Title: AnchorMoE: Interpretable Time Series Classification via Anchor-Routed MoE

Original: arXiv:2606.03631v1 Announce Type: cross Abstract: Multivariate time series classification (MTSC) is pivotal in high-stakes domains, such as clinical diagnosis and industrial fault detection, where safe deployment necessitates transparent decision-making. However, isolating the temporal segments that drive model predictions is challenging because discriminative signals in real-world time series are typically sparse, heterogeneous, and heavily obscured by background noise. This paper, therefore, proposes AnchorMoE, an interpretable-by-construction classification framework. Built upon a Mixture-of-Experts (MoE) architecture, AnchorMoE encodes multi-view representations of local patches and routes them to specialized experts, ensuring that the final prediction is formulated as an exact additive decomposition over the input segments, facilitating ante-hoc transparency rather than relying on post-hoc estimations. To maintain the reliability of this decomposition under sparse signal distributions, we introduce a geometric orthogonality constraint that penalizes representational redundancy, compelling distinct experts to specialize in heterogeneous predictive patterns. Furthermore, an uncertainty-aware reliability gate is designed to dynamically calibrate the contribution of each segment, effectively suppressing residual background noise. Extensive experiments on real-world and synthetic benchmarks demonstrate that AnchorMoE achieves highly competitive classification performance while faithfully grounding its decisions in the raw time series.

Rewritten:

Title: AnchorMoE: Interpretable Time Series Classification via Anchor-Routed MoE

Original: arXiv:2606.03631v1 Announce Type: cross Abstract: Multivariate time series classification (MTSC) is pivotal in high-stakes domains, such as clinical diagnosis and industrial fault detection, where safe deployment necessitates transparent decision-making. However, isolating the temporal segments that drive model predictions is challenging because discriminative signals in real-world time series are typically sparse, heterogeneous, and heavily obscured by background noise. This paper, therefore, proposes AnchorMoE, an interpretable-by-construction classification framework. Built upon a Mixture-of-Experts (MoE) architecture, AnchorMoE encodes multi-view representations of local patches and routes them to specialized experts, ensuring that the final prediction is formulated as an exact additive decomposition over the input segments, facilitating ante-hoc transparency rather than relying on post-hoc estimations. To maintain the reliability of this decomposition under sparse signal distributions, we introduce a geometric orthogonality constraint that penalizes representational redundancy, compelling distinct experts to specialize in heterogeneous predictive patterns. Furthermore, an uncertainty-aware reliability gate is designed to dynamically calibrate the contribution of each segment, effectively suppressing residual background noise. Extensive experiments on real-world and synthetic benchmarks demonstrate that AnchorMoE achieves highly competitive classification performance while faithfully grounding its decisions in the raw time series.

Rewritten:

In critical fields like industrial fault detection and clinical diagnosis, multivariate time series classification (MTSC) plays a vital role. Because safe implementation in these areas demands clear, transparent reasoning, understanding the decision-making process is essential. Nevertheless, identifying the specific time intervals that influence model outcomes is difficult. In practical applications, the signals that distinguish between classes are often scattered, varied, and masked by significant background noise.

To address this, we introduce AnchorMoE, a classification system designed with interpretability as a core feature. Utilizing a Mixture-of-Experts (MoE) structure, the model captures diverse perspectives of local data segments and directs them to specialized experts. This approach ensures that the ultimate prediction results from a precise additive breakdown of the input segments. Consequently, the model offers transparency before the prediction is made (ante-hoc), avoiding the need for post-hoc explanations.

To ensure this additive structure remains trustworthy when signals are sparse, we apply a geometric orthogonality constraint. This mechanism discourages redundant representations, forcing different experts to focus on distinct and varied predictive patterns. Additionally, we developed a reliability gate that accounts for uncertainty. This gate adjusts the influence of each segment in real-time, successfully minimizing the impact of lingering background noise.

Tests conducted on both synthetic and real-world datasets show that AnchorMoE delivers strong classification results. Importantly, these decisions are firmly rooted in the original time series data, ensuring that the model’s logic is both accurate and traceable.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC