Differences and Similarities Learning for Unsupervised Feature Selection

摘要

In this paper, a novel feature selection algorithm, named Feature Selection with Differences and Similarities (FSDS), is proposed. FSDS jointly exploits sample differences from global structure and similarities from local structure. To reduce the disturbance from noisy feature, a row-wise sparse constraint is also merged into the objective function. FSDS, then combines the underlying subspace features with original feature to construct a more reliable feature set. Furthermore, a joint version of FSDS (FSDS2) is introduced. To optimize the proposed two-step FSDS and the joint version FSDS2 we also design two efficient iterative algorithms. Experimental results on various datasets demonstrate the effectiveness of the proposed algorithms.