An Ensemble System with Random Projection and Dynamic Ensemble Selection

In this paper, we propose using dynamic ensemble selection (DES) method on ensemble generated based on random projection. We first construct the homogeneous ensemble in which a set of base classifier is obtained by a learning algorithm on different training schemes generated by projecting the original training set to lower dimensional down spaces. We then develop a DES method on those base classifiers so that a subset of base classifiers is selected to predict label for each test sample. Here competence of a classifier is evaluated based on its prediction results on the test sample’s \( k - \) nearest neighbors obtaining from the projected data of validation set. Our proposed method, therefore, gains the benefits not only from the random projection in dimensionality reduction and diverse training schemes generation but also from DES method in choosing an appropriate subset of base classifiers for each test sample. The experiments conducted on some datasets selected from four different sources indicate that our framework is better than many state-of-the-art DES methods concerning to classification accuracy.