SAWE: Signature-based Aggregated Window Embedding for Unsupervised Time Series Segment Clustering
Marcell Németh, Gábor Szücs · 2024
This paper introduces SAWE (Signature Aggregated Window Embedding), an advanced methodology for unsupervised time series segment clustering. Central to our approach is the use of path signature embeddings, which excel in capturing the intricate nonlinear dynamics inherent in time series data. Path signatures transform sequences of data points into a feature set that succinctly encodes both the order and magnitude of data interactions, facilitating a robust representation that is invariant to common distortions and noise. By combining these embeddings with change point detection, and aggregating them within intervals defined by detected change points, SAWE significantly reduces sensitivity to noise and enhances feature stability. The adoption of a sliding window technique further optimizes the extraction and aggregation process, resulting in superior classification accuracy and clustering metrics compared to non-aggregated, single segment embeddings (SSE), particularly effective in handling multi-class time series data.