An Experimental Evaluating Method For Clustering Ensemble

In data analysis, clustering analysis is considered as one of the most effective tools. In the last two decades, numerous outstanding clustering algorithms are proposed. It is widely applied on various domains such as data mining, pattern recognition, machine learning, information retrieval, image analysis and bioniformatics. Distrinct algorithms possess their own merits. However, not one of clustering algorithms can be adapted to all the circumstances. Clustering ensemble is believed as an algorithm that can extract the merits from distrinct algorithms, it combines the decision from base clustering algorithms and obtains the outcome that can not be stretched by any base clustering algorithm.Clustering ensemble algorithm is the hot research spot in clustering algorithm. Nowadays, many clustering ensemble algorithms are presented. Nevertheless, besides developing a decent algorithm, how to measure clustering quality is also a tough task. According to the condition that has the benchmark or not, the methods of evaluating clustering ensemble can be divided by intrinsic method and extrinsic method. Intrinsic method does not utilize benchmark, it is on basis of the definition, observing the interior structure of the data set. Intrinsic method need user to decide the distance metric, but usually the user does not have the this kind of knowledge. Extrinsic method utilizes benchmark, investigating the similarity of the clustering outcome and benchmark. Traditional extrinsic method utilizes ground truth evaluated by experts as benchmark. Nevertheless, in real-world, this genre of benchmark is not only expensive, but also rare to be obtained.This paper proposes an extrinsic method that does not use benchmark evaluated by experts. It is on the ground of the indistinct definition of benchmark, asserting that the benchmark evaluated by experts is just one of the benchmarks, benchmark can be based on all the clustering ensemble algorithms. Based on this fact, a state-of-art experimental evaluating clustering ensemble method is proposed, whose benchmark is compromised by all clustering ensemble methods. According to the framework of this method. this method is utilized to compare single-linkage clustering(SLC), iterative voting clustering(IVC) and an algebraic approach to clustering ensemble(AA)(SLC and IVC. SLC and AA, IVC and AA) on 2 synthetic and 3 UCI data sets, and draw a comparison between the results and the one from traditional extrinsic method. From the view point of traditional extrinsic method, if the clustering ensemble algorithms are strong clusterings, then this method’s result shows no difference compared with the traditional one. With the metric of not using benchmark evaluated by experts, the method possesses a prospect of wide application.