Hostility measure for multi-level study of data complexity

Abstract Complexity measures aim to characterize the underlying complexity of supervised data. These tackle factors hindering performance Machine Learning ( ML ) classifiers like overlap, density, linearity, etc. The state-of-the-art has mainly focused on dataset perspective complexity, i.e., offering an estimation whole dataset. Recently, instance also been addressed. In this paper, hostility measure , a multi-level (instance, class, and dataset) data is proposed. proposal built by estimating novel notion : difficulty correctly classifying point, or given their corresponding neighborhoods. proposed estimated at level applying k -means algorithm in recursive hierarchical way, which allows analyze how points from different classes are naturally grouped together across partitions. information aggregated provide knowledge class levels. validity evaluated through variety experiments dealing with three perspectives comparative measures. Throughout experiments, shown promising results be competitive, stable, robust.