SegTime: Precise Time Series Segmentation without Sliding Window

Time series are common in a wide range of domains and tasks such as stock market partitioning, sleep stage labelling, and human activity recognition, where segmentation, i.e. splitting time series into segments that correspond to given categories, is often required. A common approach to segmentation is to sub-sample the time series using a sliding window with a certain length and overlapping stride, to create sub-sequences of fixed length, and then classify these sub-sequences into the given categories. This reduces time series segmentation to classification. However, this approach guarantees to find only approximate breakpoints: the precise breakpoints can appear in sub-sequences, and thus the accuracy of segmentation degrades when labels change fast. Also, it ignores possible long-term dependencies between sub-sequences. We propose a neural networks approach SegTime that finds precise breakpoints, obviates sliding windows, handles long-term dependencies, and it is insensitive to the label changing frequency. SegTime does so, thanks to its bi-pass architecture with several structures that can process information in a multi-scale fashion. We extensively evaluated the effectiveness of SegTime with very promising results.