LineMVGNN: Anti-Money Laundering with Line-Graph-Assisted Multi-View Graph Neural Networks

Battle Count: 68

The paper is primarily focused on anti-money laundering and fraud detection in transaction graphs, not on quantitative trading strategies, price prediction, or portfolio optimization. However, the graph-based anomaly detection methodology could potentially be adapted for detecting market manipulation patterns or unusual trading activity. The GNN techniques for directed transaction graphs have tangential relevance to order flow analysis in trading.

The method involves multiple components: (1) two-way message passing with parameter sharing in the Dir-GNN framework, (2) line graph construction and edge feature propagation, (3) two variants (add/cat) for message combination, (4) personalized PageRank-based aggregation for final embeddings, (5) residual connections for edge updates, (6) edge sampling for scalability. The refined version (Algorithm 2) avoids explicit line graph construction but still requires scanning neighboring edges. Implementation uses PyTorch, DGL, and PyTorch Geometric. Hyperparameter tuning includes learning rate, embedding size, edge sampling threshold, and model depth.

The Ethereum Phishing Transaction Network dataset is publicly available on Kaggle. Implementation details are provided (PyTorch, DGL, PyTorch Geometric, Adam optimizer, cosine annealing with warm restarts, specific hyperparameter ranges). However, the FPT dataset is not available due to privacy regulations, and no GitHub repository is mentioned. The anomaly injection strategy for FPT follows a cited reference. Key hyperparameters (learning rate, embedding size, edge sampling threshold) are specified with grid search ranges.