EHC: Non-parametric Editing by Finding Homogeneous Clusters

Abstract

Editing is a crucial data mining task in the context of k-Nearest Neighbor classification. Its purpose is to improve classification accuracy by improving the quality of training datasets. To obtain such datasets, editing algorithms try to remove noisy and mislabeled data as well as smooth the decision boundaries between the discrete classes. In this paper, a new fast and non-parametric editing algorithm is proposed. It is called Editing through Homogeneous Clusters (EHC) and is based on an iterative execution of a clustering procedure that forms clusters containing items of a specific class only. Contrary to other editing approaches, EHC is independent of input (tuning) parameters. The performance of EHC is experimentally compared to three state-of-the-art editing algorithms on ten datasets. The results show that EHC is faster than its competitors and achieves high classification accuracy.