Uncorrelated Semi-Paired Subspace Learning

Multi-view datasets are increasingly collected in many real-world applications, and existing multi-view learning methods by leveraging the relation information across multiple views have gained great successes comparing to conventional single-view learning methods. However, most of these methods are built on the assumption that each instance having all given views, so they are not able to integrate the unpaired data properly that can often be observed in reality. In this paper, we focus on learning uncorrelated features for semi-paired subspace learning as the usefulness of uncorrelated features have been witnessed by many existing works. Specifically, we propose a generalized uncorrelated multi-view subspace learning framework, which is able to naturally integrate various learning criteria derived from the semi-paired data. To showcase the flexibility of our framework for various learning scenarios, we instantiate new semi-paired models for both unsupervised learning and semi-supervised learning. We further propose a deflation-based optimization scheme to solve the resulting challenging optimization problem. An efficient and scalable Krylov subspace method is used to solve the subproblem with global solution, and the proposed algorithm guarantees monotonic convergence. Extensive experimental results on multi-view feature extraction and multi-modality classification show that our proposed models perform competitively to or better than baselines.