Unsupervised Cross-View Feature Selection on incomplete data

Abstract

Unsupervised multi-view feature selection (UMV-FS) deals with the dimension reduction problem wherein instances are unlabeled and represented by heterogeneous features. Existing mainstream UMV-FS methods incorporate instance-wise view interactions based on graphs to guide feature selection, in which within-view selection decisions are independently learned to piece up a global feature subset. However, this strategy induces a globally sub-optimal feature selection decision in the sense that unexpected redundant features across views proliferate. Furthermore, existing studies are performed in view-complete frameworks, which hardly satisfies real-world applications. To address these issues, we propose a novel cross-view feature selection (CVFS) framework in an unsupervised manner, which can process large-scale/streaming data. This is the first attempt to approach incomplete multi-view feature selection by devising and fusing two-wise view interactions. Specifically, we incorporate the traditional instance-wise view interactions based on graphs to find discriminative features in each view and model a novel kind of feature-wise view interactions to enforce selection diversity and reduce feature redundancy. These techniques can yield a globally optimal feature subset across all views. Comprehensive experiments validate the effectiveness and efficiency of the proposed CVFS.