Multi-level correlation learning for multi-view unsupervised feature selection

Recently, unsupervised multi-view feature selection has received a lot of interest. However, current methods facilitate feature selection by preserving the local consistency of multi-view data, which is defined based on similarities of data within each view, but ignore the global topological consistency in data, which is defined based on the cross-view topological similarities of data between views and is essential for revealing the distribution of multi-view data. In light of this, this paper proposes a novel multi-view unsupervised feature selection method with multi-level correlation learning, termed Multi-Level Correlation Learning for Multi-View Unsupervised Feature Selection (MLCL). It simultaneously derives the global topological correlation structure from the cross-view topological similarities of data and the local geometric correlation structure from the local similarities of data within each view, to take advantage of both global and local consistencies of multi view data. An effective optimization algorithm is then developed to resolve the optimization problem for the proposed model. Extensive experiments on eight publicly available datasets show that the proposed MLCL outperforms several state-of-the-art unsupervised multi-view feature selection models.

Automated summary

This paper proposes a multi-level correlation learning method for unsupervised feature selection in multi-view data, which takes into account both the global topological consistency and the local geometric consistency, and outperforms other methods according to extensive experiments.