Genetic Programming-based AutoML for EEG Signal Classification - A Comparative Study
I. M. Miranda, Claus C. Aranha, André C. P. L. F. de Carvalho, Luís P. F. Garcia · 2022
End-to-end Machine Learning (ML) applications using complex data often need to investigate several alternatives for the data modeling pipeline before a good solution is found. This process, which is time-consuming and subjective, can benefit from an automated solution design by using Automated Machine Learning (AutoML). End-toend AutoML allows automated data preparation, modeling, and evaluation of ML pipelines, increasing the chances of arriving at a good solution. AutoML can implement this optimization with different strategies. Among them, Genetic Programming (GP) stands out for its ability to create pipelines of arbitrary format, allowing high interpretability and the customization of information from the data context. This paper proposes and compares two approaches of end-to-end AutoML optimized with GP for a time series classification problem, the classification of Electroencephalogram (EEG) signals. We selected this dataset because of the signals’ high complexity, spatial and temporal co-variance, and nonstationarity. For the AutoML experiments, four different domain-based data characterization measures are evaluated. The analysis of the data characterization measures shows that using only spectral or time-domain features does not lead to pipelines with good predictive performance. Our experimental results also reveal how AutoML can generate more accurate and interpretable solutions than the literature’s complex and ad hoc models. The proposed approach makes it easier to analyze dimensional reduction through fitness convergence, tree depth, and extracted features.