Abstract: Aiming at  the problems that traditional graph neural networks were difficult to handle large-scale complex data and it was difficult to 
 fully utilize the features of nodes and edges in the network as well as the connections between them, we proposed a feature-enhanced graph attention network (FEGAT). The method first modelled the data as a directed graph of nodes and edges. Then, a graph neural network model with an attention mechanism was designed to dynamically aggregate neighbor node information through edge feature weighting and attention coefficients, and fused multi-dimensional edge features by adopting an independent weighting and parallel concatenation strategy for multi-dimensional edge features. Finally, it combined XGBoost algorithm to complete node classification. To verify the effectiveness of the model, experiments were carried out on the anti-money laundering AMLSim dataset. The results show that the model significantly improves precision, recall and F₁-score in imbalanced data scenarios.

Key words: graph neural network, attention mechanism, graph representation learning, money laundering detection