Cancer classification using collaborative representation classifier based on non-convex lp-norm and novel decision rule

Sparse representation classification (SRC) and collaborative representation classification (CRC) are the most promising classifiers for classifying high dimensional data. However, they may suffer from outliers and noises, as l₂-norm on signal fidelity is not effective enough to represent the test sample in that case. Recent studies show that non-convex l_p-norm minimization can boost the performance of classifiers compared with l₁- and l₂-norm minimization in classification. In this paper, we present an improved collaborative representation classification method for the accurate identification of cancer subtype. We improve CRC method by adopting non-convex l_p-norm on the signal fidelity term and introducing a new classification decision rule. We compute the coding coefficients over training samples for test sample via generalized iterated shrinkage algorithm (GISA) and classify the test sample into the subclass which has the maximum sum of coefficient (SoC). Extensive experiments on eight publicly available gene expression profile (GEP) datasets demonstrate the superiority of our proposed method.