Why the Counterintuitive Phenomenon of Likelihood Rarely Appears in Tabular Anomaly Detection with Deep Generative Models?

2026-02-10

AI Summary

This study investigates why the counterintuitive phenomenon—where anomalous samples receive higher likelihood scores than inliers—is rarely observed when applying deep generative models to tabular anomaly detection. By formalizing a general definition of this phenomenon, we systematically evaluate its occurrence across 47 tabular datasets and 10 computer vision/natural language processing embedding datasets, complemented by theoretical and empirical analyses linking its prevalence to data dimensionality and feature correlations. Leveraging invertible flow-based models with exact likelihoods, large-scale experiments on the ADBench benchmark compare against 13 baseline methods and demonstrate that this phenomenon is indeed uncommon in tabular data, thereby validating the reliability and practicality of using likelihood scores alone for anomaly detection in the tabular domain.

Technology Category

• Data Mining & Knowledge Management: Anomaly/Outlier Detection
• Machine Learning: Deep Generative Models & Autoencoders
• Reasoning under Uncertainty: Relational Probabilistic Models

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

Deep generative models with tractable and analytically computable likelihoods, exemplified by normalizing flows, offer an effective basis for anomaly detection through likelihood-based scoring. We demonstrate that, unlike in the image domain where deep generative models frequently assign higher likelihoods to anomalous data, such counterintuitive behavior occurs far less often in tabular settings. We first introduce a domain-agnostic formulation that enables consistent detection and evaluation of the counterintuitive phenomenon, addressing the absence of precise definition. Through extensive experiments on 47 tabular datasets and 10 CV/NLP embedding datasets in ADBench, benchmarked against 13 baseline models, we demonstrate that the phenomenon, as defined, is consistently rare in general tabular data. We further investigate this phenomenon from both theoretical and empirical perspectives, focusing on the roles of data dimensionality and difference in feature correlation. Our results suggest that likelihood-only detection with normalizing flows offers a practical and reliable approach for anomaly detection in tabular domains.

Problem

• tabular anomaly detection
• deep generative models
• likelihood
• counterintuitive phenomenon
• normalizing flows

Innovation

• likelihood-based anomaly detection
• normalizing flows
• counterintuitive phenomenon
• tabular data
• feature correlation