Dictionary Learning Based on Laplacian Score in Sparse Coding

Sparse coding, which is represented a vector based on sparse linear combination of a dictionary, is widely applied on signal processing, data mining and neuroscience. How to get a proper dictionary is a problem, which is data dependent and computational cost. In this paper, we treat dictionary learning in the unsupervised learning view and proposed Laplacian score dictionary (LSD). This new method uses local geometry information to select atoms for dictionary. Comparison experiments with competitive clustering based dictionary learning methods are established. We also compare LSD with full-training-data-dictionary and others classic methods in the experiments. The results on binary classes datasets and multi class datasets from UCI repository demonstrate the effectiveness and efficiency of our method.