A Spatio-temporal Prediction Method based on GCNBiLSTM for Digital Financial Fraud Risk

Financial fraud risk prediction faces significant challenges in modeling complex spatiotemporal correlations, as conventional methods struggle to simultaneously capture topological structures of transaction networks and dynamic temporal patterns. To address this, we propose a novel spatiotemporal fusion model integrating Graph Convolutional Networks (GCN) and Bidirectional Long Short-Term Memory (BiLSTM). The model first employs GCN layers to learn spatial dependencies among transaction orders, encoding network topology, timestamps, and features as graph signals. Subsequently, BiLSTM processes the GCN outputs to capture long-term temporal patterns of fraudulent behaviors. This architecture preserves local structural information while modeling global fraud patterns across time steps. Experimental results demonstrate that our co-optimized framework for spatial feature extraction and temporal dependency learning significantly improves the accuracy of financial fraud risk prediction.