Energy-Based Models Boost Tabular Anomaly Detection

This paper re-examines the utility of Energy-Based Models (EBMs), specifically the Deep Boltzmann Machine (DBM), for tabular anomaly detection, a field often dominated by density-proxy and reconstruction-based methods. The core hypothesis is that the DBM's mean-field energy provides a distinct and complementary perspective to reconstruction-based scores, such as those derived from Autoencoders. Through extensive experiments on two diverse tabular benchmarks, UCI Bank Marketing and NSL-KDD, the DBM's mean-field energy demonstrated superior performance, matching or statistically outperforming strong baselines like the Autoencoder and significantly beating seven other methods. Crucially, when the DBM energy was combined with an Autoencoder via rank fusion, it yielded statistically significant improvements on both datasets. This suggests that classical EBMs, exemplified by the DBM, offer a non-redundant and valuable addition to the anomaly detection toolkit, especially when integrated with existing reconstruction-based techniques.