Revisiting Energy-based Tabular Anomaly Detection: Energy and Reconstruction are Complementary

Abstract: Tabular anomaly detection is dominated by classical density-proxy methods (Isolation Forest, OCSVM, LOF), reconstruction-based detectors (Autoencoders, VAEs), and modern non-parametric scorers (COPOD, ECOD, Deep SVDD), all of which approximate the inlier distribution only indirectly; explicit energy-based models are largely absent. Motivated by the recent revival of EBMs in deep learning (e.g., Energy-Based Transformers, JEPA), we revisit the classical Deep Boltzmann Machine (DBM) for this task and hypothesize that its mean-field energy combines more effectively with a reconstruction-based score than same-lineage pairs do. We evaluate a two-hidden-layer DBM on two tabular benchmarks spanning distinct domains (UCI Bank Marketing and NSL-KDD) against eight classical and modern baselines across twenty random seeds. The DBM mean-field energy matches the strongest baseline (the Autoencoder) on Bank Marketing and statistically beats it on NSL-KDD, while significantly outperforming the remaining seven on both datasets. When fused with the Autoencoder via rank fusion, the DBM energy yields a statistically significant improvement on both datasets (AUROC=+0.014, p<0.01 on Bank Marketing; +0.002, p<0.001 on NSL-KDD); every non-DBM-derived base model instead fails to improve or significantly degrades the AE-paired ensemble. Our position is that classical EBMs, exemplified by the DBM, deserve a place in the tabular anomaly detection toolbox as a non-redundant complementary view to the reconstruction-based scores that dominate current practice.

• The paper shows that Deep Boltzmann Machine mean-field energy provides a complementary joint-distribution signal to autoencoder reconstruction, matching or surpassing strong baselines across Bank Marketing and NSL-KDD.
• Rank fusion and leakage-free convex fusion of DBM energy with autoencoder scores significantly improve AUROC, reaching 0.745 on Bank Marketing and 0.9953 on NSL-KDD.
• Only DBM-derived scores consistently enhanced autoencoder performance, while entropy-corrected variational free energy often weakened detection, highlighting the importance of latent posterior behavior.

Motivation and position

Tabular anomaly detection is dominated by three method families that all approximate the inlier distribution $p_{\text{in}}(x)$ only indirectly: classical density proxies (Isolation Forest, OCSVM, LOF), reconstruction-based detectors (Autoencoders, VAEs), and modern non-parametric scorers (COPOD, ECOD, Deep SVDD). None of these provides an explicit, differentiable scalar energy over the joint configuration of all input features. This paper, accepted at ICONIP 2026 (2608.14186), revisits the Deep Boltzmann Machine (DBM) as a canonical deep energy-based model (EBM) for this task, motivated by the broader revival of explicit energy formulations in modern deep learning (Energy-Based Transformers, the JEPA family). The central position is deliberately modest but specific: the DBM's mean-field energy is a non-redundant complementary view to reconstruction-based scores, not a replacement for them.

The mechanistic argument is structural. The DBM's mean-field energy $F_{\text{before}}(v) = E_\theta(v, \mu)$ aggregates evidence over the joint hidden-unit configuration via bidirectional mean-field inference, whereas Autoencoder (AE) reconstruction error is computed coordinate-wise on a feed-forward bottleneck. From this, the paper derives a testable hypothesis: a cross-lineage hybrid of the two scores should outperform either component, while same-lineage hybrids should not.

Experimental design

The evaluation covers two tabular benchmarks spanning distinct domains: UCI Bank Marketing (45,211 rows, 16 attributes, 11.7% anomaly rate) and the DoS subset of NSL-KDD (74,826 connections, 41 attributes, 10% anomaly rate). All inputs are binarized (one-hot categoricals, quantile-binned numerics), yielding visible dimensions of $D=69$ and $D=144$. The DBM has two hidden layers ($H_1=48$, $H_2=24$), selected via a single-seed architecture ablation showing that moderate compression outperforms expansion stacks ($[96,128]$: AUROC 0.663 vs. 0.737) and aggressive bottlenecks; notably, configurations whose deepest layer is at or below the source-attribute count terminate fine-tuning at epoch 3 and never recover. Training follows the standard two-phase protocol: layer-wise RBM pretraining via Persistent Contrastive Divergence, then joint fine-tuning on the mean-field energy.

The baseline panel comprises eight established detectors from all three lineages, and all results are reported over 20 random seeds with paired two-sided $t$-tests on per-seed metrics. The NSL-KDD inclusion is explicitly justified as a high-saturation control: if DBM complementarity persists when every method sits near the AUROC ceiling, the gains on the weaker Bank Marketing benchmark cannot be dismissed as artifacts of weak baselines.

Single-method results

The DBM mean-field energy is competitive with, and on one dataset superior to, the strongest baseline:

• Bank Marketing: DBM statistically ties the Autoencoder on all metrics (AUROC $p=0.10$; DBM $0.713 \pm 0.009$ vs. AE $F_{\text{before}}(v) = E_\theta(v, \mu)$0) while significantly outperforming the remaining seven baselines on AUROC ($F_{\text{before}}(v) = E_\theta(v, \mu)$1).
• NSL-KDD: DBM statistically beats the AE on all three metrics: $F_{\text{before}}(v) = E_\theta(v, \mu)$2AUROC $F_{\text{before}}(v) = E_\theta(v, \mu)$3 ($F_{\text{before}}(v) = E_\theta(v, \mu)$4), $F_{\text{before}}(v) = E_\theta(v, \mu)$5PR-AUC $F_{\text{before}}(v) = E_\theta(v, \mu)$6 ($F_{\text{before}}(v) = E_\theta(v, \mu)$7), $F_{\text{before}}(v) = E_\theta(v, \mu)$8F1 $F_{\text{before}}(v) = E_\theta(v, \mu)$9 ($D=69$0), with AUROC 0.995 vs. 0.993.

That an unmodified 2009-era architecture matches or exceeds the AE—and the modern scorers COPOD, ECOD, and Deep SVDD—on both datasets is the paper's first substantive claim. The implication is that explicit joint energy is a viable primary anomaly score for tabular data, not merely a diagnostic.

Hybrid scoring

Two leakage-free fusion protocols are evaluated: parameter-free rank fusion and a convex combination of $D=69$1-standardized scores with leave-one-seed-out (LOSO) tuning of $D=69$2. Both significantly outperform either single method on both datasets:

Score	BM AUROC	NSL-KDD AUROC
AE alone	$D=69$3	$D=69$4
DBM alone	$D=69$5	$D=69$6
Convex (LOSO)	$D=69$7 ($D=69$8 vs. AE, $D=69$9)	$D=144$0 ($D=144$1, $D=144$2)
Rank fusion	$D=144$3 ($D=144$4, $D=144$5)	$D=144$6 ($D=144$7, $D=144$8)

The chosen $D=144$9 is stable across seeds (0.4 in 18/20 BM seeds; 0.6 in 17/19 NSL-KDD seeds), and the convex AUROC sweeps form smooth concave curves peaking strictly above both endpoints—consistent with genuinely complementary signal rather than an averaging artifact. The gains are larger on the low-ceiling BM benchmark, as expected.

Only DBM-derived partners augment the AE

The strongest control in the paper addresses the objection that any diverse pair of scores should beat either alone. Rank-fusing the AE with each of eight candidate partners yields a stark pattern: only the two DBM-derived scores ($H_1=48$0 and the DBM's own reconstruction error) significantly improve the AE-paired ensemble; every non-DBM partner (VAE, LOF, IF, OCSVM, COPOD, ECOD, Deep SVDD) either fails to improve or significantly degrades it, on both datasets. On Bank Marketing, for example, fusing with the VAE costs $H_1=48$1 ($H_1=48$2) and OCSVM $H_1=48$3 ($H_1=48$4), while DBM energy gains $H_1=48$5 ($H_1=48$6).

This is the paper's most consequential result: DBM-derived scores as a family carry information the AE's reconstruction error does not, and the effect is not generic score diversity. Among the two DBM signals, $H_1=48$7 is the cleaner representative of the joint-configuration view—on BM it outperforms DBM-reconstruction as an AE partner by roughly an order of magnitude ($H_1=48$8 vs. $H_1=48$9), though the paired comparison is borderline ($H_2=24$0–0.09), and on NSL-KDD the two converge because the AUROC ceiling collapses any margin.

Mean-field energy vs. proper variational free energy

A subtle design choice receives its own analysis: the paper scores with $H_2=24$1 rather than the proper variational free energy $H_2=24$2. The justification is a systematic class asymmetry in posterior entropy: anomalies drive mean-field posteriors toward less-confident configurations, with $H_2=24$3 nats (positive in 20/20 seeds, $H_2=24$4). Subtracting $H_2=24$5 therefore removes more from anomaly scores than inlier scores, partially cancelling the discriminative signal. Empirically, $H_2=24$6 degrades PR-AUC and F1 significantly ($H_2=24$7) and AUROC borderline ($H_2=24$8) on BM.

The paper scopes this claim carefully. The entropy asymmetry reproduces in a depth-1 RBM ablation and on NSL-KDD ($H_2=24$9 to $[96,128]$0 nats, 20/20 seeds), so the mechanism is generic to latent-variable EBMs scored through variational bounds—but the magnitude of the ranking effect is dataset-dependent, and the authors report observing both signs of the difference in preliminary runs on further benchmarks. The argument also does not extend to latent-free energy networks such as DSEBM, where no entropy term arises.

Computational cost

The cost accounting is candid. DBM training is expensive—176 s on BM and 468 s on NSL-KDD, roughly $[96,128]$1–$[96,128]$2 the AE and three orders of magnitude above the classical proxies—and the GPU does not help at this scale (683 s on the Metal backend, the model being too small to amortize dispatch overhead). Inference, however, runs the other way: scoring requires 10 mean-field sweeps and one energy evaluation with no sampling, making it $[96,128]$3–$[96,128]$4 cheaper than the AE and an order of magnitude cheaper than OCSVM. The practical consequence: where training budget binds, the DBM is the wrong tool; where a detector is fitted once and served, rank-fusing its energy into an AE deployment raises per-sample scoring cost by about a third.

Interpretability: pairwise column interactions

An appendix extends the per-column intervention $[96,128]$5 to pairwise interventions, defining the non-additive component $[96,128]$6, which is identically zero for any column-separable score and thus structurally unavailable to feed-forward reconstruction detectors. Of 91 column pairs on BM, 31 remain significant under Bonferroni correction, with stable signs across seeds; the top pairs recover plausible joint structure (job × education, month × day_of_week). The authors correctly frame this as a diagnostic rather than a competing predictor—per-pair statistics reach at most AUROC 0.71, well below the full score.

Limitations and open questions

The paper is explicit about its boundaries. The empirical study spans only two datasets, so the complementarity claim is evidence of cross-domain generalization rather than a universal one; whether it holds across other ADBench datasets is left open. The baseline panel covers established practice but not the current frontier—tabular-specific contrastive designs (ICL), diffusion-time estimation, and non-parametric-transformer detectors are excluded, leaving untested whether a recent deep tabular detector would also prove a productive AE partner. Architecturally, the Bernoulli visible layer forces quantile binning of continuous attributes, discarding within-bin ordinal information and introducing sensitivity to bin count; a Gaussian or Gaussian–Bernoulli mixture visible layer is proposed but not evaluated. Finally, the DBM-reconstruction vs. $[96,128]$7 comparison on BM is borderline rather than decisive, and the entropy-term effect size is acknowledged to vary by dataset in preliminary runs.

Conclusion

This paper makes a controlled, statistically grounded case that the DBM's mean-field energy is a viable and non-redundant anomaly score for tabular data. Its three findings—parity or superiority over the strongest single baseline across two domains, significant hybrid gains over either component, and the exclusive productivity of DBM-derived partners in AE-paired ensembles—collectively support treating explicit joint energy as a complementary perspective to coordinate-wise reconstruction. The evidence base is narrow (two datasets, eight baselines, a Bernoulli visible layer), and the authors scope their claims accordingly, leaving the breadth of the complementarity effect across the modern tabular AD landscape as the principal open question.