Improved K-Means Clustering for Initial Center Selection in Training Radial Basis Function Networks.

Radial Basis Function networks accuracies mainly affected by its center selection from dataset. K-means (KM) clustering is a widely in numerous field for data classification and centers selection. However, initial centers selection poses high impact on KM clustering outcome. It suffers from its immense reliance on the initial centers selection algorithm from the dataset. KM algorithm has been enhanced for its performance from diverse perspectives over the years. Nonetheless, a good balance between quality and efficiency of the centers selected by the algorithm is not attained. To overcome this issue, this paper proposed an improvement on KM clustering algorithm in getting initial centers and reduce its sensitivity to initial centers. This paper introduce the use of improved K-means (KM) clustering that consider the each point distance as probability for selecting the initial centers with radial basis function network (RBFN) training algorithm. The proposed approach uses improved KM for centers selection in RBFN training algorithm shows accuracy improvement in predictions and with simpler network architecture compared to the conventional RBFN. The proposed network called IKM-RBFN was tested against the conventional RBFN, KM-RBFN, back-propagation neural network and long short-term memory neural network in FOREX EURUSD pair price predictions. The results are compared to proposed method on its root mean square error (RMSE) and mean absolute error (MAE) results. The proposed method shows promising results in improving RMSE accuracy over 20% in compared to other tested networks.