Rethinking Weak Supervision in Anomaly Detection: A Comprehensive Benchmark

Researchers launched WSADBench, an open-source benchmark for weakly supervised anomaly detection.

The benchmark tests 36 algorithms across four modalities while varying label quantity, granularity, and quality.

Across more than 700,000 experiments, they found limited gains from unlabeled data and strong sensitivity to label noise.

Specialized WSAD methods were competitive mainly in extreme label-scarcity settings, while broader models often performed better as supervision increased.