TSMix: time series data augmentation by mixing sources

Abstract

Data augmentation for time series is challenging because of the complex multi-scale relationships spanning ordered continuous sequences: one cannot easily alter a single datum and expect these relationships to be preserved. Time series datum are not independent and identically distributed random variables. However, modern Function as a Service (FaaS) infrastructure yields a unique opportunity for data augmentation because of the multiple distinct functions within a single data source. Further, common strong periodicity afforded by the human diurnal cycle and its link to these data sources enables mixing distinct functions to form pseudo-functions for improved model training. Herein we propose time series mix (TSMix), where pseudo univariate time series are created by mixing combinations of real univariate time series. We show that TSMix improves the performance on held-out test data for two state-of-the-art forecast models (N-BEATS and N-HiTS) and linear regression.