Learning patterns of states from multi-channel time series using genetic programming

Abstract

A state in time series is time series data stream maintaining a certain pattern over a period of time, for example, holding a steady value, being above a certain threshold and oscillating regularly. Automatic learning and discovery of these patterns of time series states can be useful in a range of scenarios of monitoring and classifying stream data, for example, activity recognition based on body sensor readings. In this study, we present our genetic programming (GP)-based time series analysis method on learning various types of states from multi-channel data streams. This evolutionary learning method can handle relatively complex scenarios using only raw input. This method does not require prior knowledge of the relationships between channels. It does not require manually defined feature to be constructed. The evaluation using both artificial and real-world multi-channel time series data shows that this method on raw input can outperform classic learning methods on pre-defined features. The analysis shows patterns can be discovered by the GP method.