Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Capturing Temporal Components for Time Series Classification

Venkata Ragavendra Vavilthota, Ranjith Ramanathan, Sathyanarayanan N. Aakur

TL;DR

The paper tackles robust time series classification across varying time scales by decomposing signals into atomic temporal components via a multi-scale change space, then learning compositional representations with a masked autoencoder over the components using a Bidirectional LSTM encoder. It introduces a multi-task objective that combines self-supervised MAE with supervised cross-entropy, enabling effective classification and coherent segmentation. Empirical results on 85 UCR datasets show competitive accuracy against non-ensemble baselines and strong performance on long sequences, with the approach extending naturally to unsupervised segmentation tasks. The method is lightweight, scalable, and opens avenues for hierarchical, multivariate, and multimodal time series representations with minimal handcrafting.

Abstract

Analyzing sequential data is crucial in many domains, particularly due to the abundance of data collected from the Internet of Things paradigm. Time series classification, the task of categorizing sequential data, has gained prominence, with machine learning approaches demonstrating remarkable performance on public benchmark datasets. However, progress has primarily been in designing architectures for learning representations from raw data at fixed (or ideal) time scales, which can fail to generalize to longer sequences. This work introduces a \textit{compositional representation learning} approach trained on statistically coherent components extracted from sequential data. Based on a multi-scale change space, an unsupervised approach is proposed to segment the sequential data into chunks with similar statistical properties. A sequence-based encoder model is trained in a multi-task setting to learn compositional representations from these temporal components for time series classification. We demonstrate its effectiveness through extensive experiments on publicly available time series classification benchmarks. Evaluating the coherence of segmented components shows its competitive performance on the unsupervised segmentation task.

Capturing Temporal Components for Time Series Classification

TL;DR

The paper tackles robust time series classification across varying time scales by decomposing signals into atomic temporal components via a multi-scale change space, then learning compositional representations with a masked autoencoder over the components using a Bidirectional LSTM encoder. It introduces a multi-task objective that combines self-supervised MAE with supervised cross-entropy, enabling effective classification and coherent segmentation. Empirical results on 85 UCR datasets show competitive accuracy against non-ensemble baselines and strong performance on long sequences, with the approach extending naturally to unsupervised segmentation tasks. The method is lightweight, scalable, and opens avenues for hierarchical, multivariate, and multimodal time series representations with minimal handcrafting.

Abstract

Analyzing sequential data is crucial in many domains, particularly due to the abundance of data collected from the Internet of Things paradigm. Time series classification, the task of categorizing sequential data, has gained prominence, with machine learning approaches demonstrating remarkable performance on public benchmark datasets. However, progress has primarily been in designing architectures for learning representations from raw data at fixed (or ideal) time scales, which can fail to generalize to longer sequences. This work introduces a \textit{compositional representation learning} approach trained on statistically coherent components extracted from sequential data. Based on a multi-scale change space, an unsupervised approach is proposed to segment the sequential data into chunks with similar statistical properties. A sequence-based encoder model is trained in a multi-task setting to learn compositional representations from these temporal components for time series classification. We demonstrate its effectiveness through extensive experiments on publicly available time series classification benchmarks. Evaluating the coherence of segmented components shows its competitive performance on the unsupervised segmentation task.
Paper Structure (10 sections, 4 equations, 2 figures, 4 tables)

This paper contains 10 sections, 4 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Framework
  4. Discovering Temporal Components of Signals
  5. Capturing Signal Compositionality
  6. Experimental Evaluation
  7. Experimental Setup
  8. Quantitative Evaluation
  9. Extension to Unsupervised Time Series Segmentation
  10. Discussion and Future Work

Figures (2)

  • Figure 1: Overall architecture of the proposed approach is illustrated here. First, we introduce a multi-scale state change detection model to segment sequential data into components and then use a sequence-based encoder to learn compositional representations for time series classification.
  • Figure 2: Qualitative visalization of (a) a successful segmentation and (b) unsuccessful segmentation on the GreatBarbet2 and SuddenCardianDeath1 datasets, respectively. The first row shows ground truth segments, and the second shows predicted segments.