Dynamic Ensemble Algorithm Post-Selection Using Hardness-Aware Oracle

Dynamic Ensemble Selection (DES) algorithms have obtained better performance in many tasks compared to monolithic classifiers and static ensembles. However, it is reasonable to assume that no DES algorithm is the optimal solution in different scenarios since diversity plays an important role. Thus, this paper addresses this research gap by proposing a novel approach called Hardness-aware Oracle with Dynamic Ensemble Selection (HaO-DES) that operates as a post-selection strategy, evaluating and selecting the best DES techniques per instance. Each DES technique ensemble is evaluated using a new measure called Hardness-aware Oracle (HaO). HaO extends the traditional Oracle concept by assessing a DES technique based on how the classifiers in the selected ensemble work together, contrasting with the individual classifier evaluation in the traditional assessment. We performed experiments over 30 databases, using three base classifiers (Perceptron, Logistic Regression, and Naive Bayes) in homogeneous and heterogenous pools' configurations, to assess HaO-DES with four DES approaches (KNORA-U, KNOP, DES-P, and META-DES). We use three performance metrics to evaluate the experiments: accuracy, F-score, and Matthews Correlation Coefficient (MCC). The results show that our approach outperforms or obtains similar results against the four individual DES approaches, mainly when considering heterogeneous pool settings. We also demonstrated the HaO-DES efficiency in choosing suitable DES techniques in different situations.