Machine Learning and Knowledge Discovery

Abstract

In many domains there will exist different representations or “views” describing the same set of objects. Taken alone, these views will often be deficient or incomplete. Therefore a key problem for exploratory data analysis is the integration of multiple views to discover the underlying structures in a domain. This problem is made more difficult when disagreement exists between views. We introduce a new unsupervised algorithm for combining information from related views, using a late integration strategy. Combination is performed by applying an approach based on matrix factorization to group related clusters produced on individual views. This yields a projection of the original clusters in the form of a new set of “meta-clusters” covering the entire domain. We also provide a novel model selection strategy for identifying the correct number of meta-clusters. Evaluations performed on a number of multi-view text clustering problems demonstrate the effectiveness of the algorithm.