Learning sparse features for classification by mixture models

Non-negative matrix factorization (NMF) can discover sparse features for classification via mixture models and the sparseness of features controls the learning rate of the basis function parameters. But the original NMF in which the basis vectors are unit ones in L₁ norm, does not increase the sparseness of learned features. This paper generalizes NMF to L_p-NMF where the basis vectors are unit ones in L_p norm. Experiments demonstrate how p affects the sparseness of learned features and the final classification accuracy. And the results show that L₂-NMF is superior one for practical implementation.