Integrating Multi-view Data: A Hypergraph Based Approach

Abstract

In this paper, we have proposed a novel framework to integrate multiple views of data using hypergraph based partitioning approach. Representing each view as separate graphs, we partitioned them into clusters using standard clustering technique. Each clustering solution is represented as a separate hypergraph in which a hyperedge encodes a cluster. Concatenating all these hyperedges result to an adjacency matrix (H) of a hypergraph with n vertices which is equal to the number of objects in each data, and E edges which is equal to total number of clusters in the representative clustering solutions. A similarity matrix (S) is compiled from the adjacency (H) by taking entry wise average. Each entry of S merely denotes the fraction of clustering in which two objects are members of the same group/cluster. Clustering the S matrix results in meta-clusters which essentially inherit the characteristics of all data views in the original data. We have performed a simulation study on three data views to validate the framework. For each view, we tune the parameter of data, and investigate whether the results conform the changes. We have also applied the proposed framework in a real life dataset to identify meta-clusters. Moreover, we have analyzed the resulting meta-clusters to validate our proposed method.