In this paper, we propose a robust MVL method (RML) that simultaneously performs representation fusion and alignment to overcome the limitations of multi-view learning (MVL) that integrates various types of data. RML uses a multi-view transformer fusion network to transform heterogeneous multi-view data into homogeneous word embeddings and obtains fused representations through a sample-level attention mechanism. In addition, we propose a multi-view contrastive learning framework that utilizes simulation-based perturbation to simulate incomplete data conditions, and aligns two fused representations obtained from noisy data and unavailable data through contrastive learning to learn discriminative and robust representations. RML is a self-supervised learning method and can be used as a plug-and-play module for multi-view unsupervised clustering, noisy label classification, and cross-modal hashing search. Experimental results verify the effectiveness of RML.
