RTTAD: Risk-Aware Test-Time Adaptation for Unsupervised Tabular Anomaly Detection
Researchers have introduced RTTAD, a novel method for unsupervised tabular anomaly detection that addresses the challenge of normality shifts in practical scenarios. Traditional methods often fail due to limited training data diversity and the risks associated with indiscriminate test-time adaptation, which can lead to anomaly contamination. RTTAD employs a synergistic two-stage mechanism to holistically tackle these issues. During the training phase, it utilizes collaborative dual-task learning to capture multi-level representations and establish a robust normal prior. In the testing phase, a Test-Time Contrastive Learning (TTCL) module manages adaptation risk by selectively updating the model using high-confidence pseudo-normal samples while constraining anomalous ones. Additionally, TTCL incorporates a k-nearest neighbor-based contrastive objective to refine embedding distributions, enhancing the model's discriminative capacity. Extensive experiments conducted on 15 tabular datasets demonstrate that RTTAD achieves state-of-the-art overall detection performance, offering a significant improvement over existing methods by effectively balancing training-phase learning with test-time optimization.
Wire timeline
RTTAD: Risk-Aware Test-Time Adaptation for Unsupervised Tabular Anomaly Detection
Researchers have introduced RTTAD, a novel method for unsupervised tabular anomaly detection that addresses the challenge of normality shifts in practical scenarios. Traditional methods often fail due to limited training data diversity and the risks associated with indiscriminate test-time adaptation, which can lead to anomaly contamination. RTTAD employs a synergistic two-stage mechanism to holistically tackle these issues. During the training phase, it utilizes collaborative dual-task learning to capture multi-level representations and establish a robust normal prior. In the testing phase, a Test-Time Contrastive Learning (TTCL) module manages adaptation risk by selectively updating the model using high-confidence pseudo-normal samples while constraining anomalous ones. Additionally, TTCL incorporates a k-nearest neighbor-based contrastive objective to refine embedding distributions, enhancing the model's discriminative capacity. Extensive experiments conducted on 15 tabular datasets demonstrate that RTTAD achieves state-of-the-art overall detection performance, offering a significant improvement over existing methods by effectively balancing training-phase learning with test-time optimization.
cs.AI updates on arXiv.org