Illicit transaction detection is a key task in blockchain anti-money laundering, yet it remains challenging due to temporal graph evolution, severe class imbalance, and the difficulty of preserving informative transaction attributes during representation learning. Existing graph-based approaches have demonstrated the importance of temporal modeling on Bitcoin transaction networks, but their performance may still degrade when structural evidence becomes unstable or when minority illicit nodes are weakly supported by local graph context. To address these issues, we propose FG-EGCN, a feature-gated temporal graph model for illicit Bitcoin transaction detection. The proposed method combines temporal graph encoding with a residual feature branch and an adaptive gating mechanism, enabling the model to balance temporal-relational evidence and transaction-level attribute information in a node-aware manner. In addition, focal-loss-based optimization is adopted to improve sensitivity to rare illicit samples under highly imbalanced label distributions. We evaluate FG-EGCN on the Elliptic Bitcoin transaction dataset under a chronological train-test split. Experimental results show that FG-EGCN achieves strong performance among the compared graph neural models and delivers more robust illicit transaction detection under temporal distribution shift. Timestep-level analysis further demonstrates improved temporal stability, while ablation studies confirm the contribution of temporal evolution, feature preservation, adaptive gating, and imbalance-aware optimization. Moreover, representation visualization and gate analysis show that the proposed model produces a more structured and interpretable embedding space, in which illicit nodes are more compact and better separated from the dominant node population. These results indicate that combining temporal graph learning with feature preservation and adaptive fusion provides an effective and robust solution for illicit Bitcoin transaction detection.
Han et al. (Sun,) studied this question.