Towards Uncertainty Quantification for Time Series Segmentation

Huiping Cao, Erick Draayer

Abstract: Time Series Segmentation (TSS) is a data mining task widely used in many applications to generate a set of change points for a time series. Current TSS performance analyses focus on accuracy and, therefore, fail to fully evaluate the reliability and originality of a segmentation. We investigate using uncertainty quantification (UQ) to fully evaluate TSS performance. We propose UQ-TSS, a framework to quantify uncertainties surrounding TSS. UQ-TSS captures uncertainties from different sources in an integrative manner. It incorporates a novel TS augmentation algorithm to address inherent uncertainty in the data. It uses ensemble learning in a novel way to create samples and estimate the probability distributions of changepoint presence and locations. We demonstrate the ability of UQ-TSS to guide hyperparameter selection, refine segmentations, and determine an algorithm’s suitability for segmenting without the need for ground truth. We validate these claims through extensive experimentation using several well-established TSS algorithms and datasets.