Disentangling Tabular Data Towards Better One-Class Anomaly Detection

TL;DR

Disent-AD tackles tabular one-class anomaly detection by learning intrinsic correlations among normal attributes through two non-overlapping CorrSets discovered with a two-head self-attention module. It jointly optimizes a disentangling loss $L_d$ and a reconstruction loss $L_r$ to restore data from each CorrSet, producing an anomaly score $\phi(x) = \sum_{h=1}^2 (x - \hat{x}^{s_h})^2$. Evaluated on 20 tabular datasets, it achieves an average improvement of $6.1\%$ in AUC-PR and $2.1\%$ in AUC-ROC over prior methods, with strong ablation results showing the necessity of two-subset disentanglement and the proposed losses. The approach uses patch-splitting preprocessing to aid disentanglement and demonstrates robustness to anomaly contamination, marking a significant advance in robust, unsupervised tabular anomaly detection. This work pioneers applying disentanglement concepts to tabular one-class anomaly detection and points to future extensions to non-tabular data domains.

Abstract

Tabular anomaly detection under the one-class classification setting poses a significant challenge, as it involves accurately conceptualizing "normal" derived exclusively from a single category to discern anomalies from normal data variations. Capturing the intrinsic correlation among attributes within normal samples presents one promising method for learning the concept. To do so, the most recent effort relies on a learnable mask strategy with a reconstruction task. However, this wisdom may suffer from the risk of producing uniform masks, i.e., essentially nothing is masked, leading to less effective correlation learning. To address this issue, we presume that attributes related to others in normal samples can be divided into two non-overlapping and correlated subsets, defined as CorrSets, to capture the intrinsic correlation effectively. Accordingly, we introduce an innovative method that disentangles CorrSets from normal tabular data. To our knowledge, this is a pioneering effort to apply the concept of disentanglement for one-class anomaly detection on tabular data. Extensive experiments on 20 tabular datasets show that our method substantially outperforms the state-of-the-art methods and leads to an average performance improvement of 6.1% on AUC-PR and 2.1% on AUC-ROC. Codes are available at https://github.com/yjnanan/Disent-AD.

Figures (8)

• Figure 1: Visualization of MCM yin2024mcm's fifteen soft masks and our two attention maps on the Thyroid dataset (six attributes). The presented masks and attention maps are averages derived from all training samples. In MCM's masks, darker colors indicate greater masking ratios while a higher attention weight in our attention maps.
• Figure 2: Illustration of our strategy. Red regions indicate anomaly attributes.
• Figure 3: Illustration of our framework. The model consists of three modules: an encoder, a two-head self-attention, and a decoder. The network is trained under the constraint of a disentangling loss and a reconstruction loss.
• Figure 4: Averaged reconstruction error and disentangling loss of test samples during training on the Thyroid dataset.
• Figure 5: AUC-PR and AUC-ROC vs. subset numbers.