Skip to Main Content
Multi-view semi-supervised learning methods exploit the combination of multiple data views and unlabeled data in order to learn better predictive functions with limited labeled data. However, their applicability is limited since typically one data view is readily available but additional views may be costly to obtain. Here we explore a new research direction at the intersection of active learning and multi-view semi-supervised learning: active view completion. The goal is to actively select which instances to obtain missing view data for, for the purposes of enabling effective multi-view semi-supervised learning. Recent work has shown an active selection strategy for view completion can be more effective than a random one. Here a better understanding of active approaches is sought, and it is demonstrated that the effectiveness of an active selection strategy over a random one can depend on the relationship between views. We present new algorithms, theoretical results, and experimental study to elucidate the conditions for and extent to which active approaches can be beneficial in this scenario.