A Research Of Multi-class Clustering Problem Based On Maximum Margin And Binary Tree Structure Problem

Clustering algorithm, as an important technology for data analysis, has been widely used indata processing, text analysis, image search, etc. So it has received extensive attention. SupportVector Machine (SVM) is a new supervised machine learning algorithm and it has excellentclassfication and generalize ability. We promote the hyper-plan theory of SVM to clusteringalgorithm and propose a new clustering algorithm-maximum margin clustering algorithm. It ismainly according to the unsigned sample and want to make the cluster margin of the two clustermaximum through finding a group of optimal sample label. This algorithm has a good clusteringperformance. However, the maximum margin clustering has a high clustering efficiency justbased on a two class problem. In this paper, we propose this two class clustering method to amulti-class clustering algorithm from the structure of a binary tree. We divise a multi-classclustering into many two class clustering by the structure of a binary tree and finish thegeneralization of maximum margin clustering.In this paper, we propose a multi-class clustering algorithm through the structure of a binarytree based on maximum margin clustering. Firstly, we general introduce the traditional clusteringalgorithms and propose the existing problems of these algorithms. Meanwhile, a in-depthanalysis of SVM theory has been conducted and the concept of maximum margin has been putforward. The SVM good supervised classification performance has been discussed and weextend it to the unsupervised clustering. Secondly, we make a detail account for the specificprinciple and implementation of the maximum margin algorithm. Then we propose an improvedalgorithm through analysing the shortage of this algorithm and extend this improved algorithminto a multi-class clustering through a structure of a binary tree. A detail introduce of the stepsand algorithm process for this algorithm has been illustrated. Finally, we show the feasibility ofthe algorithm through the experimental results and find this algorithm has lower clustering theerror rate by comparing the clustering the error with the traditional k-means clustering clustering.