Radial basis function neural networks for datasets with missing values

Abstract

Radial Basis Function Neural Networks (RBFNN) are among the most popular supervised learning methods and showed significant results in various applications. Despite is applicability, RBFNNs basic formulation can not handle datasets with missing attributes. Aiming to overcome this problem, in this work, the RBFNN is modified to deal with missing data. For that, the expected squared distance approach is used to compute the RBF Kernel. The proposed approach showed promising results when compared to standard missing data strategies.