Edinburgh Napier University

  Meta-learning, a concept from the area of automated machine learning, aims at providing decision support for data scientists by recommending a suitable setting (a machine learning algorithm or its hyper-parameters) to be used for a given dataset. Such a recommendation is based the assumption that an optimal setting for a certain dataset would also be suitable for other, similar datasets. Similarity of datasets is computed from their characteristics, named meta-features, several types of which have been developed thus far. This paper introduces a novel perspective on PCA meta-features which, despite their good descriptive characteristics and easy computation, are rarely used in meta-learning. A novel meta-learning approach utilizing DTW, a well-known similarity measure for time-series, is proposed for computing dataset similarities based on the series of cumulative variances explained by their respective principal components. The results from a large-scale experiment, comparing the proposed approach to multiple baselines on 50 real-world datasets, show the potential of combining PCA and DTW in meta-learning and encourage further investigation in this direction.