on artificial intelligence, comprising:

[0008] Obtaining enterprise multi-source heterogeneous data to construct an enterprise feature vector; constructing a bidirectional causal graph structure based on the enterprise feature vector, the bidirectional causal graph structure using a directed acyclic graph to represent the two-way risk propagation relationship between enterprises; for each enterprise node, calculating the risk state of the node according to the risk state of its directly associated enterprise; using the NOTEARS optimization algorithm, calculating the risk propagation adjacency matrix between enterprises based on the risk state of the node;

[0009] Based on the enterprise feature vector and the risk propagation adjacency matrix, a metric space of enterprise risk distribution is constructed using the Wasserstein distance, distribution robustness optimization is performed in the metric space, and a triple loss function is used to train an enterprise risk representation vector; the enterprise risk representation vector is input into an integrated learning framework to obtain an initial score of enterprise fraud risk;

[0010] According to the risk propagation adjacency matrix and the initial score, a final fraud risk score considering the risk propagation effect is calculated, and a financing guarantee anti-fraud decision result is generated.

[0011] Based on the enterprise feature vector, a bidirectional causal graph structure is constructed, and the bidirectional causal graph structure uses a directed acyclic graph to represent the bidirectional risk propagation relationship between enterprises; for each enterprise node, the risk state of the node is calculated according to the risk state of the directly associated enterprise; using the NOTEARS optimization algorithm, the risk propagation adjacency matrix between enterprises is calculated based on the risk state of the node, which comprises the following steps:

[0012] Based on the enterprise feature vector, the risk correlation degree between enterprises is calculated, and an enterprise risk correlation degree matrix is constructed; a directed acyclic graph structure is constructed according to the enterprise risk correlation degree matrix, and the directed acyclic graph structure represents the bidirectional risk propagation relationship between enterprises;

[0013] For each enterprise node in the directed acyclic graph structure, the risk state of the enterprise node is calculated, and the risk state of the enterprise node is obtained by weighted combination of the feature risk evaluation value and the associated enterprise influence value of the enterprise node, and the associated enterprise influence value is the weighted sum of the risk state of the directly associated enterprise of the enterprise node and the corresponding risk propagation weight;

[0014] The risk state of the enterprise node is composed of an enterprise risk state matrix, and an optimization objective function is constructed based on the enterprise risk state matrix, the optimization objective function includes a risk state fitting term, a risk propagation time sequence dynamic term and a acyclicity constraint term; the optimization objective function is solved using the NOTEARS optimization algorithm, and the risk propagation adjacency matrix between enterprises satisfying the directed acyclic constraint is obtained, and the risk propagation adjacency matrix represents the topological structure of the risk propagation between enterprises.

[0015] The steps of constructing the optimization objective function and solving the optimization objective function comprise:

[0016] The risk state fitting term is calculated by combining the risk state difference between enterprises and the importance weight, the risk propagation time sequence dynamic term is calculated by combining the time window decay weight and the historical risk propagation intensity, and the acyclicity constraint term is calculated by the loop connection of the risk propagation relationship between enterprises;

[0017] constructing an augmented Lagrangian function comprising the optimization objective function, the augmented Lagrangian function introducing a Lagrange multiplier term and a penalty term;

[0018] optimizing and solving the augmented Lagrangian function by using an L-BFGS algorithm introducing a second-order Hessian correction term, the second-order Hessian correction term constructing a diagonal block matrix based on the risk propagation timing characteristic, and adjusting a gradient update direction through the diagonal block matrix;

[0019] when the change value of the optimization objective is less than a set threshold value and a loop constraint is satisfied, obtaining an optimized timing risk propagation adjacency matrix; normalizing the propagation intensity in the optimized timing risk propagation adjacency matrix, and determining the risk propagation relationship between enterprises through a timing confidence score.

[0020] based on the enterprise feature vector and the risk propagation adjacency matrix, constructing a metric space of enterprise risk distribution using a Wasserstein distance, and performing distribution robustness optimization in the metric space, the steps of training an enterprise risk representation vector using a triplet loss function including:

[0021] weighting the enterprise feature vector through the risk propagation adjacency matrix, and mapping the weighted enterprise feature vector to a probability density space using a Gaussian kernel function to obtain an enterprise risk probability distribution;

[0022] constructing a Wasserstein distance metric space based on the enterprise risk probability distribution, wherein the Wasserstein distance is obtained by solving the p-th norm of the optimal transport plan between enterprise risk probability distributions;

[0023] constructing a triplet loss function based on the Wasserstein distance, dividing the classes of enterprise risk distribution pairs according to a preset Wasserstein distance threshold, and constructing a training sample set; by optimizing the triplet loss function, learning to obtain a representation vector reflecting the characteristics of enterprise risk distribution;

[0024] introducing an adversarial perturbation to the representation vector, setting a double-layer optimization objective, maximizing the adversarial perturbation in the inner layer, and minimizing the triplet loss in the outer layer, enhancing the robustness of the representation vector, and obtaining a final enterprise risk representation vector; based on the final enterprise risk representation vector, calculating the distance between enterprises to obtain consistency and discriminability indexes, and evaluating the enterprise risk representation result.

[0025] introducing an adversarial perturbation to the representation vector, setting a double-layer optimization objective, maximizing the adversarial perturbation in the inner layer, and minimizing the triplet loss in the outer layer, enhancing the robustness of the representation vector, and obtaining a final enterprise risk representation vector, the steps including:

[0026] calculating a sensitive direction in the feature vector space based on gradient information of the feature vector; generating an initial adversarial perturbation in the sensitive direction, an amplitude of the initial adversarial perturbation being determined based on a historical risk fluctuation range of the enterprise;

[0027] designing a dynamic weight triplet loss function, the dynamic weight being determined based on an Euclidean distance between pairs of enterprise risk feature vectors; adjusting a direction and an amplitude of the initial adversarial perturbation according to the Euclidean distance to generate a final adversarial perturbation;

[0028] constructing a double-layer optimization framework for adversarial training, an inner-layer optimization finding an optimal perturbation direction by maximizing a triplet loss function under adversarial perturbation, and an outer-layer optimization improving robustness of the feature vector by minimizing a weighted combination of an original triplet loss function and an adversarial triplet loss function, a weight coefficient of the weighted combination being dynamically adjusted according to a perturbation influence degree in a training process;

[0029] evaluating robustness of the feature vector based on a validation sample set, and calculating stability of a feature result under different perturbation amplitudes, when the stability meets a preset threshold, obtaining an enterprise risk feature vector with adversarial robustness.

[0030] inputting the enterprise risk feature vector into an ensemble learning framework to obtain an initial score of enterprise fraud risk, the steps comprising:

[0031] constructing a feature importance matrix based on enterprise risk feature vectors of historical fraud samples, and performing dimension reduction on the feature importance matrix by using a principal component analysis method to obtain a feature combination mode having identification effect on fraud;

[0032] projecting the enterprise risk feature vector to a subspace corresponding to the feature combination mode, constructing a decision tree ensemble structure based on gradient boosting, each decision tree corresponding to a feature combination mode; and performing weighted fusion on prediction results of each decision tree to obtain an initial score of enterprise fraud risk.

[0033] calculating a final fraud risk score considering a risk propagation effect based on the risk propagation adjacency matrix and the initial score, and generating a financing guarantee anti-fraud decision result, the steps comprising:

[0034] constructing an enterprise risk propagation network, determining a risk propagation path between enterprises based on a risk propagation adjacency matrix, and calculating a risk propagation time window according to a business association degree and historical transaction data between enterprises to obtain risk propagation intensities under different time windows;

[0035] For each enterprise node, a first-order risk propagation influence value is calculated based on the initial score of its adjacent enterprise nodes and the corresponding risk propagation intensity; a distance attenuation coefficient is introduced, which decreases with the increase of the propagation path, and a second-order risk propagation influence value is calculated; the first-order risk propagation influence value and the second-order risk propagation influence value are weighted and combined to obtain the cumulative risk influence coefficient of the enterprise;

[0036] The initial score is adjusted according to the cumulative risk influence coefficient to obtain a final fraud risk score considering the risk propagation effect; based on the final fraud risk score, the decision result of the financing guarantee anti-fraud is generated in combination with the guarantee limit of the enterprise.

[0037] The second aspect of the embodiment of the application provides a financing guarantee anti-fraud system based on artificial intelligence, which comprises:

[0038] The first unit is configured to acquire enterprise multi-source heterogeneous data to construct an enterprise feature vector; a bidirectional causal graph structure is constructed based on the enterprise feature vector, and the bidirectional causal graph structure uses a directed acyclic graph to represent the bidirectional risk propagation relationship between enterprises; for each enterprise node, the risk state of the node is calculated according to the risk state of its directly associated enterprise; a NOTEARS optimization algorithm is used to calculate the risk propagation adjacency matrix between enterprises based on the risk state of the node;

[0039] The second unit is configured to use the Wasserstein distance to construct a measurement space of enterprise risk distribution based on the enterprise feature vector and the risk propagation adjacency matrix, perform distribution robustness optimization in the measurement space, and train to obtain an enterprise risk representation vector by using a triplet loss function; the enterprise risk representation vector is input into an ensemble learning framework to obtain an initial score of enterprise fraud risk;

[0040] The third unit is configured to calculate a final fraud risk score considering the risk propagation effect according to the risk propagation adjacency matrix and the initial score, and generate a financing guarantee anti-fraud decision result.

[0041] The third aspect of the embodiment of the application,

[0042] An electronic device is provided, comprising:

[0043] A processor;

[0044] A memory for storing processor-executable instructions;

[0045] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0046] The fourth aspect of the embodiment of the application,

[0047] A computer readable storage medium is provided, and computer program instructions are stored on the computer readable storage medium, and the computer program instructions are executed by a processor to implement the method.

[0048] The beneficial effects of the present application are as follows:

[0049] The present application accurately captures the complex risk propagation relationship between enterprises by constructing the enterprise bidirectional causal graph structure and the risk propagation adjacency matrix, effectively identifies associated fraudulent behaviors, and improves the accuracy of financing guarantee anti-fraud.

[0050] The present application introduces the Wasserstein distance to construct the enterprise risk distribution measurement space and performs distribution robustness optimization, so that the model can adapt to the changes of data distribution and the evolution of enterprise behavior, and enhances the robustness and adaptability of the anti-fraud system.

[0051] The present application adopts the ensemble learning framework combined with risk propagation effect analysis, comprehensively considers the enterprise's own characteristics and associated risk factors, realizes the accurate evaluation and early warning of financing guarantee fraud behavior, and reduces the credit risk and loss rate of financial institutions. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The flowchart of the financing guarantee anti-fraud method based on artificial intelligence of the embodiment of the present application is shown in

[0053] Figure 2 The flowchart of the enterprise risk representation vector construction based on the Wasserstein distance and the adversarial robustness optimization is shown in DETAILED DESCRIPTION

[0054] In order to make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be described clearly and completely in combination with the drawings in the embodiment of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0055] The technical scheme of the present application will be described in detail in specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0056] Figure 1 The flowchart of the financing guarantee anti-fraud method based on artificial intelligence of the embodiment of the present application is shown in Figure 1 As shown, the method comprises:

[0057] An enterprise multi-source heterogeneous data is acquired to construct an enterprise feature vector; a bidirectional causal graph structure is constructed based on the enterprise feature vector, and the bidirectional causal graph structure uses a directed acyclic graph to represent a bidirectional risk propagation relationship between enterprises; for each enterprise node, a risk state of the node is calculated according to a risk state of a directly associated enterprise of the node; a NOTEARS optimization algorithm is used to calculate a risk propagation adjacency matrix between the enterprises based on the risk state of the node.

[0058] Based on the enterprise feature vector and the risk propagation adjacency matrix, a measurement space of enterprise risk distribution is constructed using a Wasserstein distance, distribution robustness optimization is performed in the measurement space, and a triple loss function is used to train an enterprise risk representation vector; the enterprise risk representation vector is input into an ensemble learning framework to obtain an initial score of enterprise fraud risk.

[0059] According to the risk propagation adjacency matrix and the initial score, a final fraud risk score considering the risk propagation effect is calculated, and a financing guarantee anti-fraud decision result is generated.

[0060] In an optional implementation, based on the enterprise feature vector, a bidirectional causal graph structure is constructed, and the bidirectional causal graph structure uses a directed acyclic graph to represent a bidirectional risk propagation relationship between enterprises; for each enterprise node, a risk state of the node is calculated according to a risk state of a directly associated enterprise of the node; and a NOTEARS optimization algorithm is used to calculate a risk propagation adjacency matrix between the enterprises based on the risk state of the node.

[0061] Based on the enterprise feature vector, a risk correlation degree between the enterprises is calculated to construct an enterprise risk correlation degree matrix; and a directed acyclic graph structure is constructed according to the enterprise risk correlation degree matrix, and the directed acyclic graph structure represents a bidirectional risk propagation relationship between the enterprises.

[0062] For each enterprise node in the directed acyclic graph structure, a risk state of the enterprise node is calculated, and the risk state of the enterprise node is obtained by weighted combination of a feature risk evaluation value of the enterprise node and an associated enterprise influence value, and the associated enterprise influence value is a weighted sum of a risk state of a directly associated enterprise of the enterprise node and a corresponding risk propagation weight.

[0063] The risk states of the enterprise nodes are combined to form an enterprise risk state matrix, an optimization objective function is constructed based on the enterprise risk state matrix, the optimization objective function includes a risk state fitting term, a risk propagation time sequence dynamic term and an acyclic constraint term; a NOTEARS optimization algorithm is used to solve the optimization objective function to obtain a risk propagation adjacency matrix between the enterprises that satisfies a directed acyclic constraint, and the risk propagation adjacency matrix represents a topological structure of risk propagation between the enterprises.

[0064] Exemplarily, an enterprise feature vector is obtained, which includes multi-dimensional feature information of the enterprise, such as financial indicators, business conditions, industry classification, and the like. These features can include asset-liability ratio, liquidity ratio, net profit ratio, revenue growth rate, and the like financial indicators, as well as enterprise size, establishment time, number of employees, and the like basic information. For example, for a manufacturing enterprise A, its feature vector contains multi-dimensional features such as asset-liability ratio 0.45, liquidity ratio 1.2, net profit ratio 0.08, revenue growth rate 0.12, enterprise size (large), establishment time 15 years, number of employees 1000, and the like.

[0065] The risk correlation degree can be determined by measuring the similarity between enterprise feature vectors, such as cosine similarity or Euclidean distance. For example, for two enterprises A and B, if they have 5-dimensional feature vectors [0.45, 1.2, 0.08, 0.12, 5] and [0.48, 1.1, 0.07, 0.15, 4] respectively, the risk correlation degree between them is calculated as 0.92. The risk correlation degree calculation results between all enterprises are combined to form an enterprise risk correlation degree matrix, which is an nxn square matrix, where n is the total number of enterprises.

[0066] A directed acyclic graph structure is constructed, and a risk correlation degree threshold is set, for example 0.8. When the risk correlation degree between two enterprises is greater than this threshold, a connection is established in the graph. For pairs of enterprises with a correlation degree higher than the threshold, the direction of risk propagation is determined according to factors such as enterprise size, industry influence, and the like. For example, if enterprise A is larger than enterprise B, the risk propagation direction from A to B is set. In this way, a directed graph is constructed, and topological sorting is used to ensure that the graph structure is acyclic, forming a directed acyclic graph.

[0067] For each enterprise node in the constructed directed acyclic graph, the risk state thereof is calculated. The risk state of the enterprise node is composed of two parts: a self-characteristic risk assessment value and a related enterprise influence value. The self-characteristic risk assessment value can be calculated by assigning weights to the key indicators of the enterprise. For example, the asset-liability ratio, the current ratio, and the net profit ratio are respectively assigned weights of 0.4, 0.3, and 0.3, and the weighted average is obtained to obtain the risk assessment value. If the asset-liability ratio risk score of the enterprise is 0.5, the current ratio risk score is 0.3, and the net profit ratio risk score is 0.2, then the self-risk assessment value is 0.5x0.4+0.3x0.3+0.2x0.3=0.35. The calculation method of the related enterprise influence value is to multiply the risk state of the directly related enterprise by the corresponding risk propagation weight and then sum them. For example, if enterprise A is directly related to enterprises B and C, and the risk states of enterprises B and C are 0.4 and 0.5 respectively, and the propagation weights are 0.3 and 0.2 respectively, then the related enterprise influence value is 0.4x0.3+0.5x0.2=0.22. The risk state of enterprise A is the weighted combination of the self-risk assessment value and the related enterprise influence value, such as 0.35x0.7+0.22x0.3=0.311.

[0068] The risk states of all enterprise nodes form an enterprise risk state matrix, which is an n-dimensional vector, where n is the total number of enterprises. When constructing the optimization objective function based on the enterprise risk state matrix, the risk state matrix is directly used as the data input of the optimization objective. The risk state fitting term calculates the error by comparing the predicted risk state of the model with the actual value in the risk state matrix; the risk propagation timing dynamic term uses the time sequence change data of the risk state matrix; and the acyclic constraint term imposes a constraint on the adjacency matrix generated by the model to ensure that the graph structure represented thereby does not contain a loop. The three terms together constitute the optimization objective function, guiding the algorithm to find the optimal risk propagation structure.

[0069] The NOTEARS optimization algorithm is used to solve the above optimization objective function. The core innovation of the NOTEARS algorithm is to convert the discrete constraints of the directed acyclic graph into a continuously differentiable function form, which enables the application of standard continuous optimization techniques. The algorithm first defines a special continuous function to represent whether the graph structure contains a loop. When the function value is zero, it indicates that the graph structure does not contain any loop. Then, the algorithm converts the original problem into a constrained optimization problem, and further converts it into an unconstrained optimization problem by introducing Lagrange multipliers. During the solving process, the algorithm uses the gradient descent method to iteratively update the element values of the risk propagation adjacency matrix. After each update, the loop constraint condition is checked. When the change of the objective function value is less than the preset threshold and the loop constraint is satisfied, the algorithm stops iteration. In practical applications, a suitable learning rate can be set to control the step size of each update, the maximum number of iterations can be set to avoid excessive calculation time, and the convergence threshold can be used to determine whether the algorithm has reached convergence. This optimization method can efficiently solve the inter-enterprise risk propagation adjacency matrix that satisfies the directed acyclic constraint. For example, the learning rate can be set to 0.01, the maximum number of iterations to 1000, and the convergence threshold to 1e-6.

[0070] After optimization, the inter-enterprise risk propagation adjacency matrix is obtained, and each element in the matrix represents the risk propagation intensity from enterprise i to enterprise j. For example, in a case involving 5 enterprises, the final risk propagation adjacency matrix is: [[0, 0.3, 0.5, 0, 0], [0, 0, 0.2, 0.4, 0], [0, 0, 0, 0.1, 0.3], [0, 0, 0, 0, 0.2], [0, 0, 0, 0, 0]]. This matrix indicates that enterprise 1 propagates risk to enterprises 2 and 3 with intensities of 0.3 and 0.5 respectively, enterprise 2 propagates risk to enterprises 3 and 4 with intensities of 0.2 and 0.4 respectively, and so on. By analyzing this matrix, the key path and core enterprise of risk propagation can be identified, providing decision support for risk prevention and control.

[0071] This method accurately depicts the inter-enterprise risk propagation relationship by constructing a bidirectional causal graph structure and a risk correlation matrix, and solves the directed acyclic constraint problem using the NOTEARS optimization algorithm. This method breaks through the limitations of traditional models in expressing complex inter-enterprise correlations, and can more accurately identify risk propagation paths and key node enterprises, providing a comprehensive risk network view for financial institutions. By calculating the risk state of enterprise nodes and constructing the risk propagation adjacency matrix, hidden risk transmission patterns can be captured, improving the accuracy of fraud risk identification and reducing false negatives and false positives.

[0072] In an alternative embodiment, the step of constructing an optimization objective function and solving includes:

[0073] The risk state fitting term is calculated by the combination of the risk state difference between enterprises and its importance weight, the risk propagation time sequence dynamic term is calculated by the combination of the time window decay weight and the historical risk propagation intensity, and the acyclic constraint term is calculated by the loop connection of the risk propagation relationship between enterprises;

[0074] An augmented Lagrange function including the optimization objective function is constructed, and the augmented Lagrange function introduces a Lagrange multiplier term and a penalty term;

[0075] An L-BFGS algorithm introducing a second-order Hessian correction term is used to optimize and solve the augmented Lagrange function, and the second-order Hessian correction term is based on the risk propagation time sequence characteristics to construct a diagonal block matrix, and the diagonal block matrix is used to adjust the gradient update direction;

[0076] When the change value of the optimization objective is less than a set threshold and the acyclic constraint is met, an optimized time sequence risk propagation adjacency matrix is obtained; the propagation intensity in the optimized time sequence risk propagation adjacency matrix is normalized, and the risk propagation relationship between enterprises is determined through time sequence confidence scoring.

[0077] For example, risk state data of a plurality of enterprises in different time windows is obtained, and a time sequence risk state matrix is constructed. Each row in the matrix represents an enterprise, each column represents a time point, and the matrix element represents the risk state value of the corresponding enterprise at the corresponding time point. For example, monthly risk state data of an enterprise in the past 12 months can be collected to form a time sequence risk state matrix including all target enterprises.

[0078] An optimization objective function is constructed based on the time sequence risk state matrix, including three key components: a risk state fitting term, a risk propagation time sequence dynamic term, and an acyclic constraint term.

[0079] The calculation method of the risk state fitting term is as follows: for any two enterprises i and j, the Euclidean distance between the corresponding row vectors in the time sequence risk state matrix is calculated as the risk state difference. Then, a weight is assigned to each pair of enterprise relationships based on the importance of the enterprise, and the importance weight is determined according to the enterprise asset size, industry influence, etc. For example, the enterprises are divided into three categories: large, medium and small, and weights of 0.8, 0.5 and 0.2 are assigned respectively. The importance weight between two enterprises can be the average of the weights of the two enterprises. The risk state fitting term is obtained by summing the product of the risk state difference of all enterprise pairs and the corresponding importance weight.

[0080] The calculation of the risk propagation time sequence dynamic term involves time window decay weight and historical risk propagation intensity. The time window decay weight adopts an exponential decay function, which is: weight = α (T-t)where a is the decay coefficient (e.g. 0.9), T is the total number of time points, and t is the index of a specific time point. For example, if there are 12 time points, the weight of the 1st time point is 0.9 11 , and the weight of the 12th time point is 0.9 0 = 1. The historical risk propagation intensity is determined by calculating the cross-correlation coefficient r(τ) of the risk indicators of the two enterprises, where τ represents the time delay. The maximum correlation coefficient r max and the corresponding delay τ max are selected. If r max > 0.6 and τ max > 0, it is considered that there is risk propagation from the leading enterprise to the lagging enterprise, and the initial propagation intensity is set as r max × (1 - e^(-τ max )). The timing dynamic term of risk propagation is obtained by calculating the weighted sum of the time window decay weight of each time point and the corresponding historical risk propagation intensity.

[0081] The purpose of the acyclic constraint term is to ensure that the final risk propagation network does not have circular dependencies. Circular connection refers to the formation of a closed loop in the risk propagation path of enterprises, such as enterprise A affecting enterprise B, enterprise B affecting enterprise C, and enterprise C affecting enterprise A. Detecting circular connections can be achieved by analyzing the properties of the risk propagation adjacency matrix. The specific method is to check the results of successive power operations of the adjacency matrix. When the diagonal elements of the matrix of a certain power appear non-zero values, it indicates the existence of a circular path. The acyclic constraint term is calculated by summing the penalty values of all cycles.

[0082] When constructing the augmented Lagrangian function, the original optimization objective function f(W) and the acyclic constraint h(W) = 0 are combined, which is in the form of L(W, λ, μ) = f(W) + λh(W) + μ / 2 × h(W) 2 , where W is the risk propagation adjacency matrix, λ is the Lagrange multiplier, and μ is the penalty factor. The initial value of the Lagrange multiplier λ is set to 0, and it is updated to λ + μh(W) after each iteration. The initial value of the penalty factor μ is set to 1.0, and it is multiplied by 1.5 after each iteration until it reaches an upper limit of 100. This method ensures that the algorithm gradually increases the punishment for the acyclic constraint, guiding the solution to converge to the feasible region that satisfies the constraint.

[0083] When solving the augmented Lagrangian function using the L-BFGS algorithm, a second-order Hessian correction term is introduced to improve the convergence speed and stability. The n enterprises can be divided into k groups, with each group having a size of approximately n / k. For example, 20 enterprises can be divided into 4 groups, with each group having 5 enterprises. The grouping basis can be industry category or historical transaction correlation between enterprises. For each group of enterprises, a Hessian submatrix is constructed, and the gradient difference (y k ) and parameter difference (s k ) of the last m = 10 iterations are stored.), the Hessian matrix of the new iteration is composed of three parts: the Hessian matrix of the last iteration, plus a positive correction term composed of the ratio of the outer product of the gradient difference vector and the inner product of the gradient difference vector and the parameter difference vector, minus a negative correction term composed of the ratio of the multiple matrix product of the Hessian matrix of the last iteration, the parameter difference vector and its transpose, and the inner product of the parameter difference vector and the product of the Hessian matrix and the parameter difference vector. This iterative calculation method avoids direct calculation of the second derivative, improving the efficiency of the algorithm. These block Hessian matrix components form a diagonal block matrix, which is used to adjust the gradient update direction and accelerate the optimization process.

[0084] The termination condition in the optimization process is set as follows: when the change in the optimization objective function value of two consecutive iterations is less than the preset threshold 0.0001 and the ring constraint value is less than 0.01, the algorithm is considered to have converged, and the optimized time sequence risk propagation adjacency matrix is obtained.

[0085] The propagation intensity in the optimized adjacency matrix is normalized by dividing the risk propagation intensity of each enterprise by its maximum value, so that all intensity values are mapped to the interval of zero to one, facilitating subsequent analysis and comparison. Finally, the risk propagation relationship between enterprises is determined through time sequence confidence scores, which consider both the stability and persistence of the propagation intensity.

[0086] Stability is evaluated by calculating the coefficient of variation (CV) of the propagation intensity within a historical time window. The coefficient of variation is defined as the standard deviation divided by the mean, and is used to measure the relative dispersion of data. For example, the average value of the propagation intensity of a certain enterprise within a 12-month window is 0.45, and the standard deviation is 0.09, so the coefficient of variation CV = 0.09 / 0.45 = 0.2, indicating that the propagation intensity is relatively stable. Persistence is evaluated by counting the number of time windows in which the propagation intensity exceeds a certain threshold. The threshold is set to 0.3, i.e. when the monthly propagation intensity is greater than 0.3, it is considered that there is significant risk propagation in that month. The persistence ratio is defined as the ratio of the number of months with propagation intensity exceeding the threshold to the total number of months. For example, if the propagation intensity exceeds 0.3 in 9 out of 12 months, the persistence ratio is 9 / 12 = 0.75. The time sequence confidence score considers both stability and persistence, and the calculation formula is: time sequence confidence score = 0.5 x (1-min(1, CV / 0.5)) + 0.5 x min(1, N s / N total ); where CV is the coefficient of variation, N s is the number of time windows in which the propagation intensity exceeds the threshold, and N total is the total number of time windows.

[0087] Taking the above example as an example, the stability score = 0.5 x (1-0.2 / 0.5) = 0.3, the persistence score = 0.5 x 0.75 = 0.375, and the total score = 0.675. When the confidence score exceeds 0.6, it is confirmed that there is a significant risk propagation relationship.

[0088] A complete practical case: suppose that 5 enterprises (A, B, C, D, E) are analyzed for 12 months of risk propagation relationship. In the initial adjacency matrix before optimization, the propagation strength from enterprise A to B is estimated to be 0.4, and after 35 iterations of the optimization algorithm, the propagation strength from enterprise A to B is finally adjusted to 0.65, with a coefficient of variation of 0.15, and the propagation strength exceeds 0.3 for 10 months in 12 months, the time series confidence score is 0.5 x (1-0.15 / 0.5) + 0.5 x (10 / 12) ≈ 0.35 + 0.42 = 0.77, which exceeds the threshold value 0.6, so it is confirmed that there is a significant risk propagation relationship from A to B.

[0089] The method innovatively constructs an optimization objective function containing a risk state fitting term, a risk propagation time series dynamic term and a acyclic constraint term, and solves it by augmented Lagrange function. The L-BFGS algorithm with second-order Hessian correction term significantly improves the calculation efficiency and optimization accuracy, and can effectively handle the complex calculation of large-scale enterprise risk network. Through the normalization processing and time series confidence score of the optimized adjacency matrix, the method can provide risk propagation strength change analysis in the time dimension, help financial institutions to predict risk transmission trends, and realize the forward-looking control and accurate decision-making of risk.

[0090] In an optional implementation, based on the enterprise feature vector and the risk propagation adjacency matrix, a metric space of enterprise risk distribution is constructed using the Wasserstein distance, and the step of training the enterprise risk representation vector using a triplet loss function by performing distribution robustness optimization in the metric space includes:

[0091] The enterprise feature vector is weighted by the risk propagation adjacency matrix, and the weighted enterprise feature vector is mapped to a probability density space using a Gaussian kernel function to obtain an enterprise risk probability distribution;

[0092] A Wasserstein distance metric space is constructed based on the enterprise risk probability distribution, wherein the Wasserstein distance is obtained by solving the p-order norm of the optimal transport plan between enterprise risk probability distributions;

[0093] A triple loss function is constructed based on the Wasserstein distance, and a training sample set is constructed according to a preset Wasserstein distance threshold value to divide the categories of enterprise risk distribution pairs; a representation vector reflecting the characteristics of enterprise risk distribution is learned by optimizing the triple loss function;

[0094] An adversarial perturbation is introduced to the representation vector, and a double-layer optimization objective is set, in which the adversarial perturbation is maximized in the inner layer and the triple loss is minimized in the outer layer, so as to enhance the robustness of the representation vector and obtain a final enterprise risk representation vector; a consistency index and a discriminability index are obtained by calculating the distance between enterprises based on the final enterprise risk representation vector, and the enterprise risk representation result is evaluated.

[0095] In combination with Figure 2 A flowchart of the enterprise risk representation vector construction based on the Wasserstein distance and adversarial robustness optimization is described. For example, the enterprise feature vector is weighted by the risk propagation adjacency matrix. The specific method is to perform dot product operation on the enterprise feature vector and the corresponding risk propagation adjacency matrix row vector to obtain a weighted feature vector considering the influence of risk propagation. The element value in the risk propagation adjacency matrix reflects the risk transmission strength between enterprises. The weighting process is essentially to integrate the risk characteristics of related enterprises into the feature representation of the target enterprise according to the risk propagation strength. For example, the original feature vector of a certain enterprise A contains 20 feature dimensions such as asset-liability ratio, current ratio, profitability, etc., which have been standardized to the 0-1 interval. Assuming that the weights of enterprises B, C and D related to enterprise A in the risk propagation adjacency matrix are 0.6, 0.3 and 0.1 respectively, the weighted feature vector of enterprise A is calculated as 0.6 times the original feature vector of enterprise A plus 0.3 times the original feature vector of enterprise B plus 0.1 times the original feature vector of enterprise C.

[0096] A Gaussian kernel function is applied to the weighted enterprise feature vector for nonlinear transformation to map the feature vector to a probability density space. The form of the Gaussian kernel function is K(x,y)=exp(-||x-y|| ² / 2σ ²), where x and y are feature vectors, ||x-y|| is the Euclidean distance, and σ is the bandwidth parameter. The bandwidth parameter is determined according to the distribution characteristics of the feature vectors, and can be calculated using the Silverman rule, i.e., σ = 0.9 x min (standard deviation, quartile range / 1.34) x n^(-1 / 5), where n is the sample size. For example, for a 20-dimensional weighted feature vector of an enterprise, each dimension feature can be regarded as a data point. If the standard deviation of these feature values is 0.15 and the sample size is 20, then the bandwidth parameter σ = 0.9 x 0.15 x 20^(-1 / 5) ≈ 0.102. After applying the Gaussian kernel function, the original feature vector is transformed into a probability density function, which can more comprehensively represent the distribution characteristics of enterprise risk.

[0097] Based on the enterprise risk probability distribution, a Wasserstein distance measurement space is constructed. The Wasserstein distance measures the difference between two probability distributions by calculating the optimal transport scheme between them. For the risk probability distributions P and Q of two enterprises, their support sets can be obtained by uniformly sampling in the feature space, usually sampling 50-100 points as discrete support points. The element c ij of the transport cost matrix C represents the cost of moving the support point i of distribution P to the support point j of distribution Q, which is calculated as the square of the Euclidean distance between the two points. The optimal transport plan can be solved using the Sinkhorn iterative algorithm, with a regularization parameter ε = 0.01 and an iteration number of 100 to obtain the transport matrix T. The Wasserstein distance is calculated as the sum of the element-wise product of the transport matrix T and the cost matrix C. For example, for the risk probability distributions of two enterprises, if their support sets each contain 50 points, the 50x50 transport cost matrix is constructed with element values ranging from 0 to 1. After solving using the Sinkhorn algorithm, the sum of the elements of the transport matrix is 1. After multiplying with the cost matrix, the value of the Wasserstein distance is 0.35, indicating that the similarity of the risk distributions of the two enterprises is relatively high.

[0098] The Wasserstein distance calculated is used to construct a triplet loss function. A triplet consists of an anchor enterprise a, a positive sample enterprise p, and a negative sample enterprise n. The positive sample enterprise is an enterprise similar to the anchor enterprise in terms of risk distribution, and the negative sample enterprise is an enterprise with a larger difference in risk distribution from the anchor enterprise. Specifically, set the Wasserstein distance threshold to 0.5. When the Wasserstein distance between two enterprises is less than 0.5, it is considered a similar enterprise pair, and when it is greater than 0.5, it is considered an dissimilar enterprise pair. Filter enterprises with known risk states from historical data as training samples to construct a large number of triplet samples. For example, filter 100 enterprises from financial historical data, randomly select an anchor enterprise, and then select enterprises with a Wasserstein distance less than 0.5 and greater than 0.5 as positive and negative samples, respectively, to construct 10,000 triplet training samples.

[0099] The design goal of the triplet loss function is to make the distance between similar enterprise pairs in the representation space less than that of dissimilar enterprise pairs, and the gap exceeds the specified margin value. The specific form is max(0, d(f a ,f p )-d(f a ,f n )+margin), where d represents the Euclidean distance between representation vectors, f a , f p , f n represent the representation vectors of the anchor, positive sample, and negative sample enterprises, respectively, and margin represents the expected margin value, set to 0.2. A loss value greater than 0 indicates that the current representation vector does not meet the expected distance relationship and needs to be adjusted. The loss function is optimized by gradient descent method, with an initial learning rate of 0.01, the learning rate is reduced to 0.8 times of the original every 5000 batches, a total of 30000 batches are trained, and the representation vector that can reflect the characteristics of enterprise risk distribution is obtained.

[0100] An adversarial perturbation is introduced to the learned representation vector to enhance robustness. The adversarial perturbation δ is generated by calculating the gradient ∇_fL of the representation vector f to the loss function L, and the direction is consistent with the gradient direction, and the amplitude is limited to 5% of the norm of the representation vector, that is, ||δ||≤0.05×||f||. The way to generate an adversarial sample is f'=f+δ. Set a double-layer optimization objective: the inner-layer optimization adjusts the perturbation δ by 5-step gradient ascent, and the step size is 0.01, and the goal is to maximize the loss function L(f+δ); the outer-layer optimization adjusts the model parameters θ by 20-step gradient descent, and the step size is 0.005, and the goal is to minimize the weighted sum of the original loss function L(f) and the adversarial loss function L(f+δ) 0.5×L(f)+0.5×L(f+δ). In practical applications, the representation vector of a certain enterprise is subjected to adversarial perturbation, and a perturbation with an amplitude of 0.03 is added to the original representation vector. The representation vector obtained after double-layer optimization has stronger robustness to data perturbation while maintaining the original semantic information.

[0101] When evaluating the results of enterprise risk representation, consistency indicators and discriminability indicators are calculated. The consistency indicator measures the degree of aggregation of enterprises with similar risk distributions in the representation space, and the calculation method is the average of the Euclidean distances of all enterprise pairs with a Wasserstein distance less than 0.5 in the representation space. The discriminability indicator measures the degree of separation of enterprises with different risk distributions in the representation space, and the calculation method is the average of the Euclidean distances of all enterprise pairs with a Wasserstein distance greater than 0.5 in the representation space. The Euclidean distance between representation vectors is calculated as the square root of the square sum of the difference between the two vectors. Ideally, the consistency indicator should be small, and the discriminability indicator should be large. For example, in the experimental evaluation, the representation results of 100 enterprises were analyzed, the average representation distance of similar enterprise pairs was 0.15, the average representation distance of dissimilar enterprise pairs was 0.85, and the distance ratio was 0.15 / 0.85=0.18, indicating that the representation vector had good discrimination ability.

[0102] For example, for risk characterization analysis of enterprises in a certain financial data set, a total of 200 enterprises' historical risk data are included. Each enterprise's original feature vector contains 30 financial indicators and associated transaction features. After weighting by the risk propagation adjacency matrix, the feature vector is mapped to the probability density space using a Gaussian kernel function with a bandwidth parameter of 0.1. 75 support points are sampled to represent each enterprise's risk probability distribution, the Wasserstein distance between enterprise pairs is calculated to form a 200x200 distance matrix, and the average distance is 0.62. A distance threshold of 0.4 is set to construct about 15,000 triplet training samples. By optimizing the triplet loss function, a 128-dimensional enterprise risk characterization vector is learned. By introducing an adversarial perturbation with a magnitude of 3% of the vector norm, the robustness of the characterization vector is significantly improved after double-layer optimization. The consistency index on the validation set is 0.18, the discriminability index is 0.79, and the distance ratio is 0.23, indicating that the model can effectively distinguish enterprises with different risk patterns.

[0103] The metric space constructed by the Wasserstein distance can accurately capture the subtle differences in enterprise risk distribution, the triplet loss function effectively learns the characterization vector of the risk distribution, and the adversarial training mechanism significantly enhances the robustness of the characterization. Compared with traditional methods, this method can more accurately identify the risk similarity between enterprises, has stronger adaptability to data noise and distribution changes, provides a more reliable risk characterization basis for financing guarantee anti-fraud decision-making, reduces the fraud risk misjudgment rate, and improves the accuracy and stability of financial risk management.

[0104] In an optional implementation, the step of introducing an adversarial perturbation to the characterization vector, setting a double-layer optimization objective, maximizing the adversarial perturbation in the inner layer, and minimizing the triplet loss in the outer layer to enhance the robustness of the characterization vector, and obtaining the final enterprise risk characterization vector includes:

[0105] Based on the gradient information of the characterization vector, a sensitive direction in the characterization vector space is calculated. An initial adversarial perturbation is generated in the sensitive direction, and the magnitude of the initial adversarial perturbation is determined based on the historical risk fluctuation range of the enterprise.

[0106] A dynamic weight triplet loss function is designed, and the dynamic weight is determined based on the Euclidean distance between enterprise risk characterization vectors. The direction and magnitude of the initial adversarial perturbation are adjusted based on the Euclidean distance to generate the final adversarial perturbation.

[0107] A double-layer optimization framework for adversarial training is constructed. The inner-layer optimization finds the optimal perturbation direction by maximizing the triplet loss function under adversarial perturbation. The outer-layer optimization improves the robustness of the representation vector by minimizing the weighted combination of the original triplet loss function and the adversarial triplet loss function. The weight coefficient of the weighted combination is dynamically adjusted according to the degree of influence of the perturbation during the training process.

[0108] The robustness of the representation vector is evaluated based on the validation sample set. The stability of the representation result under different perturbation amplitudes is calculated. When the stability meets the preset threshold, the enterprise risk representation vector with adversarial robustness is obtained.

[0109] For example, the gradient information of the representation vector is calculated by deriving the triplet loss function. The sensitive direction of the representation vector refers to the gradient direction of the loss function with respect to the representation vector. A small change in this direction will cause a significant change in the loss function value. During calculation, a batch of triplet samples is selected. The gradient of the loss function with respect to the representation vector is calculated for each sample. The average of the gradient vectors is taken as the sensitive direction. The quantitative representation of the sensitive direction is a vector with the same dimension as the representation vector. The absolute value of the vector element represents the sensitivity of the corresponding dimension. For example, for a 128-dimensional enterprise risk representation vector, the 57th and 93rd dimensions of the calculated sensitive direction vector are 0.23 and -0.31, respectively, indicating that these two dimensions have a greater impact on the change of the loss function.

[0110] The initial adversarial perturbation is generated in the sensitive direction, and the perturbation amplitude is determined based on the historical risk fluctuation range of the enterprise. The historical risk fluctuation range is obtained by calculating the standard deviation of the risk indicators of the enterprise in the past 12 months. The risk indicator refers to the risk state value of the enterprise node recorded in the enterprise risk state matrix, which is obtained by weighted combination of the self-characteristic risk evaluation value and the associated enterprise influence value, ranging from 0 to 1, with a higher value indicating higher risk. The specific calculation method is to extract the monthly risk state values of the past 12 months for each enterprise, calculate the standard deviation of these values, and obtain the historical risk fluctuation range of the enterprise. For example, the risk state values of an enterprise in the past 12 months are [0.35, 0.38, 0.42, 0.45, 0.47, 0.50, 0.48, 0.46, 0.43, 0.41, 0.39, 0.37], and the standard deviation is calculated to be 0.047, i.e. the historical risk fluctuation range of the enterprise. A personalized perturbation amplitude is set for each enterprise, with the perturbation amplitude set to between 30% and 50% of the historical risk standard deviation, ensuring that the perturbation is small enough not to change the semantic information of the representation, while effectively testing the robustness of the model. In specific operation, the direction of the perturbation vector is set to the sensitive direction, and the perturbation amplitude is set to 5% of the norm of the representation vector. For example, the historical risk standard deviation of an enterprise is 0.047, and 40% of it is 0.0188, which is about 5% of the norm of the representation vector of the enterprise, so the corresponding adversarial perturbation amplitude can be set to 0.05.

[0111] The design of the dynamic weight triple loss function is based on the Euclidean distance between the risk representation vectors of enterprises. The smaller the Euclidean distance between the anchor enterprise and the positive sample enterprise in the triple, the more similar their risk conditions are; the greater the Euclidean distance between the anchor enterprise and the negative sample enterprise, the more different their risk conditions are. The Euclidean distance is calculated as the square root of the sum of the squares of the difference between the corresponding elements of the two representation vectors. For example, the Euclidean distance between two 128-dimensional representation vectors is calculated as the square root of the sum of the squares of the difference between the corresponding elements in 128 dimensions, with a value range generally between 0 and 2 for normalized representation vectors. The dynamic weight is calculated according to the distance ratio: when the anchor and the positive sample distance is close to the anchor and the negative sample distance, a higher weight is given to the triple, because such samples are more critical to the improvement of the model's discrimination ability. The weight calculation formula is the inverse of the distance ratio, where the distance ratio is the anchor and the positive sample distance divided by the anchor and the negative sample distance. For example, if the anchor and the positive sample distance is 0.2, and the anchor and the negative sample distance is 0.8, then the weight of the triple is 1-(0.2 / 0.8)=0.75.

[0112] The initial adversarial perturbation is adjusted based on the Euclidean distance to generate a final adversarial perturbation. The adjustment methods include direction adjustment and amplitude adjustment. Direction adjustment updates the sensitive direction of the gradient of the triplet loss of the representation vector each time after iteration. The amplitude coefficient of the direction update is set to 0.8, indicating that the new direction is composed of 80% of the current direction and 20% of the newly calculated gradient direction, ensuring the smoothness of the direction adjustment. Amplitude adjustment is based on the difficulty of the triplet sample. The difficulty is measured by the ratio of the distance between the anchor and the positive sample to the distance between the anchor and the negative sample. Specifically, the difficulty coefficient is defined as r = p / n, where p is the distance between the anchor and the positive sample, and n is the distance between the anchor and the negative sample. According to the difficulty coefficient, the perturbation amplitude adjustment factor is set: when r > 0.8, it is determined as a difficult-to-distinguish triplet, the perturbation amplitude is increased, and the adjustment factor is 1 + (r-0.8) x 2.5; when r < 0.4, it is determined as an easy-to-distinguish triplet, the perturbation amplitude is reduced, and the adjustment factor is 0.7 + (r-0.2) x 0.75; when 0.4 ≤ r ≤ 0.8, it belongs to a moderately difficult triplet, and the original perturbation amplitude is maintained, and the adjustment factor is 1.0. For example, in a certain triplet, the distance between the anchor and the positive sample is 0.35, and the distance between the anchor and the negative sample is 0.40, the difficulty coefficient is 0.35 / 0.40 = 0.875 > 0.8, which belongs to a difficult-to-distinguish triplet, and the adjustment factor is 1 + (0.875-0.8) x 2.5 = 1.1875, so the initial perturbation amplitude is increased by 18.75%, from 0.05 to 0.059. Another example, in a certain triplet, the distance between the anchor and the positive sample is 0.15, and the distance between the anchor and the negative sample is 0.70, the difficulty coefficient is 0.15 / 0.70 = 0.214 < 0.4, which belongs to an easy-to-distinguish triplet, and the adjustment factor is 0.7 + (0.214-0.2) x 0.75 = 0.71, so the initial perturbation amplitude is reduced by 29%, from 0.05 to 0.036.

[0113] The two-level optimization framework of adversarial training includes inner optimization and outer optimization. The inner optimization maximizes the triplet loss under adversarial perturbation by gradient ascent method, and iteratively updates the direction and amplitude of adversarial perturbation. The triplet loss under adversarial perturbation refers to the triplet loss value calculated after adding adversarial perturbation to the original representation vector. The inner optimization is set to 5 iterations, and each iteration moves a step of 0.01 along the gradient direction, while ensuring that the perturbation amplitude does not exceed the set upper limit, which is 10% of the norm of the representation vector. The outer optimization minimizes the weighted combination of the original triplet loss and the adversarial triplet loss by gradient descent method, and updates the model parameters. The model parameters include the weight and bias parameters of the representation vector learning network, usually thousands to tens of thousands of parameters, depending on the complexity of the network structure. The outer optimization is set to 20 iterations, with a learning rate of 0.005, using the Adam optimizer, where the hyperparameters of the Adam optimizer are set as β1 = 0.9, β2 = 0.999, and ε = 10 -8The weight coefficients of the weighted combination are initially set to 0.5:0.5, indicating that the original loss and the adversarial loss each account for 50%, and are dynamically adjusted as the training progresses: if the adversarial perturbation has a greater impact on the model performance, such as a performance decrease of more than 8% on the validation set, the weight of the adversarial loss is increased to 60%-70%; if the impact is small, such as a performance decrease of less than 3% on the validation set, the weight of the adversarial loss is reduced to 30%-40%. For example, the weight of the adversarial loss and the original loss is 0.5 at the beginning of training, and after 10,000 iterations, if the adversarial perturbation causes a 10% decrease in performance on the validation set, the weights are adjusted to 0.3:0.7, i.e., the weight of the original loss is 0.3 and the weight of the adversarial loss is 0.7.

[0114] The robustness evaluation of the representation vector is based on the validation sample set. The validation sample set is composed of enterprise data that does not participate in training, usually 20% of the total data set. Random perturbations of different amplitudes are added to the enterprise representation vectors in the validation set, with the perturbation amplitude increasing from 1% to 10% of the norm of the representation vector, a total of 10 perturbation levels. Random perturbations are generated by adding random noise following a normal distribution N(0, σ 2 ) to each dimension of the representation vector, where σ is the perturbation amplitude. For each perturbation level, the cosine similarity of the representation vector before and after perturbation is calculated, and the cosine similarity is calculated as the inner product of the two vectors divided by the product of the respective norms, with a value range of -1 to 1, and a larger value indicating that the two vectors are more similar in direction. At the same time, the change rate of the triplet loss function value calculated based on the perturbed representation vector is calculated, and the change rate is calculated as the difference between the perturbed loss value and the original loss value divided by the original loss value. The stability of the representation result is defined as the weighted average of the cosine similarity and the loss change rate at different perturbation levels, with weights of 0.6 and 0.4 respectively. The stability score is calculated as 0.6 x average cosine similarity + 0.4 x (1 - average loss change rate), with a value range of 0 to 1, and a larger value indicating better stability. When the average cosine similarity is higher than 0.9 and the loss change rate is lower than 15%, i.e., the stability score is higher than 0.84, the representation vector is considered to have sufficient robustness. For example, for the representation vectors of 100 enterprises in the validation set, at a perturbation amplitude of 5%, the average cosine similarity is 0.94 and the loss change rate is 8%, and the stability score is 0.6 x 0.94 + 0.4 x (1 - 0.08) = 0.564 + 0.368 = 0.932, indicating that the model has good robustness.

[0115] In the application example, the risk representation vectors of 200 enterprises are subjected to adversarial robustness enhancement. The enterprises come from different industries, including manufacturing, service, finance, etc., each enterprise has 30 original features, and the 128-dimensional risk representation vector is obtained through the foregoing method. The average risk state value of each enterprise in the past 12 months is between 0.2 and 0.8, and the average historical risk standard deviation is 0.06. The initial representation vector is 128-dimensional, the sensitive direction is calculated by the foregoing method, and the initial adversarial perturbation is generated. The average Euclidean distance between the anchor enterprise and the positive sample enterprise in the triple sample is 0.25, and the average Euclidean distance between the anchor enterprise and the negative sample enterprise is 0.73, based on which a dynamic weight triple loss function is designed. In the double-layer optimization framework, the inner-layer optimization generates the optimal adversarial perturbation through 5 iterations, and the average perturbation amplitude is 6.5% of the norm of the representation vector. The outer-layer optimization updates the model parameters through 20 iterations, the model includes a three-layer fully connected neural network, the total number of parameters is about 50,000, and a total of 15,000 batches of training are performed, each batch contains 64 triple samples. In the training, the weight of the weighted combination is adjusted from the initial 0.5:0.5 to 0.4:0.6. After the training is completed, the robustness of the representation vector is evaluated on the validation set, the average cosine similarity under a perturbation amplitude of 5% is 0.93, the loss change rate is 10.5%, and the stability score is 0.894, which meets the preset threshold requirement, and the enterprise risk representation vector with adversarial robustness is obtained.

[0116] The present application effectively enhances the robustness of the enterprise risk representation vector by introducing an adversarial perturbation based on the historical risk fluctuation range of the enterprise, combining a triple loss function with dynamic weights and a double-layer optimization framework. Compared with traditional methods, this technology can better cope with data noise and distribution bias, making the risk representation stable in the face of small changes in enterprise financial data, improving the generalization ability of the model in different scenarios, providing a more reliable risk measurement basis for financing guarantee anti-fraud, reducing the misjudgment rate, and improving the accuracy and reliability of financial risk prevention and control.

[0117] In an optional implementation, the step of inputting the enterprise risk representation vector into an ensemble learning framework to obtain an initial score of the enterprise fraud risk includes:

[0118] A feature importance matrix is constructed based on the enterprise risk representation vectors of historical fraud samples, and a principal component analysis method is used to reduce the dimension of the feature importance matrix to obtain a feature combination mode that has an identifying effect on fraud;

[0119] The enterprise risk representation vector is projected into a subspace corresponding to the feature combination mode, a decision tree ensemble structure based on gradient boosting is constructed, and each decision tree corresponds to a feature combination mode; the prediction results of each decision tree are weighted and fused to obtain an initial score of the enterprise fraud risk.

[0120] Exemplarily, labeled fraudulent enterprise samples and normal enterprise samples are extracted from the historical database, which have completed the construction of risk characterization vectors. For each enterprise's risk characterization vector, it contains features in dimensions of financial indicators, business conditions, legal risks, etc., totaling up to 300 feature dimensions. For these features, the feature importance is evaluated by calculating the correlation between each feature and the fraud label. Specifically, for numerical features, the point two column correlation coefficient between them and the fraud label is calculated; for categorical features, the information gain value is calculated. In this way, a feature importance matrix of size m x n is generated, where m represents the number of samples, n represents the number of features, and the elements in the matrix represent the contribution of the feature to fraud identification.

[0121] Suppose there are 1000 enterprise samples with known fraud status, each with 200 valid features. The feature importance matrix calculated by the above method is a 1000 x 200 matrix. The values in the matrix are between 0 and 1, with larger values indicating higher importance of the feature in identifying fraud. For example, the importance of "financial anomaly indicators" is 0.85, while the importance of "enterprise age" is 0.32.

[0122] The feature importance matrix is processed for dimensionality reduction, and the principal component analysis method is applied. The feature importance matrix is standardized, the covariance matrix between features is calculated, and the eigenvalues and eigenvectors of the covariance matrix are calculated. Sort the eigenvalues from large to small, and select the first k principal components whose cumulative contribution rate reaches 95% as the new feature combination method. These principal components reflect the most effective directions in the original feature space for fraud identification.

[0123] Suppose that through principal component analysis, 15 main feature combinations (principal components) are obtained, which capture the most critical variation directions in the original 200-dimensional feature space for fraud identification. Each principal component is a linear combination of the original features, and the weights reflect the importance of each original feature in the principal component. For example, the first principal component is mainly composed of "cash flow anomalies", "asset-liability ratio anomalies", and "frequent related transactions", with weights of 0.6, 0.5, and 0.4, respectively.

[0124] Enter the decision tree ensemble construction phase, and project the risk characterization vector of the enterprise to be evaluated into the k principal component space obtained earlier. For each principal component direction, a decision tree based on the gradient boosting algorithm is constructed. Each decision tree focuses on fraud judgment from the features extracted in the corresponding principal component direction. In the training process, the fraud label of the historical sample is used as the target variable, and the structure and parameters of the decision tree are continuously optimized through gradient boosting. The generation process of the decision tree considers parameters such as feature split gain, tree depth control, and minimum sample size in the leaf node to prevent overfitting.

[0125] The configuration of each decision tree is as follows: the maximum depth is set to 6, the minimum leaf node sample number is 20, the learning rate is set to 0.1, and the split quality is evaluated using a logarithmic loss function. For each principal component direction decision tree, the optimal tree parameter configuration is determined through cross-validation.

[0126] Finally, the initial score is obtained by weighted fusion. Each decision tree corresponds to a principal component direction, and a fraud score (between 0 and 1) is given for the enterprise to be evaluated. According to the proportion of explained variance of each principal component (i.e., feature combination method), the weight is determined, and the principal component with higher explained variance corresponds to higher weight of the decision tree. The sum of the weighted scores of all decision trees is the initial score of the enterprise fraud risk.

[0127] Suppose the risk representation vector of the enterprise A to be evaluated is projected into the space of 15 principal components after principal component analysis. The fraud scores given by the decision trees corresponding to the 15 directions are [0.75, 0.82, 0.68, 0.71, 0.65, 0.60, 0.58, 0.55, 0.52, 0.48, 0.45, 0.42, 0.40, 0.38, 0.35], and the proportion of explained variance of the 15 principal components is [0.20, 0.15, 0.12, 0.10, 0.08, 0.07, 0.06, 0.05, 0.04, 0.03, 0.03, 0.02, 0.02, 0.02, 0.01]. Through weighted fusion calculation: 0.75x0.20+0.82x0.15+0.68x0.12+...+0.35x0.01, the initial score of the fraud risk of enterprise A is 0.65. This score indicates that enterprise A has a higher fraud risk and needs further review.

[0128] To improve accuracy, sample resampling techniques can also be used to handle data imbalance problems during implementation, and cross-validation can be used to optimize model parameters to ensure the stability and reliability of the scoring results.

[0129] This method uses principal component analysis to reduce the dimensionality of the feature importance matrix, extracts the most valuable feature combination for fraud identification, and effectively reduces feature redundancy and computational complexity. The gradient boosting-based decision tree ensemble structure can learn specific features for different feature combinations, capturing the multi-dimensional features and non-linear relationships of enterprise risk. By weighting the prediction results of each decision tree, this method fully utilizes the advantages of ensemble learning, improves the accuracy and generalization ability of fraud risk assessment, and provides more accurate initial scores for anti-fraud decisions, reducing the business losses caused by false positives.

[0130] In an alternative embodiment, according to the risk propagation adjacency matrix and the initial score, the step of calculating a final fraud risk score considering the risk propagation effect to generate a financing guarantee anti-fraud decision result comprises:

[0131] The enterprise risk propagation network is constructed, the risk propagation path between enterprises is determined based on the risk propagation adjacency matrix, and the risk propagation time window is calculated according to the business association degree and the historical transaction data between enterprises to obtain the risk propagation intensity under different time windows;

[0132] For each enterprise node, a first-order risk propagation influence value is calculated based on the initial score of the adjacent enterprise node and the corresponding risk propagation intensity; a distance attenuation coefficient is introduced, which decreases with the increase of the propagation path, to calculate a second-order risk propagation influence value; the first-order risk propagation influence value and the second-order risk propagation influence value are combined by weighting to obtain a cumulative risk influence coefficient of the enterprise;

[0133] According to the cumulative risk influence coefficient, the initial score is adjusted to obtain a final fraud risk score considering the risk propagation effect; based on the final fraud risk score, the decision result of financing guarantee anti-fraud is generated in combination with the guarantee quota of the enterprise.

[0134] Exemplarily, the enterprise risk propagation network is constructed based on a risk propagation adjacency matrix. The adjacency matrix is an n x n matrix, where n is the number of enterprises, and the element aij in the matrix represents the risk propagation strength from enterprise i to enterprise j, with a value ranging from 0 to 1, and the greater the value, the greater the risk propagation strength. The risk propagation path between enterprises is determined by the non-zero elements in the matrix, for example, when a12=0.4, a23=0.5, it indicates that there is a risk propagation path from enterprise 1 to enterprise 2, and then to enterprise 3. The business association degree between enterprises is calculated by analyzing historical transaction data, mainly considering three aspects: transaction amount proportion, transaction frequency and transaction persistence. The transaction amount proportion refers to the proportion of the transaction amount with a certain enterprise in the total transaction amount; the transaction frequency refers to the number of transactions per unit time; and the transaction persistence refers to the time span of continuous transactions. The weighted combination value of these three indicators is used to measure the business association degree, and the weights are set to 0.5, 0.3 and 0.2 respectively. The risk propagation time window is determined by analyzing the propagation time of historical risk events, and the time window is divided into short-term (1-3 months), medium-term (4-6 months) and long-term (7-12 months). For each time window, the attenuation coefficient of risk propagation strength is calculated, which is 1.0 for the short-term window, 0.8 for the medium-term window, and 0.6 for the long-term window. For example, the business association degree between enterprise A and enterprise B is 0.7, and the corresponding element value in the risk propagation adjacency matrix is 0.6, then the risk propagation strength in the short-term window is 0.6 x 1.0 = 0.6, in the medium-term window is 0.6 x 0.8 = 0.48, and in the long-term window is 0.6 x 0.6 = 0.36.

[0135] The calculation of the first-order risk propagation influence value is based on the neighbor nodes directly connected to the enterprise. For enterprise i, the first-order risk propagation influence value is the sum of the initial score of all directly connected enterprises j and the corresponding risk propagation strength. The initial score of enterprise j refers to the initial score of enterprise fraud risk obtained through the foregoing integrated learning framework, and the value is between 0 and 1. The corresponding risk propagation strength is the element value aij in the adjacency matrix, indicating the risk propagation strength from enterprise j to enterprise i. For example, enterprise A has three directly connected enterprises B, C and D, whose initial scores are 0.8, 0.6 and 0.4 respectively, and the corresponding risk propagation strengths are 0.7, 0.5 and 0.3 respectively. Then the first-order risk propagation influence value of enterprise A is 0.8*0.7+0.6*0.5+0.4*0.3=0.87. The second-order risk propagation influence value considers the enterprise nodes indirectly connected, i.e. the case where enterprise i is connected to enterprise k through an intermediate enterprise j. A distance attenuation coefficient is introduced in the calculation, which decreases with the increase of the propagation path, and an exponential attenuation form is adopted, and the attenuation coefficient is set to 0.5. The second-order risk propagation influence value is the sum of the product of the initial score of all second-order connected enterprises k and the corresponding second-order risk propagation strength, where the second-order risk propagation strength is ajk*aji*0.5. For example, enterprise A is indirectly connected to enterprises E and F through enterprise B, the initial scores of E and F are 0.9 and 0.7 respectively, the risk propagation strength from B to E is 0.8, and the risk propagation strength from B to A is 0.7; B is indirectly connected to enterprise G through C, the initial score of G is 0.5, the risk propagation strength from C to G is 0.6, and the risk propagation strength from C to A is 0.5. Then the second-order risk propagation influence value of enterprise A is 0.9*0.8*0.7*0.5+0.7*0.8*0.7*0.5+0.5*0.6*0.5*0.5=0.347.

[0136] The cumulative risk influence coefficient of the enterprise is obtained by weighted combination of the first-order and second-order risk propagation influence values, the weight of the first-order influence is set to 0.7, and the weight of the second-order influence is set to 0.3, so as to reflect that the risk of the directly associated enterprise has greater influence on the current enterprise. Continuing the above example, the cumulative risk influence coefficient of enterprise A is 0.87*0.7+0.347*0.3=0.713. When the risk propagation network is more complex or the risk propagation of a certain industry needs to be considered for a longer distance, the third-order or higher-order risk propagation influence value can also be calculated, and the weights of the influence values of each order can be adjusted appropriately. For example, in the industry with high concentration of supply chain, the weights can be set to first-order 0.6, second-order 0.3 and third-order 0.1.

[0137] The calculation of the final fraud risk score is based on the initial score and the cumulative risk impact coefficient. The adjustment method adopts the weighted average method, that is, the final score is equal to the product of the initial score and (1+the cumulative risk impact coefficient), and then normalized processing is performed to ensure that the final score is between 0 and 1. The normalization processing adopts the min-max scaling method, that is, if the adjusted score exceeds 1, take 1; if it is lower than 0, take 0. For example, the initial score of enterprise A is 0.6, and the cumulative risk impact coefficient is 0.713, so the adjusted score is 0.6x(1+0.713)=1.028, and after normalization, it is 1.0. In order to avoid over-amplification of risk, the upper limit of risk impact can be set, for example, the maximum value of the cumulative risk impact coefficient is limited to 1.5, and even if the calculated value exceeds 1.5, it is calculated as 1.5. In addition, the risk impact can also be adjusted according to the anti-risk ability of the enterprise itself, and the anti-risk ability can be evaluated comprehensively through factors such as enterprise size, asset condition and industry status.

[0138] The generation of the financing guarantee anti-fraud decision result is based on the final fraud risk score and the guarantee limit of the enterprise. The fraud risk score is divided into four levels: low risk (0-0.3), low-medium risk (0.3-0.5), medium-high risk (0.5-0.7) and high risk (0.7 or more). For different risk levels, corresponding guarantee limit adjustment strategies are set. For low-risk enterprises, 100% of the application guarantee limit can be provided; for low-medium-risk enterprises, 80% of the guarantee limit can be provided; for medium-high-risk enterprises, 50% of the guarantee limit can be provided and the value of the collateral can be required to be increased; for high-risk enterprises, the guarantee application is rejected. For example, the final fraud risk score of enterprise A is 0.65, which belongs to medium-high risk, and the application guarantee limit is 1 million yuan, so according to the decision strategy, 500,000 yuan of guarantee limit can be provided, and the value of the collateral can be required to be increased by not less than 3 million yuan. In actual operation, other factors such as the historical credit record of the enterprise, the industry sentiment and the macroeconomic environment can also be used to fine-tune the decision result. In addition, for enterprises with risk scores near the critical value, such as 0.49 or 0.51, a fuzzy decision method can be used to determine the final decision after considering multiple factors.

[0139] The method realizes accurate measurement of risk propagation intensity of different time scales by constructing an enterprise risk propagation network and calculating a risk propagation time window, and adjusts the initial score by the cumulative risk impact coefficient. The method considers the amplification effect of risk propagation on enterprise fraud risk, so that the final fraud risk score is more comprehensive and accurate. The decision result is generated in combination with the guarantee limit of the enterprise, which provides risk-sensitive guarantee strategies for financial institutions and effectively controls business risks.

[0140] In a second aspect of the embodiment of the present application, a financing guarantee anti-fraud system based on artificial intelligence is provided, comprising:

[0141] The first unit is configured to acquire enterprise multi-source heterogeneous data to construct an enterprise feature vector, construct a bidirectional causal graph structure based on the enterprise feature vector, use a directed acyclic graph to represent a bidirectional risk propagation relationship between enterprises, calculate a risk state of each enterprise node according to a risk state of a directly associated enterprise of the node, and use a NOTEARS optimization algorithm to calculate a risk propagation adjacency matrix between enterprises based on the risk state of the node.

[0142] The second unit is configured to construct a metric space of enterprise risk distribution using a Wasserstein distance based on the enterprise feature vector and the risk propagation adjacency matrix, perform distribution robustness optimization in the metric space, train an enterprise risk representation vector using a triplet loss function, and input the enterprise risk representation vector into an ensemble learning framework to obtain an initial score of enterprise fraud risk.

[0143] The third unit is configured to calculate a final fraud risk score considering a risk propagation effect based on the risk propagation adjacency matrix and the initial score, and generate a financing guarantee anti-fraud decision result.

[0144] The third aspect of the embodiment of the present application provides an electronic device, comprising:

[0145] a processor;

[0146] a memory for storing processor-executable instructions;

[0147] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0148] The fourth aspect of the embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions are executed by a processor to implement the method described above.

[0149] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for executing various aspects of the present application.

[0150] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An artificial intelligence based anti-fraud method for financing guarantee characterized in that, The method comprises the following steps: obtaining enterprise multi-source heterogeneous data to construct an enterprise feature vector; constructing a bidirectional causal graph structure based on the enterprise feature vector, wherein the bidirectional causal graph structure uses a directed acyclic graph to represent the bidirectional risk propagation relationship between enterprises; for each enterprise node, calculating the risk state of the node according to the risk state of the directly associated enterprise; using a NOTEARS optimization algorithm to calculate the risk propagation adjacency matrix between enterprises based on the risk state of the node; based on the enterprise feature vector and the risk propagation adjacency matrix, using the Wasserstein distance to construct a metric space of enterprise risk distribution, performing distribution robustness optimization in the metric space, and training an enterprise risk representation vector using a triplet loss function; inputting the enterprise risk representation vector into an ensemble learning framework to obtain an initial score of the enterprise fraud risk; based on the risk propagation adjacency matrix and the initial score, calculating a final fraud risk score considering the risk propagation effect, and generating a financing guarantee anti-fraud decision result.

2. The method of claim 1, wherein, constructing a bidirectional causal graph structure based on the enterprise feature vector, wherein the bidirectional causal graph structure uses a directed acyclic graph to represent the bidirectional risk propagation relationship between enterprises; for each enterprise node, calculating the risk state of the node according to the risk state of the directly associated enterprise; the step of using a NOTEARS optimization algorithm to calculate the risk propagation adjacency matrix between enterprises based on the risk state of the node comprises: calculating the risk correlation degree between enterprises based on the enterprise feature vector to construct an enterprise risk correlation degree matrix; constructing a directed acyclic graph structure based on the enterprise risk correlation degree matrix, wherein the directed acyclic graph structure represents the bidirectional risk propagation relationship between enterprises; for each enterprise node in the directed acyclic graph structure, calculating the risk state of the enterprise node, wherein the risk state of the enterprise node is obtained by weighted combination of the self-feature risk evaluation value and the associated enterprise influence value, and the associated enterprise influence value is the weighted sum of the risk state of the directly associated enterprise of the enterprise node and the corresponding risk propagation weight; composing the risk state of the enterprise node into an enterprise risk state matrix, constructing an optimization objective function based on the enterprise risk state matrix, wherein the optimization objective function comprises a risk state fitting term, a risk propagation time sequence dynamic term and a acyclicity constraint term; using a NOTEARS optimization algorithm to solve the optimization objective function to obtain a risk propagation adjacency matrix between enterprises that satisfies the directed acyclic constraint, wherein the risk propagation adjacency matrix represents the topological structure of the risk propagation between enterprises.

3. The method of claim 2, wherein, the steps of constructing an optimization objective function and solving the optimization objective function comprise: the risk state fitting term is calculated by combining the risk state difference between enterprises with the importance weight, the risk propagation time sequence dynamic term is calculated by combining the time window decay weight with the historical risk propagation intensity, and the acyclicity constraint term is calculated by the loop connection of the risk propagation relationship between enterprises; constructing an augmented Lagrangian function comprising the optimization objective function, wherein the augmented Lagrangian function introduces a Lagrange multiplier term and a penalty term; The L-BFGS algorithm with a second-order Hessian correction term is used to optimize and solve the augmented Lagrange function, and the second-order Hessian correction term is based on a diagonal block matrix constructed according to the risk propagation timing characteristics, and the diagonal block matrix is used to adjust the gradient update direction; When the change value of the optimization target is less than a set threshold and the loop constraint is satisfied, an optimized time sequence risk propagation adjacency matrix is obtained; the propagation intensity in the optimized time sequence risk propagation adjacency matrix is normalized, and the risk propagation relationship between enterprises is determined through time sequence confidence scoring.

4. The method of claim 1, wherein, Based on the enterprise feature vector and the risk propagation adjacency matrix, a measurement space of enterprise risk distribution is constructed using the Wasserstein distance, and distribution robustness optimization is performed in the measurement space. The steps of training the enterprise risk representation vector using a triplet loss function include: The enterprise feature vector is weighted by the risk propagation adjacency matrix, and the weighted enterprise feature vector is mapped to a probability density space using a Gaussian kernel function to obtain an enterprise risk probability distribution; Based on the enterprise risk probability distribution, a Wasserstein distance measurement space is constructed, wherein the Wasserstein distance is obtained by solving the p-th norm of the optimal transport plan between enterprise risk probability distributions; Based on the Wasserstein distance, a triplet loss function is constructed, the classes of enterprise risk distribution pairs are divided according to a preset Wasserstein distance threshold, and a training sample set is constructed; by optimizing the triplet loss function, a representation vector reflecting the characteristics of enterprise risk distribution is learned; Adversarial perturbations are introduced to the representation vector, and a double-layer optimization objective is set, the adversarial perturbation is maximized in the inner layer, and the triplet loss is minimized in the outer layer, the robustness of the representation vector is enhanced, and a final enterprise risk representation vector is obtained; based on the final enterprise risk representation vector, the distance between enterprise pairs is calculated to obtain consistency and discriminability indexes, and the enterprise risk representation result is evaluated.

5. The method of claim 4, wherein, The steps of introducing adversarial perturbations to the representation vector and setting a double-layer optimization objective include: Based on the gradient information of the representation vector, a sensitive direction in the representation vector space is calculated; an initial adversarial perturbation is generated in the sensitive direction, and the amplitude of the initial adversarial perturbation is determined based on the historical risk fluctuation range of the enterprise; A dynamic weight triplet loss function is designed, and the dynamic weight is determined based on the Euclidean distance between enterprise risk representation vector pairs; the direction and amplitude of the initial adversarial perturbation are adjusted according to the Euclidean distance to generate a final adversarial perturbation; A double-layer optimization framework for adversarial training is constructed, the inner-layer optimization finds the optimal perturbation direction by maximizing the triplet loss function under adversarial perturbation, and the outer-layer optimization improves the robustness of the representation vector by minimizing the weighted combination of the original triplet loss function and the adversarial triplet loss function, and the weight coefficient of the weighted combination is dynamically adjusted according to the perturbation influence degree in the training process; The robustness of the representation vector is evaluated based on the validation sample set, the stability of the representation result under different perturbation amplitudes is calculated, and when the stability meets a preset threshold, an enterprise risk representation vector with adversarial robustness is obtained.

6. The method of claim 1, wherein, The step of inputting the enterprise risk representation vector into an ensemble learning framework to obtain an initial score of enterprise fraud risk includes: Based on the enterprise risk representation vector of the historical fraud sample, a feature importance matrix is constructed, and a principal component analysis method is used to reduce the dimension of the feature importance matrix to obtain a feature combination mode that has an identifying effect on fraud; The enterprise risk representation vector is projected into the subspace corresponding to the feature combination mode, a decision tree ensemble structure based on gradient boosting is constructed, and each decision tree corresponds to a feature combination mode; the prediction results of each decision tree are weighted and fused to obtain an initial score of enterprise fraud risk.

7. The method of claim 1, wherein, According to the risk propagation adjacency matrix and the initial score, a final fraud risk score considering the risk propagation effect is calculated, and a financing guarantee anti-fraud decision result is generated. An enterprise risk propagation network is constructed, the risk propagation paths between enterprises are determined based on the risk propagation adjacency matrix, and the risk propagation time window is calculated according to the business association degree and historical transaction data between enterprises to obtain the risk propagation intensity under different time windows; For each enterprise node, based on the initial score of its adjacent enterprise nodes and the corresponding risk propagation intensity, a first-order risk propagation influence value is calculated; a distance attenuation coefficient is introduced, which decreases with the increase of the propagation path, and a second-order risk propagation influence value is calculated; the first-order risk propagation influence value and the second-order risk propagation influence value are weighted and combined to obtain a cumulative risk influence coefficient of the enterprise; According to the cumulative risk influence coefficient, the initial score is adjusted to obtain a final fraud risk score considering the risk propagation effect; based on the final fraud risk score, in combination with the guarantee limit of the enterprise, a financing guarantee anti-fraud decision result is generated.

8. An artificial intelligence based anti-fraud system for financing guarantee, for implementing the method of any of the preceding claims 1-7, characterized in that, It includes: A first unit for acquiring enterprise multi-source heterogeneous data to construct an enterprise feature vector; A bidirectional causal graph structure is constructed based on the enterprise feature vector, and the bidirectional causal graph structure uses a directed acyclic graph to represent the bidirectional risk propagation relationship between enterprises; For each enterprise node, the risk state of the node is calculated according to the risk state of its directly associated enterprises; A NOTEARS optimization algorithm is used to calculate the risk propagation adjacency matrix between enterprises based on the risk state of the node; A second unit for constructing an enterprise risk distribution measurement space using a Wasserstein distance based on the enterprise feature vector and the risk propagation adjacency matrix, performing distribution robustness optimization in the measurement space, and training an enterprise risk representation vector using a triplet loss function; input the enterprise risk characterization vector into an integrated learning framework to obtain an initial score of enterprise fraud risk; a third unit configured to calculate a final fraud risk score considering the risk propagation effect according to the risk propagation adjacency matrix and the initial score, and generate a financing guarantee anti-fraud decision result.

9. An electronic device, comprising: comprise: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 7. The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 7.

Images