PaZoe: classifying time series with few labels

Mikhail Kamalov, Aurélie Boisbunon, Carlo Fanara, Ingrid Grenet, Jonathan Daeden

Abstract: Semi-Supervised Learning (SSL) on graph-based datasets is a rapidly growing area of research, but its application to time series is difficult due to the time dimension. We propose a flexible SSL framework based on the stacking of PageRank, PCA and Zoetrope Genetic Programming algorithms into a novel framework: PaZoe. This self-labelling framework shows that graph-based and non-graph based algorithms jointly improve the quality of predictions and outperform each component taken alone. We also show that PaZoe outperforms state-of-the-art SSL algorithms on three time series datasets close to real world conditions. A first set was generated in house, taking data from industrial graded equipment in order to mimick DC motors during operation. Two other datasets, which include the recording of gestures, were taken from the public domain.