ERIS: Energy-Guided Shift-Robust Time Series

• The paper introduces ERIS, which guides the disentanglement of domain-specific and label-relevant features using an energy-based semantic signal.
• It employs a joint objective that integrates domain-specific energy, label-specific energy, weight-level orthogonality, and adversarial regularization to enhance robustness under distribution shifts.
• Empirical evaluations on multi-domain sensor benchmarks demonstrate that ERIS outperforms competing methods in accuracy and calibration, validating its shift-robust design.

Energy-Regularized Information for Shift-Robustness (ERIS) is an end-to-end framework for out-of-distribution (OOD), or shift-robust, time series classification that is designed to learn invariant representations by guiding feature disentanglement with an energy-based semantic signal rather than relying on unguided separation constraints alone. In the formulation studied by the paper, training and test distributions differ, $$\mathbb{P}(\mathcal{X}^{tr}, \mathcal{Y}^{tr}) \neq \mathbb{P}(\mathcal{X}^{te}, \mathcal{Y}^{te})$$, and the central difficulty is that models tend to entangle domain-specific variation with label-relevant structure, producing spurious correlations that fail under domain shift. ERIS addresses this through three coordinated mechanisms: an energy-guided calibration mechanism, a weight-level orthogonality strategy, and an auxiliary adversarial generalization mechanism (Wu et al., 19 Aug 2025).

1. Problem setting and conceptual motivation

ERIS is situated in a common OOD setting for time series classification in which a model is trained on multiple source domains and evaluated on an unseen target domain whose distribution differs from training. The domains may correspond to subjects, sensors, environments, or operating conditions. The paper emphasizes that this setting is especially difficult in sensor time series because the same class label can occur under substantially different domain conditions (Wu et al., 19 Aug 2025).

The paper identifies the primary failure mode as entanglement of domain-specific and label-relevant features. In human activity recognition, for example, a model may associate a class such as “running” with higher amplitude or stronger motion energy that is characteristic of a particular subject or device, rather than with the underlying activity pattern itself. Such correlations act as shortcuts during training but become brittle when the domain changes, for example to a different person, device, or environment.

A central argument of ERIS is that many disentanglement methods are unguided. They may enforce orthogonality or independence, but they do not specify what should be separated. ERIS instead introduces semantic guidance. In this context, “semantic” does not refer to human language; it refers to an interpretable signal correlated with domain shift. The paper uses energy, including examples such as signal variance, spectral energy, and physical intensity, as such a cue. Energy-guided training is therefore intended to indicate which aspects of representation are likely to be domain-specific and should be separated from class information.

2. End-to-end architecture and joint objective

The ERIS pipeline begins with a shared feature extractor that maps an input time series $X$ to a shared representation $F_0$. Two energy branches then operate on $F_0$: a domain-specific energy (DSE) branch and a label-specific energy (LSE) branch. A weight-level orthogonality penalty is applied to keep the two branches structurally independent, and an adversarial branch injects structured perturbations to encourage robustness. The total loss jointly optimizes these modules (Wu et al., 19 Aug 2025).

The paper defines the total objective as

$$\mathcal{L}_{\text{total}} = \lambda_1 \mathcal{L}_{\text{DSE}} + \lambda_2 \mathcal{L}_{\text{LSE}} + \mathcal{L}_{\text{ortho}} + \mathcal{L}_{\text{reg}},$$

where $\mathcal{L}_{\text{DSE}}$ is the domain-specific energy loss, $\mathcal{L}_{\text{LSE}}$ is the label-specific energy loss, $\mathcal{L}_{\text{ortho}}$ is the orthogonality regularizer, and $\mathcal{L}_{\text{reg}}$ is the adversarial regularization term. The trade-off weights are $\lambda_1$ and $X$0. The paper notes that $X$1 is typically larger than $X$2, because the label or prototype branch is the main driver of classification, while domain disentanglement must still remain active.

This design reflects the paper’s broader claim that robust generalization in time series does not arise solely from mathematical separation of features. Instead, separation must be anchored by a semantically meaningful signal, and then reinforced structurally and locally through orthogonality and adversarial regularization.

3. Energy-guided calibration and self-calibrating separation

The energy-guided calibration mechanism is the component that provides semantic direction to the disentanglement process. From the shared feature $X$3, ERIS learns two energy functions:

$X$4

and

$X$5

Lower energy corresponds to higher compatibility with a domain or label.

For domain-specific energy, the objective is to assign low energy to the true domain and high energy to incorrect domains. The loss is

$X$6

The paper interprets this as encouraging low energy for the correct domain while pushing incorrect-domain energies above a margin. It also connects DSE to uncertainty calibration: lower energy corresponds to a more confident in-domain fit, and the model’s domain uncertainty $X$7 is conceptually inversely correlated with $X$8, so high energy indicates low match quality.

For label-specific energy, ERIS combines label energy with prototype distances. It defines the consistency score

$X$9

where $F_0$0 is the prototype for class $F_0$1. Smaller $F_0$2 means that the feature is both energetically compatible with the label and close to the class prototype. The label contrastive loss is

$F_0$3

and the prototype contrastive loss is

$F_0$4

The label-specific energy loss is then

$F_0$5

The paper characterizes this mechanism as self-calibrating. Because energy is tied to both domain fit and label fit, OOD samples tend to produce higher energy, which naturally reduces confidence. In this formulation, calibration is integrated into representation learning rather than treated as a post-processing step. A common misunderstanding addressed by the paper is that energy is used only as a detector; ERIS instead uses energy as a calibrated semantic score for directing separation between domain-specific and label-relevant information.

4. Structural disentanglement and adversarial generalization

A defining feature of ERIS is that its orthogonality mechanism is imposed at the weight level rather than only at the feature level. Let $F_0$6 denote the domain branch projection matrix and $F_0$7 denote the label branch projection matrix. The orthogonality loss is

$F_0$8

It is minimized when

$F_0$9

meaning that the column spaces of the two projection matrices are orthogonal (Wu et al., 19 Aug 2025).

The paper distinguishes this from feature-level orthogonality. Feature-level orthogonality constrains only the outputs for each training sample and is described as potentially brittle and local. Weight-level orthogonality, by contrast, constrains the global parameter subspaces, making the disentanglement structural and more robust on unseen or OOD inputs. The paper further gives a lemma showing that under gradient flow the orthogonality loss decreases monotonically and converges to $F_0$0, with gradients

$F_0$1

This formalizes the claim that optimization continuously pushes the parameter spaces apart.

The auxiliary adversarial generalization mechanism complements this structural separation. Its global component uses adversarial domain confusion, in which the feature extractor is trained to fool a domain discriminator so that latent features become domain-indistinguishable. Its local component constructs a structured adversarial perturbation

$F_0$2

and then minimizes

$F_0$3

The perturbation is described as a worst-case structured change within an $F_0$4 ball rather than random noise, and the purpose is to encourage smooth decision boundaries and local invariance.

Taken together, these mechanisms define a two-level robustness strategy: structural independence between domain and label pathways, and stability under small but adversarially chosen perturbations.

5. Disentanglement logic and empirical evaluation

ERIS treats the learned representation as comprising two conceptual parts: domain-specific features and label-relevant features. Domain-specific features include variations tied to subject, sensor, environment, amplitude or energy, and other shift factors. Label-relevant features are the class semantics that should remain stable across domains. The separation is guided because energy functions as the semantic clue: features aligning with domain energy patterns are encouraged into the domain branch, while energy-agnostic class structure is preserved in the label branch.

The empirical evaluation uses four multi-domain sensor time-series benchmarks: UCIHAR, UniMiB-SHAR, Opportunity, and EMG. These are OOD-style settings in which one or more domains are held out as target domains. The comparison includes four groups of baselines: energy-based methods (EOW-Softmax, ELI, EGC, Golan), general time-series methods (TARNet, PatchTST, TS-GAC, MPTSNet), general OOD generalization methods (GroupDRO, VREx, SDL, FOOGD, EVIL), and time-series OOD methods (GILE, AdaRNN, Diversify, ITSR) (Wu et al., 19 Aug 2025).

The paper reports that ERIS achieves the best overall performance. Its average accuracy across all datasets is reported as 75.47%, compared with 71.43% for the next best entry in the reported table, ITSR. ERIS is also ranked best overall by average rank, and Wilcoxon signed-rank tests report $F_0$5 for all comparisons. On calibration, ERIS attains the lowest average expected calibration error (ECE), 0.12, compared with 0.17 for EOW-Softmax, 0.26 for ELI, 0.39 for EGC, and 0.22 for Golan.

The ablation studies reported in the paper indicate that DSE alone and LSE alone provide reasonable performance, but not the best performance; adding orthogonality substantially improves results; and the full model with DSE, LSE, AG, and orthogonality performs best. A specific conclusion emphasized by the paper is that weight-level orthogonality is pivotal: removing $F_0$6 noticeably degrades ACC, F1, precision, and recall, and feature-level orthogonality is weaker than weight-level orthogonality in both parameter separation and final performance.

Qualitative analyses are also reported. In t-SNE visualizations, ITSR shows entangled clusters with domain overlap and blurry class boundaries, while ERIS forms compact label clusters with different domains collapsed together. Feature correlation and mutual information matrices are reported to show that ERIS suppresses redundant correlations more effectively than competing orthogonal or disentanglement methods.

6. Interpretation, assumptions, and scope

The paper’s practical interpretation is that ERIS addresses OOD time series classification from three directions at once: energy guidance indicates what is domain-like and what is class-like; orthogonality forces these kinds of information into separate parameter subspaces; and adversarial training discourages vulnerability to small, structured perturbations. The relationship among the modules is presented as complementary: energy regularization provides semantic direction, disentanglement uses that direction to separate domain and label factors, orthogonality stabilizes the separation structurally, and adversarial training improves local and cross-shift robustness (Wu et al., 19 Aug 2025).

Several assumptions are implicit in the method as presented. The framework assumes access to multiple source domains during training. It relies on the premise that energy is a meaningful semantic cue for the task. It uses a specific architecture with separate branches and learned prototypes, so performance may depend on the choice of backbone and branch capacities. The adversarial mechanism adds training complexity, although inference remains simple.

The paper does not foreground many explicit failure cases, but it does point to future work on the theoretical limits of energy-based modeling for temporal uncertainty. This suggests that ERIS should be understood not merely as a set of regularizers, but as a position within a broader line of research arguing that shift-robust time series classification requires semantically guided invariance rather than purely blind disentanglement.