🤖 machine learning

AnchorMoE: Interpretable Time Series Classification via Anchor-Routed MoE

This paper proposes AnchorMoE, an interpretable-by-construction framework for multivariate time series classification that leverages a Mixture-of-Experts architecture with geometric orthogonality constraints and an uncertainty-aware gate to achieve competitive performance while ensuring transparent, additive decision-making grounded in raw signal segments.

Original authors: Tao Xie, Zexi Tan, Haoyi Xiao, Mengke Li, Yiqun Zhang, Yang Lu, Cuie Yang, Yiu-ming Cheung

Published 2026-06-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tao Xie, Zexi Tan, Haoyi Xiao, Mengke Li, Yiqun Zhang, Yang Lu, Cuie Yang, Yiu-ming Cheung

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery based on a long, messy audio recording. The recording is full of static, wind noise, and random chatter (the "background noise"), but somewhere in there, a few specific words were spoken that actually solve the case (the "discriminative signals").

Most current AI detectives use a "post-hoc" approach: they listen to the whole recording, guess the answer, and then try to point their finger at the part of the recording they think was important. The problem is, they often point at the wrong thing—like blaming the wind noise because it happened to be loud at the same time.

AnchorMoE is a new kind of detective that is built differently from the ground up. Instead of guessing first and explaining later, it is designed so that the explanation is built into the very act of solving the mystery.

Here is how it works, using simple analogies:

1. The Team of Specialists (The "Mixture of Experts")

Imagine you have a team of five different specialists.

Specialist A is great at spotting patterns in the speed of the sound.
Specialist B is great at spotting patterns in the pitch (frequency).
Specialist C looks at how the sound fits into the whole story.

In older AI models, these specialists were forced to listen to the entire recording at the same time. This meant they all heard the same background noise and ended up giving the same, muddy advice.

AnchorMoE changes the rules. It cuts the recording into small chunks (patches). It then has a smart manager who sends each chunk to the one specialist best suited to handle it.

A chunk with a weird speed spike goes to Specialist A.
A chunk with a high-pitched squeal goes to Specialist B.

2. The "Anchor" System (Keeping Specialists Honest)

Here is the tricky part: How do you make sure the specialists don't all start doing the exact same job? If they all listen to the same noise, they will all agree on the wrong answer.

AnchorMoE uses a system called Orthogonal Posterior Anchors. Think of this as a "personal space" rule for the team.

The system forces each specialist to develop a unique "mental anchor" or a specific type of evidence they are looking for.
It mathematically punishes them if their "anchors" look too similar.
The Result: Specialist A must look for speed patterns, and Specialist B must look for pitch patterns. They cannot overlap. This ensures that when they give their opinion, they are actually looking at different, unique clues, not just repeating the background noise.

3. The "Reliability Gate" (The Noise Filter)

Even with specialized experts, some chunks of the recording are just pure garbage (static or wind). If you just add up everyone's opinions, that garbage could accidentally tip the scale and change the verdict.

AnchorMoE adds a Reliability Gate. Before the final verdict is written down, this gate checks every single chunk:

"Is this chunk actually useful, or is it just noise?"
If a chunk is noisy, the gate turns its volume down to almost zero.
If a chunk is a clear clue, the gate lets it speak loudly.

This is like a bouncer at a club who checks IDs. If you look like you belong (you have a real clue), you get in. If you are just a random passerby (background noise), you are stopped at the door.

4. The Final Verdict (The "Additive" Solution)

Finally, AnchorMoE adds up the contributions of the approved chunks.

Old AI: "I think it's a cat because the whole video looks like a cat, but I'm not sure which part made me think that."
AnchorMoE: "I think it's a cat because Chunk #3 had a purr (Specialist A said yes), and Chunk #7 had whiskers (Specialist B said yes). The rest of the video was just a blurry background, so I ignored it."

Because the final answer is just a simple sum of these specific, approved chunks, you can look at the math and see exactly which parts of the data led to the decision. There is no guessing.

Why Does This Matter?

The paper claims that this method is interpretable by design.

Old way: You build a black box, then try to shine a flashlight on it to see inside (often the flashlight is dim or misleading).
AnchorMoE: The machine is built with glass walls. You can see the gears turning, and you know exactly which gear made the machine move.

The authors tested this on many different types of data (from heartbeats to industrial machines) and found that AnchorMoE is not only very accurate at solving the problems but also incredibly honest about why it solved them, successfully ignoring the "noise" that confuses other AI models.

Technical Summary: AnchorMoE

Problem Statement
Multivariate Time Series Classification (MTSC) is critical in high-stakes domains like clinical diagnosis and industrial fault detection, where transparency is as vital as accuracy. However, achieving interpretable MTSC is fundamentally challenging due to the nature of real-world signals: discriminative evidence is typically sparse, localized within brief temporal windows, and restricted to specific variables, while the remainder of the sequence consists of uninformative background noise. Furthermore, these signals are heterogeneous, comprising distinct patterns across variables, timeframes, and frequency domains.

Existing solutions rely heavily on post-hoc attribution methods (e.g., gradient-based, perturbation-based, or attention-weight analysis) to explain model decisions after prediction. These approaches suffer from structural limitations: they decouple explanation from computation, leading to potential unfaithfulness where explanations may reflect information routing rather than causal importance. Additionally, post-hoc methods lack architectural safeguards to prevent the mis-attribution of correlated background noise, a significant risk given the sparsity of true signals.

Methodology: AnchorMoE
To address these challenges, the authors propose AnchorMoE, an interpretable-by-construction framework based on a Mixture-of-Experts (MoE) architecture. Unlike standard MoE models that enforce uniform load balancing (often causing experts to learn redundant representations and absorb noise), AnchorMoE formulates the final prediction as an exact additive decomposition over input segments. The architecture consists of three core components:

Multi-View Representation Embedding (MVRE):
Before routing, raw local patches are enriched into comprehensive evidence units. MVRE encodes three distinct views:
- Temporal View: Preserves ordered local morphology via convolutional embedding and positional encoding.
- Spectral View: Summarizes frequency content across multiple bands to capture regularities independent of exact temporal alignment.
- Contextual View: Situates each patch relative to the global sequence context by computing a mask-weighted centroid and scoring patch alignment against it.
  These views are fused into a holistic patch representation.
Diversified Posterior-Anchor Routing (DPAR):
To prevent expert redundancy and ensure specialization in heterogeneous patterns, DPAR introduces a geometric orthogonality constraint.
- Posterior Anchors: For each expert, a "posterior anchor" is dynamically induced as the weighted centroid of the patch units routed to that expert.
- Orthogonality Loss: A loss function penalizes the cosine similarity between these anchors. This compels distinct experts to specialize in different predictive patterns rather than converging on homogeneous representations, thereby isolating heterogeneous signals.
Reliability-Gated Composition (RGC):
To ensure the additive decomposition remains reliable under sparse signal distributions, RGC dynamically calibrates the contribution of each segment.
- Uncertainty-Aware Gating: A reliability gate evaluates the confidence of the representation generated by each expert for a specific patch.
- Sparse Routing: The composition is restricted to the top- $K$ experts for each patch.
- Global Admission: A global admission weight is computed based on the aggregated reliability scores. This mechanism suppresses uninformative background segments and residual noise, ensuring that only high-confidence, evidence-bearing segments contribute significantly to the final additive sum.

The final prediction is formulated as a transparent sum of calibrated patch contributions ( $u_c = \sum \pi_p \ell_{p,c}$ ), allowing for direct inspection of the decision process without external estimators.

Key Contributions

Interpretable-by-Design Formulation: The paper reframes MTSC as an expert-routing problem where attribution is an exact, additive decomposition of the decision process, eliminating the need for unreliable post-hoc approximations.
Architectural Innovations: The introduction of Orthogonal Posterior Anchors to enforce expert specialization in heterogeneous patterns and an Uncertainty-Aware Reliability Gate to dynamically suppress background noise. These designs structurally prevent the additive decomposition from being corrupted by sparse signals.
Rigorous Evaluation: The authors construct synthetic datasets with controlled distractors and ground-truth masks to validate that the model's decisions are faithfully grounded in raw data, not just correlated noise.

Experimental Results
Extensive experiments were conducted on 18 real-world UEA benchmark datasets and four synthetic datasets.

Classification Performance: AnchorMoE achieved state-of-the-art performance, ranking first in average rank for both Accuracy (1.53) and F1-score (1.28) across the 18 datasets, outperforming strong baselines including Transformer-based models (PatchTST, FEDformer) and MLP architectures. It showed particular superiority on datasets with localized, non-stationary signals (e.g., Cricket, RacketSports).
Explanation Faithfulness: On real-world datasets, AnchorMoE consistently achieved optimal "Gap" scores (measuring ranking faithfulness) and high feature retention ( $I(20)$ ), outperforming nine post-hoc baselines (including Integrated Gradients, DeepLIFT, and TimeX).
Evidence Localization: On synthetic datasets with known ground-truth masks, AnchorMoE significantly outperformed post-hoc methods in AUPRC, IoU@K, and Precision@K. Notably, while gradient-based methods degraded to random performance in sparse settings, AnchorMoE maintained robust localization, successfully ignoring salient but label-irrelevant distractors.
Ablation Studies: Results confirmed that DPAR is critical for class discrimination via expert diversity, while RGC is paramount for faithful evidence localization by filtering noise.

Significance and Claims
The paper claims that AnchorMoE demonstrates that strict ante-hoc transparency does not necessitate a compromise in predictive performance. By unifying classification and explanation into a single process, the model provides a faithful decomposition of decisions directly from the forward pass. The authors emphasize that their approach resolves the structural misalignment of standard MoE architectures in MTSC, where uniform load balancing typically forces experts to absorb background noise. While the additive formulation limits the capture of higher-order segment interactions, the work establishes a robust framework for transparent decision-making in high-stakes multivariate time series applications.