Research On Fraud Detection Method Based On Knowledge-guided Graph Neural Network

With the deepening of the digitization of the financial industry,financial frauds are constantly evolving into new forms.The new type of financial fraud criminal activities is more difficult to supervise,which will seriously damage the safety of people’s property,reduce the credibility of financial platforms,impact the national financial credit system,and cause immeasurable economic losses to the country.Therefore,the effective detection of potential financial fraud is of great significance to ensuring the financial security of the country.At present,rules-based financial fraud detection requires a lot of manpower and domain expertise to establish a complete and reliable fraud detection system;and the fraud detection method based on machine learning is difficult to explain due to the lack of sufficient labeled samples,and it is difficult to effectively apply and promote to the practice of the financial fraud detection industry.Inspired by traditional fraud detection methods and machine learning fraud detection methods,this paper combines machine learning methods with domain expertise in traditional fraud detection methods to design a Knowledge-Guided Weakly Supervised Graph Neural Network,namely Weak-GNN.The introduction of machine learning methods has reduced the need for manpower;the guidance of expert knowledge has also solved the problem of sparse labeled samples,while also providing interpretability of detection results.In this paper,under the framework of Weak-GNN,we conduct an indepth study on the specific problems in the knowledge-guided fraud detection model,and the main research content is as follows:(1)Research of the implementation of the knowledge-guided weakly supervised graph neural network,and the verification method based on co-teaching modelThis article basically accomplishes the Weak-GNN.The technical implementation of the model is: modeling financial transactions into a graph structure;for the majority of unlabeled transaction samples in the graph,pseudo-labeling these samples as noise samples using formal expert knowledge,so that the Weak-GNN obtains enhanced input data;pseudo-labels marked by expert knowledge and supervised information are fully combined into a framework,and only a small number of labeled samples and expert prior knowledge accumulated in the corresponding industry are required to obtain satisfactory results.For two types of samples,namely the noise samples brought by knowledge guidance and the original labeled clean samples,in order to combine their training together,this paper draws on the noise training neural network model method,introduces the coteaching training method.Co-teaching mixes the noise and the labeled samples,and inputs them to the Co-teaching based Knowledge-Guided Weakly Supervised Graph Neural Network,called Weak-Co.During training,Weak-Co utilizes the noise-containing training method co-teaching to reduce the noise effects caused by expert knowledge guidance.In summary,this paper solves the training problem of knowledge-guided fraud detection model through Weak-Co model,and preliminarily verifies the feasibility of Weak-GNN.(2)Research of the implementation of a fraud detection model based on a jointlytraining methodSince the co-teaching method is designed for noisy data sets,noise and clean labeled samples are mixed and trained using co-teaching in Weak-Co.In order to realize the joint training of noise labels and clean labels brought about by knowledge guidance in WeakGNN,after improving the joint teaching method,the Jointly-training method is designed to train noise and clean labels jointly to improve the effectiveness of the Weak-GNN model.Therefore,this paper designs a Jointly-training based Knowledge-Guided Weakly Supervised Graph Neural Network,called Weak-Jo.Weak-Jo divides noise samples and labeled samples into two groups and trains them in two graph neural networks with independent parameters.In each epoch of training,the two networks update the parameters with part of small loss samples of themselves,as well as their peer networks,thus avoiding the problem of overfitting of a single network caused by the training of only noisy samples or labeled samples.In addition,Weak-Jo introduced a noise preprocessing method based on constant moving average,which dynamically screens out noisy samples during each training round,thereby improving the accuracy of the fraud detection model.In summary,this paper implements the jointly training of knowledge-guided noise samples and clean labeled samples through the Weak-Jo model.(3)Research of the implementation of an interpretable fraud detection model based on graph attention networksFraud detection models need to serve financial industry experts.In order to improve the value of the application and promotion of the Weak-GNN fraud detection model in the financial industry,this paper explores its interpretable ability.This paper combines expert knowledge with attention mechanism,and designs an interpretable fraud detection model to provide certain interpretability for fraud detection results.Therefore,based on the Weak-GNN,this paper designs an Explainable Knowledge-Guided Weakly Supervised Graph Neural Network,called Weak-Know.Weak-Know extends the reliability level in Weak-GNN to a reliability vector,and aggregates neighbor nodes by introducing a graph attention network.Finally,WeakKnow is provided with interpretable fraud detection results through reliability weights and confidence vectors.In summary,through the exploration of the interpretable model Weak-Know,this paper verifies that the Weak-GNN model has practical application and promotion value in the field of fraud detection,which is conducive to the promotion and application of GNN based machine learning methods in the financial industry.In summary,the main research results of this paper are reflected in the following four aspects:(1)This paper modeled financial transactions into graph structures,introduced industry expert knowledge,and designed a Knowledge-Guided Weakly Supervised Graph Neural Network.(2)Aiming at the problem of how expert knowledge is used for guidance in the WeakGNN model,the co-teaching training method is introduced,and the Co-teaching based Knowledge-Guided Weakly Supervised Graph Neural Network is designed.(3)Aiming at the joint training problem of noise label and clean label brought by knowledge guidance in the Weak-GNN model,the jointly-training method is designed,and the Jointly-training based Knowledge-Guided Weakly Supervised Graph Neural Network is designed.(4)Aiming at the practical application and promotion of the Weak-GNN fraud detection model,the Explainable Knowledge-Guided Weakly Supervised Graph Neural Network model is designed in combination with the attention mechanism in the WeakGNN fraud detection model.The research work of this paper can well meet the needs of financial fraud detection in the era of digital economy,and has great significance and wide application prospects for the prevention of financial transaction fraud.