Research On Clustering Ensembles And Diversity

Clustering ensemble is an important part of ensemble learning, which is aimed at improving the results in unsupervised clustering analysis that are otherwise affected by the special data distribution of the samples and mismatching with assumptions. Clustering ensemble achieves better results by combining multiple clustering results into a single consensus clustering.In clustering ensemble, accuracy and diversity of individual member clusterings are the primary basis in selective ensemble. Accuracy is about a single learners’effect, while diversity is in case of many learners in multi-clusterings. If there is no diversity among the member clusterings, then clustering ensemble would become meaningless. On the other hand however, it also makes clustering ensemble complex if each member clustering seeks its own views. So the degree of diversity among member clusterings usually has a direct influence on the final ensemble clustering results.Diversity is a key factor in improving the performance of an ensemble clustering algorithm. However, it should be studied deeply in order to measure the diversity of clusters more effectively. The study about clustering ensemble diversity mainly aims at two main factors. Firstly, researchers want to find a proper measurement to reflect the diversity of clusterings ensemble. Secondly, the relationship between ensemble diversity and ensemble performance is studied to facilitate the construction of multiple clustering systems.The main contributions of this thesis may be discussed in three aspects. Firstly, in order to overcome problems in pair-wise diversity measurement, an improved algorithm, adjusting the ensemble diversity is proposed. Secondly, a clustering ensemble algorithm based data relevance (DRBCE), is presented. The basic idea is to extract related data objects to make up new classes, and then combine them to gain the final ensemble result. Experimental results in this thesis confirm the validity of the presented algorithm. Thirdly, the diversity fluctuation of internal member clusterings and the relationship between ensemble diversity and ensemble performance is analyzed by extracting different diversity clusterings.