Graph Fraud Detection on Amazon (AUC, Recall, G-Mean)
0.9805AUCLGSPF
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| LGSPFOptimization=End-to-end2026.05 | 0.9805 | 91.79 | 93.1 | |
| BWGNN2026.05 | 0.9717 | 78.72 | 88.4 | |
| w/o Joint-OptDescription=LGSPF ablation without end-to-end optimization (frozen GNNs)2026.05 | 0.9703 | 92.71 | 92.33 | |
| PC-GNN2026.05 | 0.9665 | 90.18 | 90.62 | |
| w/o SemanticsDescription=LGSPF ablation removing natural language instructions2026.05 | 0.964 | 88.63 | 91.97 | |
| RGTAN2026.05 | 0.9638 | 88.98 | 88.36 | |
| GTAN2026.05 | 0.9585 | 75.61 | 86.8 | |
| CARE-GNN2026.05 | 0.92 | 86.35 | 88.24 | |
| w/o LLMDescription=LGSPF ablation replacing LLM backbone with MLP classifier2026.05 | 0.9087 | 78.12 | 88.12 | |
| GraphSAGE2026.05 | 0.8735 | 78.42 | 82.11 | |
| Instruct-LLMMode=Instruction-tuned2026.05 | 0.839 | 90.77 | 76.65 | |
| GCN2026.05 | 0.8362 | 77.51 | 77.23 | |
| GAT2026.05 | 0.7954 | 72.34 | 72.96 | |
| Zero-shot LLMMode=Zero-shot2026.05 | 0.5062 | 3.21 | 14.53 |