AutoTSC: Optimization Algorithm to Automatically Solve the Time Series Classification Problem

Nowadays Automated Machine Learning, abbreviated AutoML, is recognized as a good solution to quickly find a model without spending too much time on the tedious task of selecting an algorithm and its associated hyperparameters. AutoML is well studied on traditional classification problems but has never been explored on Time Series Classification (TSC) problems. Yet, we show in this article that this problem is different enough to require specific approaches. Indeed, while the preprocessing phases need to be defined with classical ML algorithms, most of them are embedded within the algorithms for TSC. This particularity is mainly due to the fact that we are dealing with ordered data. Over the last decade, there has been a lot of interest for this problem and many new ML algorithms oriented toward TSC (ML-TSC) have blossomed. A practitioner facing a TSC use case, and searching for the most adequate algorithm now has a wide range of options, and will generally end up trying a few algorithms with standard hyperparameters, similarly to what happens in classical ML. Clearly, this is suboptimal. In this article we propose an AutoML solution called AutoTSC based on Evolutionary Algorithms (EAs) and its associated ML-TSC search space. AutoTSC, when tested against standard datasets from the UCR archive, outperforms both a Random Search on the proposed search space and TPOT a well known standard AutoML tool.