Inductive Multi-view Multiple Clusterings

The typical clustering algorithms concern only one clustering result, which ignore the other alternative patterns of intricate real world data. As a result, approaches that exploit diverse clusterings in the same dataset have been studied in the past decade. Most previous multiple clusterings approaches are developed for single view datasets. Existing multi-view approaches address multiple clusterings problem in a transductive manner, which can not be generalized to new data. To solve these issues, this paper presents an inductive multi-view multiple clusterings (IMVMC) framework. Specifically, IMVMC aligns the heterogeneous views at first, then generates multiple common subspaces by optimizing a group of encoder deep networks. Finally, it introduces a diversity control term to minimize the redundancy among these subspaces. For purpose of improving the robustness of the model, IMVMC further introduce a regular term to constrain model parameters. Experiments on benchmark datasets verify the effectiveness of IMVMC in capturing multiple clusterings of high diversity and quality.