Prototype-Guided Classification Sub-Task Decoupling Framework: Enhancing Generalization and Interpretability for Multivariate Time Series

2026-05-21

AI Summary

This work addresses the limited interpretability and generalization of existing time series classification methods, which typically rely on end-to-end black-box mappings. To overcome these limitations, we propose PDFTime, a novel framework that reformulates classification as a prototype-guided, multi-granularity, and decoupled similarity reasoning process. By modeling class-conditional distributions in latent space and decomposing the task into multiple progressive sub-stages, PDFTime enables interpretable and incremental discrimination. The approach substantially enhances both model interpretability and generalization, achieving state-of-the-art performance on the UEA/UCR benchmark suite—attaining the highest accuracy on 80 out of 128 UCR datasets and significantly outperforming strong existing baselines.

Abstract

Time Series Classification (TSC) is a long-standing research problem that has gained increasing attention in recent years with the rapid growth of large-scale temporal data. Despite substantial progress enabled by deep learning, designing TSC models that are both accurate and interpretable remains a challenging task. Many existing approaches adopt a direct feature-to-label classification paradigm, by collapsing high-dimensional temporal embeddings into class logits via a single linear projection (often after global pooling), the paradigm conflates feature extraction and decision logic into an inseparable mapping. To address these limitations, we propose PDFTime, a prototype-guided framework that reformulates time series classification as a multi-stage decision process. Instead of direct feature-to-label mapping, PDFTime leverages learned prototypes to approximate class-conditional feature distributions in the latent space, enabling progressive discrimination through classification sub-tasks of varying granularity. To our knowledge, PDFTime is the first framework to reformulate time series classification as a decoupled, multi-stage similarity-based reasoning process, breaking the long-standing paradigm of direct, black-box feature-to-label mapping. Extensive evaluations demonstrate that PDFTime achieves state-of-the-art (SOTA) performance across UEA and UCR benchmarks. Notably, it secures the top-$1$ accuracy on 80 out of 128 datasets in the UCR archive, significantly outperforming recent strong baselines in both consistency and generalization.

Problem

• Time Series Classification
• Interpretability
• Generalization
• Multivariate Time Series
• Prototype Learning

Innovation

• prototype-guided
• sub-task decoupling
• multivariate time series classification
• similarity-based reasoning
• interpretability