Title: Reevaluating Weak Supervision for Anomaly Detection: A Holistic Benchmarking Study

Abstract:

Weakly supervised anomaly detection (WSAD) has evolved along three distinct paths: incomplete, inexact, and inaccurate supervision. Despite this proliferation, these trajectories have largely remained siloed, with no unified framework available to determine whether they tackle distinct problems or rely on shared underlying mechanisms. To bridge this gap, we present WSADBench, the inaugural benchmark designed to harmonize evaluations across diverse weakly supervised contexts. This platform assesses a wide spectrum of techniques, ranging from specialized WSAD algorithms to state-of-the-art tabular foundation models.

WSADBench implements standardized protocols to rigorously test 36 different algorithms across four data modalities. By systematically manipulating the quantity, granularity, and quality of labels, the benchmark maps the performance limits of various methodologies. Drawing on insights from more than 700,000 experiments, we identify four pivotal findings:

(i) There are strong intrinsic correlations among these weak supervision scenarios, which contradicts the assumption that current research directions are entirely isolated.

(ii) While specialized WSAD algorithms perform exceptionally well in extreme label-scarcity conditions, they are rapidly outperformed by tabular foundation models and general classification methods as supervision levels rise or when encountering out-of-distribution (OOD) scenarios.

(iii) The value of unlabeled data is inconsistent across different settings, often yielding only marginal improvements when compared to the benefits of refining existing labels.

(iv) Models demonstrate an asymmetric sensitivity to various forms of label noise.

To support ongoing research in this field, we are releasing WSADBench as an open-source benchmark, complete with associated code and datasets. The resource is available at: https://github.com/SUFE-AILAB/WSADBench.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC