LLM-Augmented GNN for Fraud Detection

Combining a fine-tuned SLM with a Skip-GNN for graph fraud detection.

Recreated and extended the FLAG framework to resolve graph neighborhood camouflage in text-attributed anomaly detection.

  • Alternating PEFT/LoRA fine-tuning of an SLM with a composite loss (BCE, KL-divergence, orthogonality), resolving context limits with semantic k-hop neighbor sampling.
  • Dual-branch Skip-GNN with multi-head attention fusion on LLM-derived embeddings, outperforming GCN/GAT baselines by 5% AUC-ROC/F1 on imbalanced graph benchmarks.

Tech Stack

  • Python, PyTorch, PyTorch Geometric, Hugging Face, PEFT