Core Concepts
A novel simplified transformer framework with cross-view attention is proposed to effectively capture the relationship between nodes/graphs and exploit the view co-occurrence for unsupervised graph-level anomaly detection.
Abstract
The paper proposes a novel method called Simplified Transformer with Cross-View Attention for Unsupervised Graph-level Anomaly Detection (CVTGAD). The key highlights are:
Graph Pre-processing Module:
Generates feature view and structure view of each graph using perturbation-free graph augmentation.
Calculates preliminary node/graph embeddings using GNN encoders (GIN and GCN).
Simplified Transformer-based Embedding Module:
Designs a simplified transformer structure with projection network, residual network, and transformer to capture the relationship between nodes/graphs from both intra-graph and inter-graph perspectives.
Introduces a cross-view attention mechanism to directly exploit the view co-occurrence between feature view and structure view, bridging the inter-view gap at node level and graph level.
Adaptive Anomaly Scoring Module:
Employs an adaptive strategy considering both node-level and graph-level cross-view contrastive losses to calculate the final anomaly score.
The proposed CVTGAD method is evaluated on 15 real-world datasets across different fields, demonstrating its superiority over 9 competitive baselines in unsupervised graph-level anomaly detection.
Stats
The average number of nodes in the datasets is 39.06, 32.63, 15.69, 42.43, 35.75, 41.22, 284.32, 29.87, 19.77, 429.63, 74.49, 16.89, 17.62, 17.92, and 17.38.
The average number of edges in the datasets is 72.82, 62.14, 16.20, 44.54, 38.36, 43.45, 715.66, 32.30, 96.53, 497.75, 2457.78, 17.23, 17.98, 18.34, and 17.72.
Quotes
"To increase the receptive field, we construct a simplified transformer-based module, exploiting the relationship between nodes/graphs from both intra-graph and inter-graph perspectives."
"We design a cross-view attention mechanism to directly exploit the view co-occurrence between different views, bridging the inter-view gap at node level and graph level."