Financial Transaction Network Risk Prediction Model Based On Graph Neural Network

To address the risks in financial systems arising from the continuous development of financial markets and the deepening of globalization. First, based on the risk indicator system of the trading market, subgraphs representing the fundamental elements of multidimensional information and risk indicators are determined. For the integrated complex network graph data, an improved node feature embedding method is used for classification, with cosine similarity measuring node attribute similarity. This allows for the extraction of node structural features, which are then concatenated with attribute features. Finally, based on the aggregated features, normalization and a softmax classifier are employed to generate a probability distribution fitting the objective function, improving the overall accuracy of the weighted sampling GraphSAGE financial transaction network risk prediction model. Experimental results show that the loss value of the graph neural network’s cross-entropy loss function is in the low range of 0.15–0.35, and the AUC values for the risks involved in the trading center and the risk indicators faced by traders are 0.8529 and 0.8845, respectively. The classification accuracy of the graph neural network for risk transaction overflow is as high as 99.93%. This validates the effectiveness and feasibility of the proposed financial transaction network risk prediction model, providing new technical means for risk management and development in financial transaction networks.