Multi-Granulation Fusion-Driven Method for Many-View Classification

Multi-view classification is an important research issue in multi-view machine learning. Existing studies indicate that the effective fusion operators play a key role for performance improvement of multi-view classification tasks. Hence, one of the hot topics for multi-view classification research is to design effective fusion operators. As the number of views increases (task with more than three views, named many-view task), the existing fusion operators confront two problems: 1) the dimension of the fused vector obtained by using the fusion operators with stronger expressiveness ability exponentially increases with the increasing number of views; while the expressive ability of the fusion operators having nothing to do with the number of views is weaker; 2) currently, most of these methods use a fusion operator to directly fuse all view features and obtain the final fused vector. When there are many views, it is difficult to model relationship among views via the strategy. To address these issues, inspired by multigranulation, this article proposes a multi-granulation fusion method (MGF) for multi-view classification. MGF hierarchically models relation among view features from three level of granularities. Specifically, firstly, it uses a fusion operator to model the relation between pair of two views at the first level of granularity, then uses a fusion operator to model the relation between a view and other views at the second level of granularity, and finally a fusion operator is used to model the relation among all views at the third level of granularity. The experimental results conducted on four large-scale multi-view datasets demonstrate that our proposed method can statistically achieve better performance than the compared methods.