Layer 1: Graph Anomaly Detection

PyOD has 8 graph anomaly detectors built on PyTorch Geometric (PyG). Install with pip install pyod[graph]. All 8 are transductive in v1 – use decision_scores_ after fit().

from pyod.models.pyg_dominant import DOMINANT
import torch
from torch_geometric.data import Data

data = Data(x=torch.FloatTensor(X), edge_index=torch.LongTensor(edge_index))
clf = DOMINANT(hidden_dim=64, epochs=100)
clf.fit(data)
scores = clf.decision_scores_                 # per-node anomaly scores

---

Detectors

Rankings from BOND benchmark (NeurIPS 2022):

Type	Detector	Year	Venue
GCN Autoencoder	DOMINANT: GCN AE, structure + attribute	2019	SDM
Contrastive	CoLA: local neighbor context	2022	WWW
Contrastive+AE	CONAD: anomalous-view injection	2022	PAKDD
Attention AE	AnomalyDAE: GAT + MLP dual AE	2020	CIKM
Motif AE	GUIDE: triangle-motif dual AE	2021	BigData
Matrix Factor.	Radar: residual analysis	2017	IJCAI
Matrix Factor.	ANOMALOUS: CUR-style, Laplacian	2018	IJCAI
Structural	SCAN: structural clustering	2007	KDD

---

Input Format

Primary: PyG Data object with x (node features) and edge_index (COO edges).

Convenience: numpy arrays for direct fit() calls:

clf.fit(X, edge_index=edge_index)
# X: numpy (n_nodes, n_features)
# edge_index: numpy (2, n_edges)

SCAN is the only structure-only detector – it works without node features. All others require data.x.