Financial fraud detection method and device based on graph compression technology, and medium

Abstract

The invention discloses a financial fraud detection method and device based on a graph compression technology, and a medium, and belongs to the technical field of graph machine learning. In order to solve the problems that key structure information capable of reflecting collaborative complex transactions is lost and fraud detection performance is affected due to the fact that node and edge synthesis needs to be carried out during coarsening in an existing method, the method comprises the steps that 1, a financial transaction network data set is obtained, and financial transaction platform users serve as nodes; the financial transaction relation or the dynamic relation between the users is used as an edge, and a financial transaction network is constructed; step 2, constructing a financial fraud model based on a graph compression technology, including three parts of initialization, optimization track generation and compression network optimization; and step 3, constructing a financial fraud detection model, training by using the compression network obtained in the step 2, and then carrying out fraud detection. According to the method, a complex transaction path, a multi-hop fund flow direction and a gang type fraud behavior can be more effectively captured, and the fraud detection accuracy is remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of graph machine learning, and particularly relates to a financial fraud detection method and device based on graph compression technology and a medium. BACKGROUND

[0002] With the continuous advancement of the digitalization process of the global financial industry, the mode of financial transactions is rapidly evolving from traditional offline channels to online payment, mobile banking, and virtual currency and other emerging forms. This change brings convenience, but also gives rise to more covert and complex financial fraud behaviors. Currently, financial fraud has evolved from early individual cases to organized, coordinated, and even gang-like patterns, often hidden in complex and dynamic financial networks composed of users and their transaction relationships through frequent transactions, intermediary accounts, and other means. In the face of increasingly complex financial network environments, financial fraud detection has become increasingly difficult, and it is urgent to build a detection model that can capture complex fraud behaviors to improve the ability to identify financial fraud behaviors.

[0003] Graph neural networks (GNNs) have become an effective and widely used network data analysis method through the aggregation of neighbor node information and iterative feature propagation mechanisms for graph representation learning. Using GNNs for financial fraud detection can effectively capture abnormal behavior patterns hidden in transaction networks, uncover potential associations between fraud gang members, identify covert transaction paths, and predict potential risks, providing a new technical path for modeling financial fraud.

[0004] However, as the scale of financial transaction networks continues to grow, traditional GNN models face huge memory consumption and training time during training and inference, limiting the scalability of fraud detection methods. Existing research has proposed fraud detection methods based on graph reduction, which reduce financial transaction networks to obtain reduced graphs that retain key graph structure and semantic information, and train GNN models on them to reduce computational cost.

[0005] Existing graph reduction-based financial fraud detection methods mostly use graph coarsening strategies, which only retain loan-free users closely related to fraud behavior and merge other loan-free user nodes to construct a reduced graph for training a GNN model. Since coarsening requires node and edge synthesis, key structural information reflecting coordinated complex transactions is lost, affecting the performance of fraud detection. Therefore, there is an urgent need to design a graph reduction method that can retain key transaction structures, reduce model computational overhead, and still have high sensitivity and accuracy for complex transaction paths, multi-hop fund flows, and gang-style fraud behaviors, thereby achieving efficient and robust financial fraud detection. SUMMARY

[0006] The existing method needs to synthesize nodes and edges during roughening, loses key structural information reflecting collaborative complex transactions, and affects fraud detection performance.

[0007] To achieve the above purpose, the following technical solutions are adopted in the present application:

[0008] A financial fraud detection method based on graph compression technology, the method comprising the following steps:

[0009] Step 1, obtaining a financial transaction network data set The financial transaction platform user is taken as a node, and the financial transaction relationship or dynamic contact relationship between users is taken as an edge to construct a financial transaction network.

[0010] The financial transaction network data set in step 1 is represented as Wherein V is a node set composed of all users, including a loan user node set V f and a non-loan user node set V b , V=V f V b , E is an edge set composed of financial transaction relationships or dynamic contact relationships between users, the users in G have two types of normal users and fraudsters, Y is a user label, and C is a label set.

[0011] Step 2, constructing a compressed network model, performing compression operation on the financial transaction network, the compression operation including initialization, optimization trajectory generation, and compressed network optimization three parts; the compressed network model includes a trajectory optimization model GNN T and a compressed network optimization model GNNs.

[0012] The network compression model in step 2 specifically includes:

[0013] Step 2.1, initialization: including compressed network initialization and compressed model initialization;

[0014] Step 2.2, generating an optimized trajectory based on curriculum learning: training the trajectory optimization model GNN T and generating an optimized trajectory;

[0015] Step 2.3, generating a compressed network based on the optimized trajectory: optimizing the compressed network optimization model GNNs and generating a final compressed network.

[0016] The initialization in step 2.1 is specifically:

[0017] The financial transaction network data set According to a given compression ratio σ, randomly select from the non-loan users in the financial transaction network dataset G. There are nodes, and a set of V nodes with loan users. f The set of nodes V that form the initial compressed network are merged. (0) There will be a set of loan user nodes V f Randomly divided into non-overlapping training node sets V train and test node set V test , and |V train |:|V test | = 7:3; Extract two endpoints from edge set E that both belong to training node set V. train The edges that form the edge set E of the initial compressed network training set. train The corresponding adjacency matrix is ​​A. train Extract two endpoints from edge set E that both belong to node set V. (0) The edges form the edge set E of the initial compressed network. (0) The corresponding adjacency matrix is ​​A. (0) Based on the labels in the financial transaction network dataset G This yields the node set V. (0) The node label Y in (0) Therefore, the compressed network is initialized to G. (0) =(V (0) E (0) Y (0) );

[0018] For trajectory optimization model GNN T The compressed network optimization model GNNs were initialized separately; both models used graph convolutional neural networks (GCNs) with the same number of layers and the same hidden layer size, and the model parameters were randomly initialized using Xavier.

[0019] Step 2.2, optimizing trajectory generation based on course learning: optimizing the trajectory model GNN. T The steps for training and generating an optimized trajectory are as follows:

[0020] The difficulty measure was used to train the node set in the financial transaction network dataset G. Difficulty measurement is performed, and the training scheduler iteratively selects user nodes to optimize the trajectory optimization model GNN. T And use a trajectory generator to save the optimized trajectory, specifically:

[0021] Using a difficulty measure, users trained on the node set in the financial transaction network dataset G were evaluated. Difficulty of recognition Take measurements:

[0022]

[0023] in, This represents the probability that the neighboring nodes of node v belong to class c. The label represents the label of node i, a neighbor of node v. = , express The set of neighbor nodes in the financial transaction network dataset G;

[0024] Use the training scheduler to schedule the training node set V train The nodes are arranged in non-decreasing order of recognition difficulty S(v), and nodes are iteratively selected from them to train the trajectory optimization model GNN. T In total Round iteration, let If a ratio is selected for the nodes in the first iteration, then the... The proportion of nodes selected in each iteration is:

[0025]

[0026] The loss function of the current training Set to:

[0027] ,

[0028] in, Represents the node embedding matrix. V represents the i-th row of Z. t Indicates from the training node set V train Selected by recognition difficulty training nodes, when At that time, V t V train ;when At that time, V t =V train ;

[0029] Use a trajectory generator to extract the trajectory optimization model GNN after the t-th round of training. T parameters And save it to the track list, where and μ is the predefined number of rounds for storing the trajectory. .

[0030] Step 2.3, generating a compressed network based on the optimized trajectory: The specific process of optimizing the compressed network optimization model GNNs and generating the final compressed network is as follows:

[0031] Step 2.3.1: Let t = 0, the total number of iterations be T, and the number of steps for the target trajectory be p;

[0032] Step 2.3.2: Utilize a compression network Perform q rounds of optimization on the compressed network optimization model GNNs;

[0033] In step 2.3.2, a compressed network is used. The GNNs are optimized in q rounds, specifically:

[0034] Based on the current iteration round t, retrieve the saved trajectory from the optimized trajectory list described in step 2.2. Let the parameters of the compressed network optimization model GNNs be adjusted. Using negative log-likelihood loss The update formula for the parameters of the compressed network optimization model GNNs is as follows: (0 ≤ r < q)

[0035]

[0036] in, It is a predefined learning rate;

[0037] Step 2.3.3: Optimize the compressed network model GNNs using the compressed network optimized in Step 2.3.2. Updated to:

[0038]

[0039] Compress the network Updated to This includes updating node features and updating the topology, specifically:

[0040] The updated node features are for the compressed network in the t-th iteration. Characteristics of users without loans The update aims to optimize the following aspects:

[0041]

[0042] in, Indicates the optimization objective. and To balance hyperparameters, This represents the trajectory matching loss. Represents the regularization term of the class distribution, and the trajectory matching loss. The calculation formula is:

[0043]

[0044] in, Represents the L2 norm. Denotes the parameters of the GNNs optimization model in the t-th iteration. Save the number of rounds for a predefined trajectory. This represents the objective parameters of the compressed network optimization model GNNs. Indicates the current iteration number. Indicates the number of steps in the target trajectory;

[0045] Represents the regularization term of the class distribution. The calculation formula is:

[0046]

[0047] in, Let be the average aggregated value of the features of the c-th class of non-loan user nodes in the financial transaction network dataset G. for The average aggregate value of the features of the c-th class of users without loans after the feature update. for The number of users in category c, where N is The number of users.

[0048] The topology update uses the updated node features. renew adjacency matrix Its elements The calculation formula is:

[0049]

[0050] in, It is a multilayer perceptron. This indicates a splicing operation. and Let i and j represent the features of node i and node j after this iteration, respectively.

[0051] Steps 2.3.4, let ;like If so, proceed to step 2.3.2; otherwise, output the compressed network. .

[0052] Step 3: Construct a financial fraud detection model GNN Q The compressed network obtained in step 2 is then used for training, and fraud detection is performed.

[0053] First, the financial fraud detection model GNN Q Initialized with trajectory optimization model GNN TA graph convolutional neural network (GCN) with the same number of layers and the same hidden layer size is used, and the parameters are randomly initialized using Xavier; then, the compressed network obtained in step 2 is... Input financial fraud detection model GNN Q Optimization is performed using negative log-likelihood loss; finally, the compressed network is... Input the optimized financial fraud detection model GNN Q It performs financial fraud detection on users to be tested and outputs the detection results.

[0054] Compared with the prior art, the present invention has the following advantages:

[0055] 1. This invention proposes using graph compression technology in fraud detection to compress the original large-scale financial transaction network into a compressed financial transaction network. This compressed network is then used to train a graph neural network for fraud detection. This significantly reduces computational costs while achieving performance comparable to or even better than the original graph.

[0056] 2. This invention proposes to use a course learning mechanism to train the optimized trajectory in accordance with the difficulty of user node identification from simple to difficult during the training process, so that the obtained optimized trajectory can better identify users with complex relationships, thereby enabling the optimized financial transaction compression network to capture complex fraudulent behavior;

[0057] 3. This invention proposes to use a dynamically updated adjacency matrix. During the optimization process of the compressed graph, after updating the node features, a multilayer perceptron is used to dynamically update the adjacency matrix of the compressed network. This can aggregate loan users to more effective information for detection, and allow the optimized compressed financial transaction network to better capture multi-hop fund flows and gang-related fraudulent activities.

[0058] 4. This invention proposes a joint optimization using trajectory matching and class distribution regularization terms, which can not only capture the trajectory information during the training of the original financial network, but also retain the distribution information of the original financial network. This improves the expressive power and training efficiency of the compressed financial transaction network during graph compression, while maintaining consistency with the structure and parameters of the original graph, thereby avoiding information loss and improving detection accuracy. Attached Figure Description

[0059] Figure 1 A flowchart of a financial fraud detection method based on graph compression technology provided in an embodiment of the present invention.

[0060] Figure 2 This is a diagram illustrating a financial fraud detection framework based on graph compression technology, provided in an embodiment of the present invention.

[0061] Figure 3 This is a schematic diagram of the optimized trajectory training module provided in an embodiment of the present invention.

[0062] Figure 4 This is a schematic diagram illustrating the joint optimization of trajectory matching and class distribution regularization terms provided in an embodiment of the present invention.

[0063] Figure 5 This invention provides a method for training and optimizing a graph compression-based financial fraud detection model. Detailed Implementation

[0064] To gain a deeper understanding of this invention, we will provide a comprehensive and detailed description. However, this invention has various implementations and is not limited to the specific examples listed herein. These examples are presented to enhance a full understanding of the disclosure of this invention.

[0065] For example, this embodiment provides a financial fraud detection scenario. In this scenario, users are modeled as nodes in a graph. User types include users with loans and users without loans. Emergency connections or dynamic connections between users are modeled as edges between users. By using graph compression technology, high-quality nodes for users without loans can be optimized to assist in classifying nodes for users with loans, greatly reducing the size of the transaction network while achieving performance comparable to the original financial network.

[0066] like Figure 1 As shown, this invention provides a method, device, and storage medium for large-scale financial fraud detection based on graph compression technology. This method can effectively capture high-risk users exhibiting complex fraudulent behavior and multi-hop transaction paths in large-scale transaction networks, while significantly reducing model computational overhead. Specifically, it includes the following steps:

[0067] Step 1: Obtain the financial transaction network dataset through the DgraphFin website. The transaction network is represented as follows: Where V is the set of nodes consisting of all users on the website, including the set of loan user nodes V. f and the set of nodes without loans V b The corresponding number of nodes is |V f | = 1225601, |V b | = 2474949, V=V f V b E is an edge set consisting of financial transaction relationships or dynamic connections between users. In G, users are divided into two categories: normal users and fraudsters. Y is a user label, C is a label set, and |C| = 4.

[0068] Step 2, as follows Figure 2The diagram illustrates the construction of a network compression model to perform compression operations on the financial transaction network. The compression operation includes three parts: initialization, optimized trajectory generation, and compressed network optimization. The compressed network model includes a trajectory optimization model, GNN. T And compressed network optimization models GNNs;

[0069] Step 2.1 Initialization, which includes compression network initialization and compression model initialization.

[0070] The initialization of the compressed network, considering that the financial transaction network and the compressed financial transaction network need to maintain a consistent class distribution, requires selecting compressed nodes from the original financial transaction network according to class proportions during the initialization of the compressed financial transaction network, thus obtaining the initial compressed network. Specifically, the financial transaction network... Based on a given compression ratio of 0.15, 371,242 nodes are randomly selected from the non-loan users in the financial transaction network dataset G, and compared with the financial transaction network dataset V. f The set of nodes V that form the initial compressed network are merged. (0) Financial Transaction Network Dataset V f Divided into training node set and test node set V test The ratio is 7:3. Two endpoints are extracted from edge set E that both belong to the training node set. The edges that form the edge set E of the initial compressed network training set. train The corresponding adjacency matrix is ​​A. train Extract two endpoints from edge set E that both belong to V. (0) The edges form the edge set E of the initial compressed network. (0) The corresponding adjacency matrix is ​​A. (0) Based on the labels in the financial transaction network dataset G This yields the node set V. (0) The node label Y in (0) Therefore, the compressed network is initialized to G. (0) =(V (0) E (0) Y (0) ).

[0071] The initialization of the compressed model, considering that trajectory matching optimization of the compressed network is required in the compressed image optimization, necessitates using the same GNN model with the same number of layers and the same hidden layer size in both trajectory optimization training and compressed network optimization. Specifically, this includes initializing the trajectory optimization model GNN. TThe document also describes the initialization of compressed network optimization models (GNNs). Both models consist of three-layer networks with 256 hidden layers, using the ReLU activation function, and outputting 2-dimensional graph convolutional neural networks (GCNs). The model parameters are initialized using Xavier randomization.

[0072] Step 2.2, as follows Figure 3 The diagram shows the use of a difficulty measure on the training node set V in the financial transaction network dataset G. train Difficulty measurement is performed, and the training scheduler iteratively selects user nodes to optimize the trajectory optimization model GNN. T The optimized trajectories are then saved using a trajectory generator. Through training with these optimized trajectories learned in the course, the resulting trajectories better identify users with complex relationships, thus enabling the optimized financial transaction compression network to more effectively capture complex fraudulent activities.

[0073] The difficulty measure was used to train the node set in the financial transaction network dataset G. Difficulty measurement is performed, and the training scheduler iteratively selects user nodes to optimize the trajectory optimization model GNN. T And use a trajectory generator to save the optimized trajectory, specifically:

[0074] Using a difficulty measure, users trained on the node set in the financial transaction network dataset G were evaluated. Difficulty of recognition Take measurements:

[0075]

[0076] in, This represents the probability that the neighboring nodes of node v belong to class c. The label represents the label of node i, a neighbor of node v. = , express The set of neighbor nodes in the financial transaction network dataset G;

[0077] Use the training scheduler to schedule the training node set V train The nodes are arranged in non-decreasing order of recognition difficulty S(v), and nodes are iteratively selected from them to train the trajectory optimization model GNN. T In total Round iteration, let = 0.7 is the node selection ratio in the first iteration, then the... The proportion of nodes selected in each iteration is:

[0078]

[0079] The loss function of the current training Set to:

[0080] ,

[0081] in, Represents the node embedding matrix. V represents the i-th row of Z. t Indicates from the training node set V train Selected by recognition difficulty training nodes, when At that time, V t V train ;when At that time, V t =V train ;in =200, =100.

[0082] Use a trajectory generator to extract the trajectory optimization model GNN after the t-th round of training. T parameters And save it to the track list, where and μ is the predefined number of rounds for storing the trajectory. .

[0083] Step 2.3, as follows Figure 4 The diagram shows the optimization of the compressed network optimization model GNNs to generate the final compressed network. Specific steps include:

[0084] Step 2.3.1: Let t = 0, the total number of iterations T = 300, and the number of steps for the target trajectory p = 10;

[0085] Step 2.3.2: Utilize a compression network Perform q=50 rounds of optimization on the GNNs;

[0086] Step 2.3.3: Use the optimized GNNs to compress the network. Updated to

[0087] ;

[0088] Steps 2.3.4, let ;like If so, proceed to step 2.3.2; otherwise, output... .

[0089] In step 2.3.2, a compressed network is used. The GNNs are optimized in q rounds, specifically:

[0090] Based on the current iteration round t, retrieve the saved trajectory from the optimized trajectory list described in step 2.2. Let the parameters of GNNs Using negative log-likelihood loss The parameter update formula for GNNs in the (t+r+1)th round (0 ≤ r < q) is:

[0091]

[0092] in, This represents a predefined learning rate.

[0093] Step 2.3.3 Compressing the network Updated to This includes updating node features and updating the topology, specifically:

[0094] The updated node features are for the compressed network in the t-th iteration. Characteristics of users without loans The update aims to optimize the following aspects:

[0095]

[0096] in, Indicates the optimization objective. and To balance hyperparameters, This represents the trajectory matching loss. Represents the regularization term of the class distribution, and the trajectory matching loss. The calculation formula is:

[0097]

[0098] in, Represents the L2 norm. Denotes the parameters of the GNNs optimization model in the t-th iteration. Save the number of rounds for a predefined trajectory. This represents the objective parameters of the compressed network optimization model GNNs. Indicates the current iteration number. Indicates the number of steps in the target trajectory;

[0099] Represents the regularization term of the class distribution. The calculation formula is:

[0100]

[0101] in, Let be the average aggregated value of the features of the c-th class of non-loan user nodes in the financial transaction network dataset G. for The average aggregate value of the features of the c-th class of users without loans after the feature update. for The number of users in category c, where N is The number of users.

[0102] The updated topology component, considering that the optimized compressed graph needs to more effectively capture multi-hop fund flows and organized fraudulent activities, requires dynamic updates to the adjacency matrix. Specifically, it utilizes the updated node features. renew adjacency matrix Its elements The calculation formula is:

[0103]

[0104] in, It is a multilayer perceptron. This indicates a splicing operation. and Let i and j represent the features of node i and node j after this iteration, respectively.

[0105] Step 3: GNN Q Initialize to be compatible with GNN T A graph convolutional neural network (GCN) with the same number of layers and the same hidden layer size is used, and the parameters are randomly initialized using Xavier. Next, the compressed network obtained in step 2 is... Input model GNN Q The negative log-likelihood loss was used for model optimization. After the model was optimized, V was... test The system performs financial fraud detection on users and outputs the detection results.

[0106] To verify the effectiveness of this invention, the financial transaction benchmark datasets DgraphFin and elliptic were selected for fraud detection.

[0107] All experiments in this embodiment were implemented using PyTorch and executed on an NVIDIA A800 GPU. Experiments on the DgraphFin and elliptic datasets followed the same experimental setup as the paper on those datasets. The datasets were standardly partitioned and fed into step 2 to build the network compression model. In this embodiment, AUC was used as the evaluation metric. The test AUC of 10-fold cross-validation was averaged, and the epoch with the highest average accuracy was selected. Then, the average AUC and standard deviation were calculated based on the selected epoch, and the final results were reported. Different base GNN models, GCN, GAT, and GraphSAGE, were used for validation on the two datasets. The models were trained using the Adam optimizer with a learning rate of 0.01 and a hidden layer dimension of 128. The model of this invention was compared with seven model reduction algorithms: Random, Cent-P, Cent-D, Graph-Skeleton, GC-VN, GC-AJC, and GCond. Table 1 below shows the performance comparison between the model of this invention and existing models, which presents the fraud detection experimental results obtained on DgraphFin and elliptic. The average and standard deviation of the 10-fold cross-validation AUC values ​​are shown as percentages, and BCR represents the compression ratio.

[0108] Table 1. Experimental results of fraud detection on DgraphFin and ellipti.

[0109]

[0110] Contents not described in detail in this specification are prior art known to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications are readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of this invention are protected.

Claims

1. A financial fraud detection method based on graph compression technology, characterized in that, The method includes the following steps: Step 1: Obtain the financial transaction network dataset A financial transaction network is constructed by using users of the financial transaction platform as nodes and the financial transaction relationships or dynamic connections between users as edges. Step 2: Construct a compressed network model to compress the financial transaction network. The compression operation includes three parts: initialization, optimized trajectory generation, and compressed network optimization. The compressed network model includes a trajectory optimization model (GNN). T And compressed network optimization models GNNs; Step 3: Construct a financial fraud detection model GNN Q The compressed network obtained in step 2 is then used for training to perform fraud detection.

2. The financial fraud detection method based on graph compression technology according to claim 1, characterized in that: The financial transaction network dataset in step 1 is represented as follows: Where V is the set of nodes consisting of all users, including the set of loan user nodes V. f and the set of nodes without loans V b V=V f V b E is an edge set consisting of financial transaction relationships or dynamic connections between users. Users in G are divided into two categories: normal users and fraudsters. Y is a user label, and C is a label set.

3. The financial fraud detection method based on graph compression technology according to claim 2, characterized in that: Step 2, which involves constructing the network compression model, specifically includes: Step 2.1, Initialization: This includes initialization of the compression network and initialization of the compression model; Step 2.2: Optimize trajectory generation based on course learning: Optimize the trajectory optimization model GNN. T Train and generate optimized trajectories; Step 2.3: Generate compressed network based on optimized trajectory: Optimize the compressed network optimization model GNNs and generate the final compressed network.

4. The financial fraud detection method based on graph compression technology according to claim 3, characterized in that: The initialization in step 2.1 specifically involves: For financial transaction network datasets According to a given compression ratio σ, randomly select from the non-loan users in the financial transaction network dataset G. There are nodes, and a set of V nodes with loan users. f The set of nodes V that form the initial compressed network are merged. (0) There will be a set of loan user nodes V f Randomly divided into non-overlapping training node sets V train and test node set V test , and |V train |:|V test | = 7:3; Extract two endpoints from edge set E that both belong to training node set V. train The edges that form the edge set E of the initial compressed network training set. train The corresponding adjacency matrix is ​​A. train Extract two endpoints from edge set E that both belong to node set V. (0) The edges form the edge set E of the initial compressed network. (0) The corresponding adjacency matrix is ​​A. (0) Based on the labels in the financial transaction network dataset G This yields the node set V. (0) The node label Y in (0) Therefore, the compressed network is initialized to G. (0) =(V (0) E (0) , Y (0) ); For trajectory optimization model GNN T The compressed network optimization model GNNs were initialized separately; both models used graph convolutional neural networks with the same number of layers and the same hidden layer size, and the model parameters were randomly initialized using Xavier.

5. The financial fraud detection method based on graph compression technology according to claim 4, characterized in that: Step 2.2, optimizing trajectory generation based on course learning: optimizing the trajectory model GNN. T The steps for training and generating an optimized trajectory are as follows: The difficulty measure was used to train the node set in the financial transaction network dataset G. Difficulty measurement is performed, and the training scheduler iteratively selects user nodes to optimize the trajectory optimization model GNN. T And use a trajectory generator to save the optimized trajectory, specifically: Using a difficulty measure, users trained on the node set in the financial transaction network dataset G were evaluated. Difficulty of recognition Take measurements: in, This represents the probability that the neighboring nodes of node v belong to class c. The label represents the label of node i, a neighbor of node v. = , express The set of neighbor nodes in the financial transaction network dataset G; Use the training scheduler to schedule the training node set V train The nodes are arranged in non-decreasing order of recognition difficulty S(v), and nodes are iteratively selected from them to train the trajectory optimization model GNN. T In total Round iteration, let If a ratio is selected for the nodes in the first iteration, then the... The proportion of nodes selected in each iteration is: The loss function of the current training Set to: ， in, Represents the node embedding matrix. V represents the i-th row of Z. t Indicates from the training node set V train Selected by recognition difficulty training nodes, when At that time, V t V train ;when At that time, V t =V train ; Use a trajectory generator to extract the trajectory optimization model GNN after the t-th round of training. T parameters And save it to the track list, where and μ is the predefined number of rounds for storing the trajectory. .

6. The financial fraud detection method based on graph compression technology according to claim 5, characterized in that: Step 2.3, generating a compressed network based on the optimized trajectory: The specific process of optimizing the compressed network optimization model GNNs and generating the final compressed network is as follows: Step 2.3.1: Let t = 0, the total number of iterations be T, and the number of steps for the target trajectory be p; Step 2.3.2: Utilize a compression network Perform q rounds of optimization on the compressed network optimization model GNNs; Step 2.3.3: Optimize the compressed network model GNNs using the compressed network optimized in Step 2.3.

2. Updated to: ； Steps 2.3.4, let ;like If so, proceed to step 2.3.2; otherwise, output the compressed network. .

7. The financial fraud detection method based on graph compression technology according to claim 6, characterized in that: In step 2.3.2, a compressed network is used. The GNNs are optimized in q rounds, specifically: Based on the current iteration round t, retrieve the saved trajectory from the optimized trajectory list described in step 2.

2. Let the parameters of the compressed network optimization model GNNs be adjusted. Using negative log-likelihood loss The update formula for the parameters of the compressed network optimization model GNNs is as follows: (0 ≤ r < q) in, It is a predefined learning rate; Step 2.3.3 Compressing the network Updated to This includes updating node features and updating the topology, specifically: The updated node features are for the compressed network in the t-th iteration. Characteristics of users without loans The update aims to optimize the following aspects: in, Indicates the optimization objective. and To balance hyperparameters, This represents the trajectory matching loss. Represents the regularization term of the class distribution, and the trajectory matching loss. The calculation formula is: in, Represents the L2 norm. Denotes the parameters of the GNNs optimization model in the t-th iteration. Save the number of rounds for a predefined trajectory. This represents the objective parameters of the compressed network optimization model GNNs. Indicates the current iteration number. Indicates the number of steps in the target trajectory; Represents the regularization term of the class distribution. The calculation formula is: in, Let be the average aggregated value of the features of the c-th class of non-loan user nodes in the financial transaction network dataset G. for The average aggregate value of the features of the c-th class of users without loans after the feature update. for The number of users in category c, where N is The number of users; The topology update uses the updated node features. renew adjacency matrix Its elements The calculation formula is: in, It is a multilayer perceptron. This indicates a splicing operation. and Let i and j represent the features of node i and node j after this iteration, respectively.

8. The financial fraud detection method based on graph compression technology according to claim 7, characterized in that: Step 3: Constructing the financial fraud detection model GNN Q The compressed network obtained in step 2 is then used for training, and fraud detection is performed specifically as follows: First, the financial fraud detection model GNN Q Initialized with trajectory optimization model GNN T A graph convolutional neural network (GCN) with the same number of layers and the same hidden layer size is used, and the parameters are randomly initialized using Xavier; then, the compressed network obtained in step 2 is... Input financial fraud detection model GNN Q Optimization is performed using negative log-likelihood loss; finally, the compressed network is... Input the optimized financial fraud detection model GNN Q It performs financial fraud detection on users to be tested and outputs the detection results.

9. A processing device, characterized in that, include: At least one memory for storing one or more programs; At least one processor is capable of executing one or more programs stored in the memory, such that when the one or more programs are executed by the processor, the processor can implement the financial fraud detection method based on graph compression technology as described in any one of claims 1-8.

10. A readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, enables the implementation of a financial fraud detection method based on graph compression technology as described in any one of claims 1-8.

Images