AnchorMoE: Interpretable Time Series Classification via Anchor-Routed MoE

Jul 16, 2026

Multivariate time series classification (MTSC) supports high-stakes decisions in areas such as clinical diagnosis and industrial fault detection, where trustworthy deployment requires transparent reasoning. Yet identifying the temporal segments that actually drive predictions is difficult because discriminative cues are sparse, heterogeneous, and masked by background noise. AnchorMoE introduces an interpretable-by-construction Mixture-of-Experts framework whose predictions are an exact additive decomposition over input segments. A geometric orthogonality constraint prevents redundant expert representations, and an uncertainty-aware reliability gate suppresses residual noise. Extensive experiments show competitive accuracy with faithful, time-localized explanations grounded in raw signals.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

Introduction

Problem: sparse, heterogeneous temporal evidence

Limitations of post-hoc attribution

AnchorMoE overview

AnchorMoE: Interpretable Time Series Classification via Anchor-Routed MoE Tao Xie Guangdong University of Technology Guangzhou, China [3123001085@mail2.gdut.edu.cn](3123001085@mail2.gdut.edu.cn) Zexi Tan Guangdong University of Technology Guangzhou, China [3123004194@mail2.gdut.edu.cn](3123004194@mail2.gdut.edu.cn) Haoyi Xiao Guangdong University of Technology Guangzhou, China [3124004599@mail2.gdut.edu.cn](3124004599@mail2.gdut.edu.cn) Mengke Li Shenzhen University Shenzhen, China [mengkeli@szu.edu.cn](mengkeli@szu.edu.cn) Yiqun Zhang∗ Guangdong University of Technology Guangzhou, China [yqzhang@gdut.edu.cn](yqzhang@gdut.edu.cn) Yang Lu Xiamen University Xiamen, China [luyang@xmu.edu.cn](luyang@xmu.edu.cn) arXiv :2606 .0363 1v 3 [ cs .LG] 10 Jul 2026 Cuie Yang Northeastern University Shenyang, China [yangcuie@mail.neu.edu.cn](yangcuie@mail.neu.edu.cn) Abstract Multivariate time series classification (MTSC) is pivotal in highstakes domains, such as clinical diagnosis and industrial fault detection, where safe deployment necessitates transparent decisionmaking. However, isolating the temporal segments that drive model predictions is challenging because discriminative signals in realworld 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. CCS Concepts • Information systems → Data mining; • Computing methodologies → Machine learning. ∗ Corresponding author. This work is licensed under a Creative Commons Attribution 4 .0 International License. KDD’26, Jeju Island, Republic of Korea © 2026 Copyright held by the owner/author(s) . ACM ISBN 979-8-4007-2259-2/2026/08 [https://doi.org/10.1145/3770855.3817786](https://doi.org/10.1145/3770855.3817786) Yiu-ming Cheung Hong Kong Baptist University Hong Kong SAR, China [ymc@comp.hkbu.edu.hk](ymc@comp.hkbu.edu.hk) Keywords Multivariate Time Series Classification, Interpretable Machine Learning, Self-Explaining Models, Mixture of Experts ACM Reference Format: Tao Xie, Zexi Tan, Haoyi Xiao, Mengke Li, Yiqun Zhang, Yang Lu, Cuie Yang, and Yiu-ming Cheung. 2026. AnchorMoE: Interpretable Time Series Classification via Anchor-Routed MoE. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2 (KDD’26), August 09–13, 2026, Jeju Island, Republic of Korea. ACM, New York, NY, USA, 12 pages. [https://doi.org/10.1145/3770855.3817786](https://doi.org/10.1145/3770855.3817786) [Resource Availability:](Resource Availability:) The source code of this paper has been made publicly available at [https://doi](https://doi). org/10.5281/zenodo.20449598 (GitHub: [https://github.com/kuxit/AnchorMoE](https://github.com/kuxit/AnchorMoE)). 1 Introduction Multivariate Time Series (MTS) capture complex, co-evolving measurements pervasive in domains like network security, industrial monitoring [23, 24], and clinical c

FAQs

What problem does AnchorMoE target in multivariate time series classification?

AnchorMoE targets the difficulty of isolating the temporal segments that truly drive model predictions when discriminative signals are sparse, heterogeneous, and obscured by noise.

How does AnchorMoE make its decisions interpretable by construction?

It uses a Mixture-of-Experts design that routes multi-view representations of local patches to specialized experts, producing predictions as an exact additive decomposition over input segments for ante-hoc transparency.

What mechanisms improve the reliability of the decomposition under sparse signals?

AnchorMoE adds a geometric orthogonality constraint to reduce representational redundancy and an uncertainty-aware reliability gate to dynamically calibrate segment contributions and suppress residual background noise.

Jul 16, 2026

Upload by Olivia Brown