Unified Time Series Explanations via Amortized Optimization and Instance-level Multi-Expert Knowledge Distillation

Deep neural networks (DNNs) achieve high accuracy on time series classification (TSC) but remain opaque, hindering deployment in sensitive domains. Existing post-hoc TSC explanation methods rely on a single attribution perspective and incur high per-instance computational cost, limiting real-time use. We propose XMA, a framework that unifies multiple post-hoc explainers at the instance level and amortizes the cost of explanation into a learned inference pass. Instance-level Multi-Expert Knowledge Distillation (IMEKD) selects, per instance, the attribution map with the highest combined faithfulness and robustness score from a pool of XAI experts. Objective-Regularized Amortized Optimization Explanation (ORAOE) trains an explainer DNN to reproduce this instance-best supervision while directly optimizing differentiable faithfulness and robustness objectives, letting the student refine beyond any individual teacher. Faithfulness-Preserving Segmentation (FPS) converts point-wise maps into contiguous segments without changing their faithfulness score. Across four synthetic datasets, MIT-ECG with cardiologist-verified ground truth, and 11 multivariate UEA benchmarks on three DNN architectures, XMA attains the best AUPRC on all synthetic and MIT-ECG benchmarks among compared methods, and the highest faithfulness on 7 of 11 and robustness on 8 of 11 UEA datasets, while amortizing explanation into a lightweight learned explainer rather than relying on per-instance optimization.

When doctors and engineers rely on AI to read signals that change over time, such as a heartbeat on an ECG monitor or a sensor stream from a machine, they need to know why the AI reached its decision, not just what it decided. Tools that produce these explanations already exist, but they have two problems: different tools disagree about which parts of the signal mattered, and the most accurate ones are slow, sometimes taking minutes per case. In a hospital monitoring many patients at once, that delay can be dangerous. We built a system that watches several existing explanation tools, learns which one works best for each individual case, and trains a single fast model to reproduce and improve on their combined judgment. Once trained, the model explains a new signal almost instantly. This efficiency matters in sensitive applications such as ECG monitoring, where faster explanations could support quicker clinical decisions. Our method produces explanations that match expert cardiologist annotations more closely than previous approaches, while running hundreds of times faster. Because the heavy training is done once, a well-resourced institution can train the model centrally and then deploy a fast, lightweight version wherever speed and limited computing power matter most.