ERIS: An Energy-Guided Feature Disentanglement Framework for Out-of-Distribution Time Series Classification (2508.14134v2)

Abstract: An ideal time series classification (TSC) should be able to capture invariant representations, but achieving reliable performance on out-of-distribution (OOD) data remains a core obstacle. This obstacle arises from the way models inherently entangle domain-specific and label-relevant features, resulting in spurious correlations. While feature disentanglement aims to solve this, current methods are largely unguided, lacking the semantic direction required to isolate truly universal features. To address this, we propose an end-to-end Energy-Regularized Information for Shift-Robustness (ERIS) framework to enable guided and reliable feature disentanglement. The core idea is that effective disentanglement requires not only mathematical constraints but also semantic guidance to anchor the separation process. ERIS incorporates three key mechanisms to achieve this goal. Specifically, we first introduce an energy-guided calibration mechanism, which provides crucial semantic guidance for the separation, enabling the model to self-calibrate. Additionally, a weight-level orthogonality strategy enforces structural independence between domain-specific and label-relevant features, thereby mitigating their interference. Moreover, an auxiliary adversarial generalization mechanism enhances robustness by injecting structured perturbations. Experiments across four benchmarks demonstrate that ERIS achieves a statistically significant improvement over state-of-the-art baselines, consistently securing the top performance rank.

• The paper's main contribution is the ERIS framework, which utilizes energy-guided calibration to effectively separate domain-specific and label-relevant features for improved out-of-distribution robustness.
• It applies weight-level orthogonality directly to the network's weight matrices, enforcing independent feature learning and reducing mutual interference between domain and label branches.
• The incorporation of adversarial generalization with structured contrastive loss yields superior classification performance, demonstrated by reduced Expected Calibration Error on benchmark datasets.

ERIS: An Energy-Guided Feature Disentanglement Framework for Out-of-Distribution Time Series Classification

Introduction

The paper addresses the significant challenge of Out-of-Distribution (OOD) generalization in Time Series Classification (TSC). It highlights the limitations of existing feature disentanglement methods, which often rely solely on mathematical constraints without incorporating semantic guidance. This lack of semantic anchoring results in the entanglement of domain-specific and label-relevant features, leading to poor OOD performance. To circumvent this, the authors propose the ERIS framework, which integrates energy-based semantic guidance and structured constraints to enable reliable feature disentanglement and enhance OOD robustness.

Core Components of ERIS

Energy-Guided Calibration Mechanism

ERIS incorporates domain-specific and label-specific energy functions, designed to provide semantic guidance during feature separation. These energy functions generate lower energy values for well-aligned features, thereby enhancing the model's ability to differentiate between domain-specific and label-relevant attributes. The framework also employs a contrastive structured loss that minimizes energy for true domains and labels while maximizing it for others, ensuring effective feature separation.

Weight-Level Orthogonality Strategy

Distinct from common feature-level orthogonal methods, ERIS applies orthogonality constraints directly to the weight matrices ($\boldsymbol{W_d}, \boldsymbol{W_l}$) of the domain and label branches. This approach enforces structural independence at the parameter level, significantly reducing mutual interference between the domain and label subspaces. The orthogonality loss $\mathcal{L}_{\mathrm{ortho}}$ ensures that projection matrices ($\boldsymbol{W_d}$ and $\boldsymbol{W_l}$) tend towards zero correlation, thereby enhancing feature disentanglement.

Adversarial Generalization (AG)

To further boost robustness, ERIS implements an adversarial generalization strategy that aligns domain distributions globally and smooths the latent space locally. This is achieved by introducing structured perturbations during training, thereby encouraging the learning of invariant features that remain robust under distribution shifts. The KL-divergence-driven adversarial perturbations optimize the feature space to resist local adversarial attacks, promoting more robust generalization.

Experimental Evaluation and Results

Extensive experiments conducted on diverse benchmark datasets (UCIHAR, UniMiB-SHAR, Opportunity, and EMG) demonstrate that ERIS consistently outperforms state-of-the-art baselines across several metrics, including accuracy, F1-score, precision, and recall. The introduction of energy guidance and orthogonality constraints proves crucial in enhancing model calibration and feature robustness, as evidenced by significant reductions in Expected Calibration Error (ECE).

Implementation Insights

ERIS’s architecture is adaptable to various neural networks including MLP, CNN, and Transformer backbones. It employs a modular design allowing easy integration of the three core components across different architectures. The framework requires careful tuning of hyperparameters, particularly the energy function margins and the balance between different loss components, to achieve optimal performance. Empirical results suggest maintaining a higher relative weight for the label-specific energy loss to ensure discriminative clustering around class prototypes.

Conclusion

ERIS offers a robust solution to the challenge of feature disentanglement in OOD settings by integrating energy-guided semantic directions with orthogonality constraints. Its ability to effectively separate domain-specific and label-relevant features has wide implications for improving time series model generalization. Future research may extend ERIS’s applications to other domains and explore its potential in real-time adaptive systems for enhanced robustness.