Online prediction of time series with assumed behavior

Abstract The prediction of future time series values is essential for many fields and applications. In some settings, the time series behavior is expected to follow distinct patterns which in turn may change over time due to a change in the user’s preferences/behavior or a change in the environment itself. In this article, we propose to leverage the assumed time series behavior by developing specialized novel online machine learning algorithms. To demonstrate the potential benefits of our approach compared to existing practices we focus on two commonly assumed time series behaviors: exponential decay and sigmoidal. We present two innovative online learning algorithms, Exponentron for the prediction of exponential decay time series and Sigmoidtron for the prediction of sigmoidal time series. We provide an extensive evaluation of both algorithms both theoretically and empirically using synthetic and real-world data. Our results show that the proposed algorithms compare favorably with the classic time series prediction methods commonly deployed today by providing a substantial improvement in prediction accuracy. Furthermore, we demonstrate the potential applicative benefit of our approach for the design of a novel automated agent for the improvement of the communication process between a driver and its automotive climate control system. Through an extensive human study with 24 drivers we show that our agent improves the communication process and increases drivers’ satisfaction.