A financial fraud detection method and system based on hyperbolic space energy calculation

Abstract

The application discloses a financial fraud detection method and system based on hyperbolic space energy calculation, and belongs to the technical field of anomaly detection. The method comprises the following steps: constructing a data set and a hyperbolic graph convolution encoder, and training and testing the hyperbolic graph convolution encoder through the data set; obtaining target graph data, processing the target graph data according to the tested hyperbolic graph convolution encoder, and obtaining node potential representation features; calculating an anomaly score value of each target node in the node potential representation features; and judging whether the target price node is abnormal according to the anomaly score value. The application proposes a node energy driven anomaly detection mechanism, innovatively defines the energy change amount of a node in a feature space and a hyperbolic space, and quantifies the deviation degree of the node, so that more concealed structural abnormal behaviors can be identified, and the detection capability for complex abnormal patterns in a real financial scene can be effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of anomaly detection, and particularly relates to a financial fraud detection method and system based on hyperbolic space energy calculation. BACKGROUND

[0002] In recent years, with the acceleration of the digitalization process of the financial industry and the continuous complication of business models, financial fraud behaviors present the development trend of scale, intelligence and high concealment, which poses a serious challenge to the safety and stability of the financial system. Under this background, how to accurately identify potential fraudulent transactions has become an important topic in the research and practice of financial technology.

[0003] With the continuous development of deep learning, graph neural networks have gradually become the focus of research in the field of fraud detection because they can model the complex association structure between account nodes and transaction behaviors in financial transaction networks. At the same time, as a typical unsupervised learning model, autoencoders rely on reconstruction error to identify samples deviating from the normal pattern and have been widely used in anomaly detection tasks.

[0004] At present, hyperbolic geometry has also shown potential in anomaly detection tasks due to its natural advantages in describing hierarchical, tree-like and power-law distributed structural data. However, the existing technology has the shortcomings of modeling non-Euclidean structures, improving representation ability and reducing supervision dependence when introducing hyperbolic geometry into the financial fraud detection scenario, which further leads to the problem of low accuracy of financial fraud detection based on hyperbolic space energy calculation. SUMMARY

[0005] The purpose of the embodiments of the present application is to provide a financial fraud detection method and system based on hyperbolic space energy calculation, which can solve the technical problem of low accuracy of financial fraud detection in the prior art.

[0006] In order to solve the above technical problems, the present application is implemented as follows:

[0007] In a first aspect, the embodiments of the present application provide a financial fraud detection method based on hyperbolic space energy calculation, which comprises:

[0008] Obtaining multiple pieces of graph structure data, and constructing a data set according to the multiple pieces of graph structure data;

[0009] Constructing a hyperbolic graph convolutional encoder, training and testing the hyperbolic graph convolutional encoder according to the data set, and obtaining the hyperbolic graph convolutional encoder after testing;

[0010] Obtaining target graph data, processing the target graph data according to the hyperbolic graph convolutional encoder after testing, and obtaining node latent representation features;

[0011] Anomaly score calculation is performed on each target node in the node potential representation feature, and an anomaly score value of each target node is obtained;

[0012] If the anomaly score value is greater than a score threshold value, the target node corresponding to the anomaly score value is determined as an abnormal risk node;

[0013] The anomaly score calculation on each target node in the node potential representation feature to obtain the anomaly score value of each target node is specifically:

[0014] Hyperbolic energy calculation is performed on the target node according to a hyperbolic energy calculation formula, and a hyperbolic energy value of the target node is obtained;

[0015] Feature space energy calculation is performed on the target node according to a feature space energy calculation formula, and a space energy value of the target node is obtained;

[0016] The anomaly score value of the target node is calculated according to the hyperbolic energy value and the space energy value of the target node;

[0017] The anomaly score value of the target node is calculated according to the hyperbolic energy value and the space energy value of the target node, and specifically:

[0018] The energy difference value of the target node is calculated according to the hyperbolic energy value and the space energy value of the target node;

[0019] Second-order calculation is performed on the energy difference value of the target node, and a Hessian matrix norm of the target node is obtained;

[0020] The node degree of the target node is obtained, and the anomaly score value of the target node is calculated according to the energy difference value, the Hessian matrix norm, and the node degree;

[0021] The hyperbolic energy calculation formula is represented as follows:

[0022] ,

[0023] wherein, the hyperbolic energy value of the i-th target node, the total number of target nodes, the i-th target node embedding vector, the embedding vector of a target node adjacent to the i-th target node, the i-th target node embedding vector, the length square of the difference value, the i-th target node embedding vector, the i-th target node embedding vector, the i-th target node embedding vector, the i-th target node embedding vector, the i-th target node embedding vector, The square of the modulus, express The square of the modulus, This represents the regularization constant, used to prevent the denominator from being zero;

[0024] The formula for calculating the energy of the characteristic space is expressed as follows:

[0025] ,

[0026] in, Indicates the first Spatial energy value of each target node;

[0027] The anomaly score of the target node is calculated based on the energy difference, the Hessian matrix norm, and the node degree, and is specifically expressed by the following formula:

[0028] ,

[0029] in, Indicates the first The anomaly score of each target node. Indicates the first The Hessian matrix norm of each target node. Indicates the first The degree of each target node.

[0030] As an optional implementation of the first aspect of this application, the step of training and testing the hyperbolic graph convolutional encoder based on the dataset to obtain the tested hyperbolic graph convolutional encoder specifically includes:

[0031] Construct a target decoder and set the target decoder at the back end of the hyperbolic graph convolutional encoder to obtain a trained network model;

[0032] The dataset is divided into a training set and a test set according to a preset ratio, and a target loss function is constructed.

[0033] The training network model is trained iteratively multiple times based on the target loss function and the training set until the parameters of the training network model converge.

[0034] The trained network model after parameter convergence is tested according to the test set to obtain the tested trained network model.

[0035] The hyperbolic graph convolutional encoder is called from the trained network model after the test to obtain the hyperbolic graph convolutional encoder after the test.

[0036] The target decoder includes a parallel feature reconstruction decoder and a structure reconstruction decoder.

[0037] As an optional implementation of the first aspect of the application, the processing procedure of the training network model on the graph structure data in the training set is specifically:

[0038] According to the hyperbolic graph convolutional encoder in the training network model, the graph structure data is processed to obtain a training latent representation;

[0039] According to the feature reconstruction decoder and the structure reconstruction decoder in the target decoder, the training latent representation is processed respectively to obtain a reconstructed node feature and a reconstructed adjacency matrix respectively.

[0040] As an optional implementation of the first aspect of the application, the hyperbolic graph convolutional encoder includes a hyperbolic space mapping layer and a hyperbolic convolutional neural network; the hyperbolic graph convolutional encoder processes the target graph data to obtain a node latent representation feature, specifically:

[0041] According to the hyperbolic space mapping layer, the target graph data is mapped and processed to obtain a hyperbolic mapping feature;

[0042] According to the hyperbolic convolutional neural network, the hyperbolic mapping feature is convoluted and processed to obtain the node latent representation feature;

[0043] The hyperbolic convolutional neural network includes a plurality of stacked hyperbolic graph convolutional layers, each of which includes an input layer, a hidden layer and an output layer.

[0044] Secondly, the application provides a financial fraud detection system based on hyperbolic space energy calculation, which includes:

[0045] A data construction module acquires a plurality of graph structure data and constructs a data set according to the plurality of graph structure data;

[0046] A network training module constructs a hyperbolic graph convolutional encoder, trains and tests the hyperbolic graph convolutional encoder according to the data set, and obtains the hyperbolic graph convolutional encoder after testing;

[0047] A first processing module acquires target graph data, processes the target graph data according to the hyperbolic graph convolutional encoder after testing, and obtains a node latent representation feature;

[0048] A second processing module calculates an anomaly score of each target node in the node latent representation feature to obtain an anomaly score value of each target node;

[0049] The comparison output module: if the anomaly score value is greater than a score threshold, the target node corresponding to the anomaly score value is determined as an abnormal risk node.

[0050] In a third aspect, an electronic device is provided, which includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, and the program or instructions, when executed by the processor, implement the steps of the method according to the first aspect.

[0051] In a fourth aspect, a readable storage medium is provided, and the readable storage medium stores a program or instructions, and the program or instructions, when executed by a processor, implement the steps of the method according to the first aspect.

[0052] In the embodiments of the present application, compared with the prior art, the following technical effects are achieved:

[0053] (1) The present application systematically introduces hyperbolic geometry into the framework of graph anomaly detection, and constructs a hyperbolic graph convolutional encoder based on the Poincaré ball model, thereby realizing efficient modeling of the hierarchical and non-Euclidean structure in the financial transaction network. Compared with the traditional graph neural network based on Euclidean space, the present application can more compactly express complex topological properties such as tree structure and power-law distribution in the hidden space, thereby significantly improving the representation efficiency and robustness of the graph structure information, and making the hyperbolic graph convolutional encoder more suitable for processing the chain, hierarchical and gang interaction relationships commonly existing in financial networks.

[0054] (2) The present application belongs to an unsupervised anomaly detection method, and does not need to use any artificially labeled fraud samples, thereby fundamentally reducing the dependence on high-quality labeled data. In view of the fact that fraud labels are scarce and lag in the financial scenario, the unsupervised paradigm of the present application can effectively solve the problem of insufficient training caused by sample imbalance and missing fraud labels in traditional supervised detection methods, and is more suitable for the changing risk scenarios in real application environments.

[0055] (3) The present application proposes a node energy driven anomaly detection mechanism, and innovatively defines a measure of the energy change of a node in the feature space and the hyperbolic space, which is used to quantify the degree of deviation. This mechanism simultaneously considers the position distribution of the node in the Poincaré ball and the hyperbolic distance relationship with the neighbor nodes, and can sensitively capture structural anomalies and local geometric deviations. Compared with the traditional autoencoder method which only relies on feature reconstruction error, the present application can identify more hidden structural abnormal behaviors, and effectively improve the detection ability of complex abnormal patterns in real financial scenarios.

[0056] (4) The application constructs a multi-dimensional comprehensive anomaly scoring mechanism, which fuses the node energy difference, the second-order gradient norm of the energy difference and the node degree to form a more discriminative anomaly score. The fusion strategy can reflect the feature anomaly, structural anomaly and topological anomaly of the node at the same time, so that the model has the ability to capture multiple types of anomaly patterns, and still maintains good generalization performance in the case of scarce anomaly samples or the emergence of new fraud patterns. BRIEF DESCRIPTION OF DRAWINGS

[0057] Fig. 1 is a flowchart of a financial fraud detection method based on hyperbolic space energy calculation provided by some embodiments of the application;

[0058] Fig. 2 is a training process diagram of training a network model in a financial fraud detection method based on hyperbolic space energy calculation provided by some embodiments of the application;

[0059] Fig. 3 is a calculation process diagram of an anomaly score in a financial fraud detection method based on hyperbolic space energy calculation provided by some embodiments of the application. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0061] The terms "first", "second", etc. in the specification and claims of the application are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the application can be implemented in an order other than those illustrated or described here. In addition, "and / or" in the specification and claims indicates at least one of the connected objects, and the character " / ", generally indicates that the front and rear associated objects are in an "or" relationship.

[0062] The financial fraud detection method and system based on hyperbolic space energy calculation provided by the embodiments of the application will be described in detail below with reference to the drawings and specific embodiments and their application scenarios.

[0063] EMBODIMENT

[0064] A financial fraud detection method based on hyperbolic space energy calculation comprises:

[0065] S100: Obtain multiple graph structure data, and construct a data set according to the multiple graph structure data;

[0066] It should be noted that the plurality of graph structure data can come from a public data set or self-captured images. If it is a self-captured image, it needs to be pre-processed before training to achieve the uniformity of the self-captured image. Then, the graph structure data in the public data set or the self-captured graph structure data is used to construct the data set of the embodiment.

[0067] S200: Constructing a hyperbolic graph convolutional encoder, training and testing the hyperbolic graph convolutional encoder according to the data set, and obtaining the tested hyperbolic graph convolutional encoder;

[0068] As an optional implementation of the embodiment, the hyperbolic graph convolutional encoder is trained and tested according to the data set in S200, and the tested hyperbolic graph convolutional encoder is obtained. Specifically, the hyperbolic graph convolutional encoder is trained and tested according to the data set in S200, and the tested hyperbolic graph convolutional encoder is obtained.

[0069] S210: Constructing a target decoder, setting the target decoder at the back end of the hyperbolic graph convolutional encoder, and obtaining a training network model;

[0070] S220: Dividing the data set into a training set and a test set according to a preset proportion, and constructing a target loss function;

[0071] S230: Iteratively training the training network model according to the target loss function and the training set until the parameters of the training network model converge;

[0072] S240: Testing the training network model with the converged parameters according to the test set, and obtaining a tested training network model;

[0073] S250: Calling the hyperbolic graph convolutional encoder from the tested training network model, and obtaining the tested hyperbolic graph convolutional encoder;

[0074] In S210, the target decoder includes a parallel feature reconstruction decoder and a structure reconstruction decoder.

[0075] It should be noted that a target decoder for the training process is constructed, the target decoder includes a parallel feature reconstruction decoder and a structure reconstruction decoder; the target decoder is arranged at the back end of the hyperbolic graph convolutional encoder to form a training network model together with the hyperbolic graph convolutional encoder; a data set is divided into a training set and a test set according to a ratio of 4:1, and a target loss function is constructed; the target loss function and the training set perform multi-round iterative training on the training network model; the target loss function calculates a loss value in the training process, and a Riemannian Adam optimizer is used for back propagation to optimize parameters of the training network model; the Riemannian Adam optimizer can better adapt to parameter optimization of the training network model, and unlike an Adam optimizer in a traditional Euclidean space, the Riemannian Adam can correctly update gradients in a hyperbolic space; through 100 rounds of iterative training, a learning rate is set to 0.0005, and a random dropout ratio is set to 0.2; when parameters of the training network model reach convergence, the training network model is tested through the test set to measure its ability in unknown samples, and a test result is evaluated through an evaluation index; if the evaluation result is qualified, the training and testing of the training network model are completed; if the evaluation result is unqualified, parameters or training strategies of the training network model need to be adjusted, and the training is performed again until the evaluation result is qualified; finally, the hyperbolic graph convolutional encoder in the training network model is called out after testing to obtain a separate hyperbolic graph convolutional encoder after testing.

[0076] As an optional implementation of the embodiment, the processing process of the graph structure data in the training set by the training network model in S230 is specifically as follows:

[0077] S231: The hyperbolic graph convolutional encoder in the training network model is used to process the graph structure data to obtain a training latent representation.

[0078] S232: The feature reconstruction decoder and the structure reconstruction decoder in the target decoder are used to process the training latent representation respectively to obtain a reconstructed node feature and a reconstructed adjacency matrix.

[0079] It should be noted that in the training process, the processing process of each graph structure data in the training set by the training network model is that the hyperbolic graph convolutional encoder in the training network model processes the graph structure data to obtain the training latent representation of the graph structure data; then the training latent representation enters the target decoder, and the feature reconstruction decoder and the structure reconstruction decoder in the target decoder process the training latent representation respectively to obtain the reconstructed node feature and the reconstructed adjacency matrix respectively; at this time, the forward processing result of the network for the graph structure data, the target loss function calculates the loss value of the graph structure data through the reconstructed node feature and the reconstructed adjacency matrix, the training network model processes a batch of graph structure data in the training set at the same time, obtains the loss value corresponding to each graph structure data in the batch of training data, then averages the loss values to obtain the average loss of the batch of training data, and the training network model is optimized through the average loss. Wherein,

[0080] As a specific embodiment of the present embodiment, the target loss function is represented by the following formula:

[0081]

[0082]

[0083] ,

[0084] Wherein, represents the loss value, represents the error of the reconstructed node feature, represents the error of the reconstructed adjacency matrix, represents the balance parameter, represents the reconstructed node feature, represents the original node feature corresponding to , represents the reconstructed adjacency matrix, represents the original adjacency matrix corresponding to .

[0085] It should be noted that the feature reconstruction decoder processes the training latent representation by multiple layers of hyperbolic graph convolution to obtain the attribute representation in the hyperbolic space Then the is mapped back to the Euclidean space through the hyperbolic mapping to obtain the reconstructed node feature ; at the same time, the structure reconstruction decoder processes the training latent representation by multiple layers of hyperbolic graph convolution to gradually restore the structure feature, and the training latent representation is processed by multiple layers of hyperbolic graph convolution in turn to obtain the structure representation in the hyperbolic space . Then the Mapping back to Euclidean space, get the structure representation ; finally, through matrix multiplication , get the reconstructed adjacency matrix , is the transpose matrix of .

[0086] S300: Obtain target graph data, process the target graph data according to the hyperbolic graph convolutional encoder after testing, and obtain node latent representation features;

[0087] As an optional implementation of the embodiment, the hyperbolic graph convolutional encoder in S300 includes a hyperbolic space mapping layer and a hyperbolic convolutional neural network;

[0088] The hyperbolic graph convolutional encoder processes the target graph data to obtain node latent representation features, specifically:

[0089] S310: Map the target graph data according to the hyperbolic space mapping layer to obtain hyperbolic mapping features;

[0090] S320: Convolve the hyperbolic mapping features according to the hyperbolic convolutional neural network to obtain node latent representation features;

[0091] Wherein, the hyperbolic convolutional neural network in S320 includes a plurality of stacked hyperbolic graph convolutional layers, each hyperbolic graph convolutional layer includes an input layer, a hidden layer and an output layer.

[0092] It should be noted that the hyperbolic graph convolutional encoder includes a hyperbolic space mapping layer and a hyperbolic convolutional neural network, the target graph data is mapped and processed by the hyperbolic space mapping layer to obtain hyperbolic mapping features; then the hyperbolic mapping features are convolved by the hyperbolic convolutional neural network to obtain node latent representation features; the hyperbolic convolutional neural network is stacked by a plurality of hyperbolic graph convolutional layers; each hyperbolic graph convolutional layer contains three parts of input layer, hidden layer and output layer, and the input layer, hidden layer and output layer all contain hyperbolic linear transformation, hyperbolic aggregation operation and hyperbolic activation function. The input layer projects the features in the Euclidean space to the hyperbolic space and performs graph convolution operation; the hidden layer gradually refines the features by hyperbolic graph convolution and superimposes nonlinear mapping, so that the hyperbolic graph convolutional encoder can capture more complex neighborhood and hierarchical structure relationship, and the output layer can obtain the hidden vector in the hyperbolic space. The input layer of the first hyperbolic graph convolutional layer processes the hyperbolic mapping features, and the output layer of the last hyperbolic graph convolutional layer outputs the node latent representation features.

[0093] S400: Calculate the anomaly score of each target node in the node latent representation features to obtain the anomaly score value of each target node;

[0094] As an optional implementation of the embodiment, the abnormal score of each target node in the node latent representation feature is calculated in S400, and the abnormal score value of each target node is obtained, specifically as follows:

[0095] S410: Hyperbolic energy of the target node is calculated according to the hyperbolic energy calculation formula, and the hyperbolic energy value of the target node is obtained.

[0096] S420: Feature space energy of the target node is calculated according to the feature space energy calculation formula, and the space energy value of the target node is obtained.

[0097] S430: The abnormal score value of the target node is calculated according to the hyperbolic energy value and the space energy value of the target node.

[0098] It should be noted that for each target node in the node latent representation feature, first, the hyperbolic energy of the target node is calculated by the hyperbolic energy calculation formula, and the feature space energy of the target node is calculated by the feature space energy calculation formula, and then the hyperbolic energy value and the space energy value of the target node are obtained, and then the abnormal score value of the target node is calculated according to the hyperbolic energy value and the space energy value of the target node; and then the abnormal score of each target node in the node latent representation feature is calculated by the above method, and the abnormal score value of each target node is obtained.

[0099] As an optional implementation of the embodiment, the abnormal score value of the target node is calculated according to the hyperbolic energy value and the space energy value of the target node in S430, specifically as follows:

[0100] S431: The energy difference value of the target node is calculated according to the hyperbolic energy value and the space energy value of the target node.

[0101] S432: The energy difference value of the target node is calculated according to the hyperbolic energy value and the space energy value of the target node.

[0102] S433: The node degree of the target node is obtained, and the abnormal score value of the target node is calculated according to the energy difference value, the Hessian matrix norm and the node degree.

[0103] It should be noted that the energy difference of the target node is calculated using its hyperbolic energy value and spatial energy value. Then, a second-order calculation is performed on this energy difference to obtain the Hessian matrix norm of the target node. The Hessian matrix norm reflects the local curvature change of the target node in the feature space. Finally, the node degree of the target node is obtained, and the anomaly score is calculated based on the energy difference, the Hessian matrix norm, and the node degree. Specifically, the Hessian matrix norm is multiplied by the node degree to obtain the first part of the result; the square root of the energy difference is taken to obtain the second part of the result; the second part of the result is multiplied by the node degree and added to the first part of the result to obtain the anomaly score. The energy difference, the Hessian matrix norm reflecting the curvature of the feature space, and the node degree are organically combined to form a comprehensive anomaly score. The energy difference reflects the change in the energy distribution of the target node before and after encoding. Anomaly target nodes are usually difficult to encode accurately by the model, and their energy differences are usually large. The Hessian matrix norm captures the local curvature of the feature space; anomaly target nodes are often located in anomalous regions of the feature space and have large curvature changes. Using node degree as a weighting factor makes the abnormal behavior of highly connected target nodes more pronounced.

[0104] As a specific implementation of this embodiment, the hyperbolic energy calculation formula is expressed as follows:

[0105] ,

[0106] in, Indicates the first The hyperbolic energy value of each target node. This represents the total number of target nodes. No. Embedded vectors of target nodes Indicates the relationship with the first The embedding vectors of the target nodes adjacent to each target node. express and The square of the modulus of the difference express The square of the modulus, express The square of the modulus, This represents the regularization constant, used to prevent the denominator from being zero;

[0107] The formula for calculating the energy of the characteristic space is expressed as follows:

[0108] ,

[0109] in, Indicates the first Spatial energy value of each target node;

[0110] The abnormal score value of the target node is calculated according to the energy difference value, the Hessian matrix norm and the node degree, and is specifically represented by the following formula:

[0111] ,

[0112] Wherein, represents the abnormal score of the i-th target node, represents the Hessian matrix norm of the i-th target node, represents the node degree of the i-th target node.

[0113] S500: If the abnormal score value is greater than the score threshold value, the target node corresponding to the abnormal score value is determined as an abnormal risk node.

[0114] ​​​According to the financial fraud detection method based on hyperbolic space energy calculation, the hyperbolic geometry system is introduced into the graph anomaly detection framework, the hyperbolic graph convolution encoder is constructed based on the Poincare ball model, and the hierarchical and non-Euclidean structure in the financial transaction network is efficiently modeled. Compared with the traditional graph neural network based on Euclidean space, the tree structure, power-law distribution and other complex topological characteristics can be more compactly expressed in the hidden space, thereby the representation efficiency and robustness of the graph structure information are significantly improved, and the hyperbolic graph convolution encoder is more suitable for processing the chain, hierarchical and gang interaction relationships commonly existing in the financial network. The financial fraud detection method based on hyperbolic space energy calculation belongs to an unsupervised anomaly detection method, and does not need to use any artificially labeled fraud samples, thereby fundamentally reducing the dependence on high-quality labeled data. Since the fraud labels are rare and lag in the financial scenario, the unsupervised paradigm can effectively solve the problem of insufficient training caused by sample imbalance and missing fraud labels in the traditional supervised detection method, and is more suitable for the changing risk scenarios in the real application environment. The financial fraud detection method based on hyperbolic space energy calculation proposes a node energy driven anomaly detection mechanism, and innovatively defines the energy change of the node in the feature space and the hyperbolic space to measure the deviation degree. The mechanism considers the position distribution of the node in the Poincare ball and the hyperbolic distance relationship with the neighbor nodes, and can sensitively capture the structural anomaly and local geometric deviation. Compared with the traditional autoencoder method which only depends on the feature reconstruction error, the financial fraud detection method based on hyperbolic space energy calculation can identify more hidden structural abnormal behaviors, and effectively improve the detection ability of complex abnormal patterns in the real financial scenario. The financial fraud detection method based on hyperbolic space energy calculation constructs a multi-dimensional comprehensive anomaly scoring mechanism, and fuses the node energy difference, the second-order gradient norm of the energy difference and the node degree to form a more discriminative anomaly score. The fusion strategy can reflect the feature anomaly, structural anomaly and topological anomaly of the node at the same time, so that the model has the ability to capture multiple types of abnormal patterns, and still maintains good generalization performance in the case of rare abnormal samples or new fraud patterns.

[0115] It should be noted that the financial fraud detection method based on hyperbolic space energy calculation provided in the embodiments of the present application can be a financial fraud detection system based on hyperbolic space energy calculation, or a control module in the financial fraud detection system based on hyperbolic space energy calculation for executing the financial fraud detection method based on hyperbolic space energy calculation. In the embodiments of the present application, the financial fraud detection system based on hyperbolic space energy calculation is taken as an example to execute the financial fraud detection method based on hyperbolic space energy calculation, and the financial fraud detection method based on hyperbolic space energy calculation provided in the embodiments of the present application is described.

[0116] The financial fraud detection system based on hyperbolic space energy calculation comprises:

[0117] The data construction module: acquire multiple pieces of graph structure data, and construct a data set according to the multiple pieces of graph structure data;

[0118] The network training module: construct a hyperbolic graph convolutional encoder, train and test the hyperbolic graph convolutional encoder according to the data set, and obtain the hyperbolic graph convolutional encoder after testing;

[0119] The first processing module: acquire target graph data, process the target graph data according to the hyperbolic graph convolutional encoder after testing, and obtain node latent representation features;

[0120] The second processing module: perform abnormal score calculation on each target node in the node latent representation features, and obtain an abnormal score value of each target node;

[0121] The comparison output module: if the abnormal score value is greater than a score threshold, the target node corresponding to the abnormal score value is determined as an abnormal risk node.

[0122] The financial fraud detection system based on hyperbolic space energy calculation in the embodiments of the present application can be a device, or a component, an integrated circuit, or a chip in a terminal. The device can be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a notebook computer, a palm computer, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), and the non-mobile electronic device can be a server, a Network Attached Storage (NAS), a personal computer (PC), and the like, which are not limited in the embodiments of the present application.

[0123] The financial fraud detection system based on hyperbolic space energy calculation in the embodiments of the present application can be a device with an operating system. The operating system can be an Android operating system, an ios operating system, or other possible operating systems, which are not limited in the embodiments of the present application.

[0124] The financial fraud detection system based on hyperbolic space energy calculation provided in the embodiments of the present application can realize Figs. 1 to 3 the processes and effects of the financial fraud detection method based on hyperbolic space energy calculation in the method embodiment, which are not repeated here to avoid repetition.

[0125] Optionally, the embodiments of the present application further provide an electronic device, comprising a processor, a memory, a program or instructions stored in the memory and executable in the processor, which, when executed by the processor, implement each process of the above-mentioned financial fraud detection method based on hyperbolic space energy calculation and achieve the same technical effects. To avoid repetition, details are not described here.

[0126] The embodiments of the present application further provide a readable storage medium, which stores a program or instructions, which, when executed by a processor, implement each process of the above-mentioned financial fraud detection method based on hyperbolic space energy calculation and achieve the same technical effects. To avoid repetition, details are not described here.

[0127] The processor is the processor of the electronic device in the above-mentioned embodiments. The readable storage medium includes a computer readable storage medium, such as a computer readable memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), a magnetic disk or an optical disk, etc.

[0128] It should be noted that, in this document, the terms "comprising", "including", or any other variant thereof are intended to cover non-exclusive inclusions, so that processes, methods, articles, or devices including a series of elements not only include those elements, but also include other elements not explicitly listed, or include elements inherent to such processes, methods, articles, or devices. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of another identical element in the process, method, article, or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to the order of performing functions as shown or discussed, but can also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method can be performed in an order different from that described, and various steps can also be added, omitted, or combined. In addition, the features described with reference to certain examples can be combined in other examples.

[0129] Those skilled in the art can clearly understand the above-mentioned embodiment method can be realized by means of software and the necessary general hardware platform, of course, also can be through hardware, but many cases the former is the better embodiment. Based on such understanding, the technical solutions of the present application essentially or say the part of the contribution to the prior art can be embodied in the form of software products, the computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disc), including a number of instructions to make a terminal (may be a mobile phone, computer, server, air conditioner, or network equipment, etc.) executes the method described in various embodiments of the present application.

[0130] The embodiments of the present application are described above in conjunction with the drawings, but the present application is not limited to the above-mentioned specific embodiments, the above-mentioned specific embodiments are only illustrative, but not limited, those skilled in the art can make many forms without departing from the purpose of the present application and the scope of the claims under the inspiration of the present application, all belong to the protection of the present application.

Claims

1. A financial fraud detection method based on hyperbolic space energy computation, characterized in that, The method comprises: obtaining multiple graph structure data, and constructing a data set according to the multiple graph structure data; constructing a hyperbolic graph convolutional encoder, training and testing the hyperbolic graph convolutional encoder according to the data set, and obtaining the hyperbolic graph convolutional encoder after testing; obtaining target graph data, processing the target graph data according to the hyperbolic graph convolutional encoder after testing, and obtaining node latent representation features; calculating an anomaly score of each target node in the node latent representation features to obtain an anomaly score value of each target node; if the anomaly score value is greater than a score threshold, the target node corresponding to the anomaly score value is determined as an abnormal risk node; the anomaly score calculation of each target node in the node latent representation features to obtain the anomaly score value of each target node is specifically: calculating the hyperbolic energy of the target node according to a hyperbolic energy calculation formula to obtain a hyperbolic energy value of the target node; calculating the feature space energy of the target node according to a feature space energy calculation formula to obtain a space energy value of the target node; calculating the anomaly score value of the target node according to the hyperbolic energy value and the space energy value of the target node; the calculation of the anomaly score value of the target node according to the hyperbolic energy value and the space energy value of the target node is specifically: calculating the energy difference value of the target node according to the hyperbolic energy value and the space energy value of the target node; performing second-order calculation on the energy difference value of the target node to obtain the Hessian matrix norm of the target node; obtaining the node degree of the target node, and calculating the anomaly score value of the target node according to the energy difference value, the Hessian matrix norm and the node degree; the hyperbolic energy calculation formula is represented as follows: ， wherein, represents a hyperbolic energy value of the th target node, represents a total number of target nodes, represents a th target node embedding vector, represents an embedding vector of a target node adjacent to the th target node, represents a square of a module of a difference, represents a square of a module of represents a square of a module of represents a regularization constant for preventing a denominator from being zero; the feature space energy calculation formula is represented as follows: ， wherein, represents the spatial energy value of the target node. the calculation of the anomaly score value of the target node according to the energy difference value, the Hessian matrix norm and the node degree is specifically represented by the following formula: ， wherein, represents an anomaly score of the th target node, represents a Huber matrix norm of the th target node, represents a node degree of the th target node. the training and testing of the hyperbolic graph convolutional encoder according to the data set to obtain the hyperbolic graph convolutional encoder after testing is specifically: constructing a target decoder, setting the target decoder at the back end of the hyperbolic graph convolutional encoder to obtain a training network model; dividing the data set into a training set and a test set according to a preset proportion, and constructing a target loss function; performing multiple rounds of iterative training on the training network model according to the target loss function and the training set until the parameters of the training network model converge; testing the training network model after parameter convergence according to the test set to obtain the training network model after testing; calling the hyperbolic graph convolutional encoder from the training network model after testing to obtain the hyperbolic graph convolutional encoder after testing; wherein the target decoder comprises a parallel feature reconstruction decoder and a structure reconstruction decoder; The hyperbolic graph convolutional encoder comprises a hyperbolic space mapping layer and a hyperbolic convolutional neural network; the hyperbolic graph convolutional encoder processes the target graph data to obtain node latent representation features, specifically as follows: The target graph data is mapped according to the hyperbolic space mapping layer to obtain hyperbolic mapping features; The hyperbolic mapping features are convoluted according to the hyperbolic convolutional neural network to obtain the node latent representation features. The hyperbolic convolutional neural network comprises a plurality of stacked hyperbolic graph convolutional layers, each of which comprises an input layer, a hidden layer and an output layer.

2. The financial fraud detection method based on hyperbolic space energy calculation of claim 1, wherein, The processing procedure of the training network model on the graph structure data in the training set is specifically as follows: The graph structure data is processed according to the hyperbolic graph convolutional encoder in the training network model to obtain training latent representations; The training latent representations are processed by the feature reconstruction decoder and the structure reconstruction decoder in the target decoder respectively to obtain reconstructed node features and reconstructed adjacency matrices respectively.

3. A financial fraud detection system based on hyperbolic space energy computation, capable of implementing a financial fraud detection method based on hyperbolic space energy computation according to any one of claims 1-2, characterized in that, The system comprises: A data construction module that acquires a plurality of graph structure data and constructs a data set according to the plurality of graph structure data; A network training module that constructs a hyperbolic graph convolutional encoder, trains and tests the hyperbolic graph convolutional encoder according to the data set, and obtains the hyperbolic graph convolutional encoder after testing; A first processing module that acquires target graph data, processes the target graph data according to the hyperbolic graph convolutional encoder after testing, and obtains node latent representation features; A second processing module that calculates an anomaly score for each target node in the node latent representation features to obtain an anomaly score value of each target node; A comparison output module that determines that the target node corresponding to the anomaly score value is an abnormal risk node if the anomaly score value is greater than a score threshold.

4. An electronic device, comprising: The processor, the memory and the program or instructions stored on the memory and executable on the processor are included, and the program or instructions are executed by the processor to implement the steps of the financial fraud detection method based on hyperbolic space energy calculation according to any one of claims 1-2.

5. A readable storage medium, characterized by, The program or instructions are stored on the readable storage medium, and the program or instructions are executed by the processor to implement the steps of the financial fraud detection method based on hyperbolic space energy calculation according to any one of claims 1-2.

Images