Reaserch On Overlapping Clustering For Multi-view Data

In the wake of the new form of Internet Plus,data has appeared around all aspects of our daily life.How to discover useful information from these ubiquitous data is the research content of Data Mining.Cluster Analysis is a technique detecting the structure or pattern hidden in data without any prior information.With the advent of Big Data era,the research object of clustering,data,has undergone great changes on quality and quantity.Based on the needs of practical application,some new branches have arisen in the field of cluster analysis,such as clustering ensemble,semi-supervised clustering,overlapping clustering,multi-view clustering and etc.Overlapping clustering works on the hypothesis that one object belongs to one or more clusters.It can discover overlapping information hidden in observed data,which plays an important role as a bridge or pivot in practical application and has great analysis value.Multi-view data describes one observation from different perspectives or different information sources.And it forms attribute split datasets,which is Multi-view clustering to study.Most overlapping clustering methods dedicate to studying the strategy of discovering overlapping observations,and ignoring the correlation distinction of overlapping observation and different clusters it belongs to.Aiming at this problem,an Overlapping Clustering approach with Correlation weight is proposed in this thesis,which takes the correlation weights of observation and different clusters into consideration and thus improves the effectiveness of overlapping partition.Experiments on overlapping datasets include multi-label learning,movie recommendation and others demonstrates that the proposed algorithm improves the performance of overlapping clustering,compared with other existing overlapping algorithms.As overlapping information appears in some multi-view datasets as well,this thesis presents an overlapping clustering approach for multi-view data on the foundation of our overlapping algorithm.Here,we hold that cluster structures in different views may be different,because overlapping objects exist not only inside of some view but also between different views.The sum of sample loss function and clustering structure loss function is served as the criterion in this algorithm.It achieves co-training of multiple views and naturally ensemble the multiple clustering results by importing the final consensus cluster structure,thus gaining a consistency clustering results.Experiments on real datasets and artificial datasets prove the proposed overlapping algorithm for multi-view data can discover the latent structure in multi-view data,and have the good convergence property.