Can Tabular Anomaly Detection Move Beyond One‑for‑One? OFA‑TAD Introduces a One‑for‑All Paradigm

Problem

Tabular anomaly detection (TAD) aims to identify rare samples that deviate from normal distributions in structured data. Existing methods follow a one‑for‑one (OFO) paradigm: each new dataset requires training a dedicated detector, incurring high computational cost and limited generalization.

One‑for‑All Objective

Goal is to train a single model on multiple source tables and apply it directly to unseen target tables without any domain‑specific fine‑tuning.

OFA‑TAD Model

Paper “Towards One‑for‑All Anomaly Detection for Tabular Data” (Li et al., 2026, arXiv:2603.14407, code: https://github.com/Shiy-Li/OFA-TAD) introduces OFA‑TAD. The model represents each sample by the distances to its top‑K nearest neighbors (“neighbor‑distance portrait”). Anomalies appear more isolated, showing jumps or long‑tail patterns in the distance sequence.

Multi‑View Distance Encoding

Four feature transformations (Raw, Standardized, MinMax, Quantile) generate distinct metric spaces.

For each view, the top‑K neighbor distance sequence is extracted and normalized via quantile scaling, yielding a unified distance language across domains.

Mixture‑of‑Experts Adaptive Fusion

View experts : each expert models one distance view using positional encoding and attention pooling to produce an anomaly score.

Gating network : predicts dynamic weights for experts based on the sample’s representations.

Weighted fusion : combines expert scores, emphasizing reliable views and suppressing noisy ones.

Synthesizing Pseudo‑Anomalies

To provide supervision in the one‑class setting, OFA‑TAD generates synthetic anomalies via four strategies:

Manifold extrapolation – global out‑of‑distribution points.

Cluster interpolation – local low‑density anomalies.

Noise injection – measurement errors.

Feature masking – missing or corrupted features.

Experimental Evaluation

Trained on 7 source datasets, evaluated on 34 target datasets spanning 14 domains. Baselines: Isolation Forest, LOF, KNN, AutoEncoder, DeepSVDD, LUNAR, MCM, DRL, DisentAD. OFA‑TAD used identical hyper‑parameters for all targets.

Overall performance : highest average rank on AUROC and AUPRC across the 34 datasets.

Ablation study : removing multi‑view encoding, MoE experts, attention pooling, or positional encoding each reduced performance; attention pooling had the largest impact.

Pseudo‑anomaly strategies : omitting any synthesis strategy lowered accuracy, confirming the benefit of diverse anomaly signals.

Context robustness : model remained effective with a small fraction of normal samples as context; performance improved and saturated as more context was added.

Dataset‑specific scaling : increasing the number of source datasets consistently improved transfer performance.

Conclusion

OFA‑TAD demonstrates a viable one‑for‑all TAD framework, achieving strong cross‑domain results without target‑domain fine‑tuning. Future directions include larger pre‑training corpora, advanced training objectives, and deeper contextual utilization.