An Enhanced K-means Algorithm Based On Shrinkage Estimation

Clustering is a method to study the similarity of data,which is widely used in many scientific fields such as statistics,meteorology,medicine,etc.This thesis studies the clustering algorithm from the perspective of improving forecasting ability and proposes an enhanced shrinkage K-means algorithm,which is a new clustering method based on James-Stein shrinkage estimation and learning vector quantization(LVQ).This new algorithm mainly considers the advantages of unsupervised clustering and supervised classification,in each iteration,we first obtain a temporal label for each data point by using the K-means algorithm,with such labeled data,we employ LVQ algorithm to get a prototype vector.Second,we shrink the clustering centers obtained in the previous stage toward the prototype vector via James-Stein estimator.These shrinkage centers thus return back to be the new clustering centers for the next iteration.The algorithm conduct “K-means-to-LVQ-Shrinkage” iterative procedures until the stop condition is achieved.This thesis also carry out extensive simulation studies and real data analysis to evaluate the performance of this new approach,and obtain encouraging results.