Convex Principal Feature Selection.

摘要: A popular approach for dimensionality reduction and data analysis is principal component (PCA). limiting factor with PCA that it does not inform us on which of the original features are important. There a recent interest in sparse (SPCA). By applying an L1 regularizer to PCA, transformation achieved. However, true feature selection may be achieved as non-sparse coefficients distributed over several features. Feature NP-hard combinatorial optimization problem. This paper relaxes re-formulates problem convex continuous minimizes mean-squared-reconstruction error (a criterion optimized by PCA) considers redundancy into account (an important property selection). We call this new method Convex Principal Selection (CPFS). Experiments show CPFS performed better than SPCA selecting maximize variance or minimize mean-squaredreconstruction error.