Semi-supervised learning for time series collected at a low sampling rate

Abstract

Although time series classification has recently received a lot of attention, two considerations make it difficult. First, time series labeling is expensive due to complicated temporal and dimensional data structures. Second, large scale time series data gathered at high sampling rates is associated with the high measurement and storage costs. To address the challenges, we propose a novel problem, semi-supervised learning for low-sampling-rate time series, utilizing few high sampling rate data that are available in most data gathering environments. Since down sampling drops class discriminative features such as short periodicities and peak events, we propose a high-resolution reconstructor with a temporal upsampler. Reconstructed data are then passed through a classifier that is regularized using sampling shift consistency to make the classification of dissimilar shifted down-samples consistent. We demonstrate that our approach outperforms standard semi-supervised learning techniques and propose a guideline that, according to empirical data, ensures high classification performance across various sampling rate time series.