Unsupervised segmentation of biosignal-based time-series

The rise of the Internet of Things (IoT) and the development of more compact and less power-hungry sensors have led to an increasing amount of data from various modalities. To analyze a large amount of continuously recorded time series data at scale, it is often beneficial to separate the data into homogenous segments before further analysis or classification. One common way to do time series segmentation is to find the transitions that separate the different segments. These transitions are in the literature referred to as changepoints, and the problem of discovering them is referred to as change point detection (CPD). Long time series data is hard and time-consuming to label; thus, having a CPD algorithm that works in an unsupervised fashion is highly beneficial. This thesis provides an overview of automatic change point detection, time series segmentation, and previous research concerning representation learning in the context of CPD. The main contribution of this thesis is the Latent Space Unsupervised Semantic Segmentation (LS-USS) model, which is an unsupervised segmentation algorithm that is hyperparameter lite, domain agnostic, and works well with multidimensional data. This model utilizes an autoencoder for extracting constrained feature encodings from time-series data before using the similarity between neighboring feature encodings in time to do changepoint detection and segmentation. The LS-USS model is compared with other state-of-the-art algorithms on multiple datasets. The results show that the model performs similarly or better than other state-of-the-art models. To demonstrate a practical application of LS-USS, a dataset containing activity data from non-depressed and depressed participants is used. Using the LS-USS algorithm as a preprocessing step before classification significantly outperforms the previous best classification results on this dataset