.png){kind=icon}
.png){kind=icon}
Return
STaTS: Structure-aware temporal sequence summarization via statistical window merging
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patrec.2026.02.014.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Time series data often exhibit latent temporal structure, transitions between locally stationary regimes, motifs, and bursts of variability, that standard representation learning methods largely ignore. Existing models typically operate on raw or fixed-window sequences, treating all time steps as equally informative, which hinders efficiency, robustness, and scalability. We propose STaTS, a lightweighast, unsupervised framework for Structure-Aware Temporal Summarization that adaptively compresses univariate and multivariate time series into compact, information-preserving token sequences. STaTS detects change points across multiple temporal resolutions via a Bayesian Information Criterion (BIC) based statistical divergence criterion and summarizes each segment using simple functions (e.g., mean) or generative models such as Gaussian Mixture Models (GMMs). This approach reduces sequence length by a factor of up to 30 while preserving core dynamics and requires no retraining. Integrated as a model-agnostic preprocessor into unsupervised encoders like TS2Vec, STaTS retains 85-90% of full-model performance across 150+ datasets, including UCR-85/128, UEA-30, and ETT benchmarks, while significantly reducing computational cost. It also improves robustness to noise and outperforms uniform and clustering-based baselines in preserving task-relevant structure.
Keywords:
Time series summarization
Time series classification
Time series forecasting
Sequence compression
Journal
IF:
3.3
Papers:
7.9K
Citations:
1.6W