DGT Framework: Effectively Captures Patterns in Evolving Graphs

To fully exploit the rich temporal and structural information in dynamic graphs and improve anomaly detection accuracy, we propose a novel approach that jointly models the temporal evolution of nodes and edges. This is implemented through the Dynamic Graph Transformer (DGT) framework.

📂 Datasets and Experiments

We conduct experiments on two dynamic graph datasets:

• DGraph-Fin: A dynamic financial social network dataset.
  🔗 DGraph-Fin Dataset
• Elliptic: A dataset for cryptocurrency transaction analysis.
  🔗 Elliptic Dataset

For each dataset, we compare Baseline models with our proposed DGT model.

---

🏋️‍♂️ Training Commands

DGraph-Fin Experiments

# Baseline: MLP
python train_fin_baseline.py --model mlp --epochs 200 --device 0

# Baseline: GCN
python train_fin_baseline.py --model gcn --epochs 200 --device 0

# Baseline: GraphSAGE
python train_fin_baseline.py --model sage --epochs 200 --device 0

# DGT
python train_fin_dgt.py

Elliptic Experiments

# Baseline: MLP
python train_elliptic_mlp.py --epoch 10 --device cuda:1

# Baseline: GCN
python train_elliptic_baseline.py --model gcn --epoch 10 --device cuda:1

# Baseline: GraphSAGE
python train_elliptic_baseline.py --model sage --epoch 10 --device cuda:1

# DGT
python train_elliptic_dgt.py --epoch 10 --device cuda:1