Incomplete multi-view learning: Review, analysis, and prospects

Multi-view data, stemming from diverse information sources, often suffer from incompleteness due to various factors such as equipment failure and data transmission issues. This challenge has given rise to the emerging field of incomplete multi-view learning (IML). To provide guidance for newcomers and researchers in this field, this survey systematically presents an in-depth analysis of IML from generative and discriminative perspectives, focusing on all missing scenarios and various learning tasks. Within these categories, discriminative methods are further classified into matrix learning-based IML and graph learning-based IML, while generative methods encompass generative adversarial networks-based IML, auto-encoder-based IML, and contrastive learning-based IML. Meanwhile, practical applications across various domains are summarized, with extensions of IML to multiple labels as well as unaligned views. To advance this field, we conclude that adapting multi-view learning for incomplete data, addressing complex and arbitrary missing scenarios, tackling high missing ratios, exploring regularization techniques, reducing noise impact, and integrating IML with other learning paradigms are valuable research directions in the future.