Attention-Guided Masking and Neighbor-Informed Reconstruction for Tabular Anomaly Detection

ICLR 2026 Conference Submission14851 Authors

Keywords: Tabular anomaly detection, One-class classification

TL;DR: We introduce a one-class anomaly detection approach that combines attention-guided masking with neighbor-informed reconstruction.

Abstract: One-class classification for tabular anomaly detection remains challenging due to the scarcity of labeled anomalies and the absence of explicit structural relationships among samples. Existing approaches have largely focused on intra-instance modeling, such as masking or reconstruction, while inter-sample modeling has received comparatively little attention. We propose AGNI (Attention-Guided Masking and Neighbor-Informed Reconstruction), a self-supervised framework that reimagines attention as a dual supervisory signal unifying these two perspectives. Specifically, Attention-Guided Masking leverages attention to identify and hide salient features, enforcing the learning of fine-grained intra-instance dependencies. At the same time, Neighbor-Informed Reconstruction repurposes the same attention scores to retrieve structurally similar neighbors, whose representations provide contextual support during reconstruction. By tightly coupling intra-instance and inter-sample objectives within a single attention space, AGNI transforms attention from a representational tool into a coordinating structural signal. Extensive experiments on 47 real-world datasets from ADBench demonstrate that AGNI achieves the best overall ranking among 15 classical and deep-learning baseline. Code is available in the supplementary material.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 14851