Skip to main content
Back

AnchorMoE Makes AI Time Series Classifiers Explain Themselves Without Approximations

AnchorMoE routes time series patches to specialized experts and explains every prediction as an exact additive sum over raw input segments — no post-hoc approximations needed.

Paper, Full Access
  • AnchorMoE is a new multivariate time series classifier whose predictions are exact additive sums over input segments — making explanations mathematically identical to the computation. [Paper]
  • It uses a Mixture-of-Experts architecture where each patch is routed to a specialized expert via learned anchor prototypes, not generic gating functions.
  • A geometric orthogonality constraint forces experts to specialize in distinct signal patterns, preventing the collapse seen in standard MoE on sparse, noisy data.
  • An uncertainty-aware reliability gate suppresses noisy patches from contributing to the final prediction, keeping attributions clean and faithful.
  • Evaluated on 18 UEA benchmark datasets plus 4 synthetic interpretability benchmarks with ground-truth annotations; all three components improve both accuracy and attribution quality.
  • No code release yet; the model is best suited for high-stakes domains (clinical, industrial) where segment-level explanations must be trustworthy, not approximate.

Most high-performing time series classifiers are black boxes. They can tell you a patient is at risk or a machine is about to fail, but they cannot tell you which part of the signal made them think so. The standard workaround is to train the model first, then bolt on an explanation method afterward , a process called post-hoc interpretability. The problem is that these after-the-fact explanations are approximations: they estimate what the model probably cared about, but they are not the actual computation the model performed. For clinical diagnosis or industrial fault detection, that gap matters enormously.

A new paper from a team of researchers introduces AnchorMoE, a multivariate time series classifier that bakes interpretability directly into its architecture. Instead of explaining predictions after the fact, the model is designed so that every prediction is literally computed as a weighted sum of contributions from individual time segments. The explanation and the prediction are the same operation.

Why time series interpretability is uniquely hard

Multivariate time series classification (MTSC) is already a tough problem. You have multiple sensor channels evolving over time, and the signal you care about , the pattern that distinguishes a seizure from normal brain activity, or a bearing fault from normal vibration , is typically buried in noise. The most effective models used for time series classification are not interpretable and hide the logic of the decision process, making them unusable in sensitive domains.

The two dominant approaches to fixing this are:

  • Post-hoc methods (SHAP, saliency maps, attention weights): These analyze and interpret the decision-making process of a trained model after it has made predictions, providing insights into how the model arrived at its outputs. The catch is that these are approximations , they can be unfaithful to what the model actually computed.
  • Ante-hoc methods