PMP: Partitioning Message Passing for Graph Fraud Detection
Metadane
- Autorzy: Wentao Zhuo, Zihao Liu, Bryan Hooi, Bingsheng He, Guang Tan, Rizal Fathony, Jia Chen
- Rok: 2024
- Źródło: The Twelfth International Conference on Learning Representations (ICLR)
- Status: to-read
- Pochodzenie: Wyekstrahowane z duan-graph-fraud-gaap-2025
- Tagi: to-read reference fraud-detection gnn message-passing partitioning specialized-gnn
Kluczowe Wnioski
- Best specialized GNN w GAAP benchmark: 58.04% average Rec@K
- Partitioning-based message passing strategy dla fraud graphs
- Spatial perspective approach (nie spectral)
- Domain-adapted information weights dla różnych typów węzłów
Notatki
Publikacja dodana automatycznie z bibliografii GAAP. Jest to najlepsza specialized GNN method w GAAP benchmark (58.04% avg Rec@K), przewyższająca GHRN (57.93%) i BWGNN (57.55%). PMP wprowadza partitioning strategy do message passing, co sugeruje focus na handling heterogeneity w fraud graphs. Dodaj PDF aby wygenerować pełne podsumowanie używając /summarize-paper zhuo-pmp-partitioning-message-2024
Context z GAAP: “H2FDetector, PMP, and DGA-GNN adapt different information weights for nodes with different properties from the domain. From the spatial perspective, methods such as CARE-GNN, Rio-GNN, PC-GNN, DiG-In-GNN, and AO-GNN select neighbors based on similarity or customized loss functions.”