A Hierarchical Dissimilarity Metric for Automated Machine Learning Pipelines, and Visualizing Search Behaviour

Angus Kenny, Tapabrata Ray, Steffen Limmer, Hemant Kumar Singh, Tobias Rodemann, Markus Olhofer

Abstract: In this study, the challenge of developing a dissimilarity metric for machine learning pipeline optimization is addressed. Traditional approaches, limited by simplified operator sets and pipeline structures, fail to address the full complexity of this task. Two novel metrics are proposed for measuring structural, and hyperparameter, dissimilarity in the decision space. A hierarchical approach is employed to integrate these metrics, prioritizing structural over hyperparameter differences. The Tree-based Pipeline Optimization Tool (TPOT) is utilized as the primary automated machine learning framework, applied on the abalone dataset. Novel visual representations of TPOT’s search dynamics are also proposed, providing some deeper insights into its behaviour and evolutionary trajectories, under different search conditions. The effects of altering the population selection mechanism and reducing population size are explored, highlighting the enhanced understanding these methods provide in automated machine learning pipeline optimization.

External IDs:dblp:conf/evoapps/KennyRLSRO24