Study Of Supervised Clustering Neural Networks

Clustering is a fundamental technology in Artificial Intelligence. In terms of ambiguity of training samples, clustering algorithms fall into three categories: unsupervised, semi-supervised and supervised. Traditional unsupervised clustering is simple but not so effective because it directly uses training data without utilizing label information. Nevertheless, supervised clustering improves its performance through labeled data.Radial basis function neural network (RBFN) is a kind of Artificial Neural Network with simple structure, good generalization ability and fast speed and is often used for application of classification, regression and time series prediction. Traditional RBFN's training involves unsupervised clustering. In this paper, we introduce supervised clustering into RBFN and create a novel supervised fuzzy clustering network based on linear regression model which shows better results in regression task. Traditional RBN regression modeling use all the training data as a whole and is thus called global modeling which is incapable of interpreting local behavior of the estimated models. An approach called the local modeling was proposed in the literature to cope with this problem; however existing local modeling methods have the problems of boundary effects as well as low speed. In this paper, we propose a novel local modeling technology based on fuzzy partition and supervised clustering which uses different algorithms according to each subset's modeling difficulty respectively and thus overcomes its antecessors' shortcomings.This paper is organized as follows: first it introduces some background knowledge about clustering and RBFN, then describes a supervised clustering algorithm based on linear regression model, finally it talks about local modeling and fuzzy partition. Experiments results are given as a proof of the improved precision and performance of proposed methods.