Identifying Label Noise in Time-Series Datasets

Reliably labeled datasets are crucial to the performance of supervised learning methods. Time-series data pose additional challenges. Data points lying on borders between classes can be mislabeled due to perception limitations of human labelers. Sensor measurements may not be directly interpretable by humans. Thus label noise cannot be manually removed. As a result, time-series datasets often contain a significant amount of label noise that can degrade the performance of machine learning models. This work focuses on label noise identification and removal by extending previous methods developed for static instances to the domain of time-series data. We use a combination of deep learning and visualization algorithms to facilitate automatic noise removal. We show that our approach can identify mislabeled instances, which results in improved classification accuracy on four synthetic and two real publicly available human activity datasets.