GPU Acceleration of Clustering Meta-feature Extraction using RAPIDS

Abstract

Although machine learning algorithms have been successful when applied to several tasks, the selection of the most suitable for a given dataset is not straightforward. The recommendation of machine learning algorithms can be automated through the use of meta-learning, but this requires efficient methods for the characterizations of datasets, i.e. meta-features extraction. In this work we propose to accelerate the extraction of clustering-based meta-features on GPUs, taking advantage of the optimized libraries and API from the RAPIDS framework. We parallelized a well-known meta-feature extraction tool (MFE) via RAPIDS to accelerate the clustering meta-features extraction process. Our experiment shows that significantly less time is required to complete the extraction, up to 10x faster than the MFE implementation. These results are promising and suggest greater feasibility for large-scale experiments involving meta-learning.