Joint Meta-Sample Extraction and Classifier Learning for Tumor Classification

Sparse representation based classification (SRC) has been extensively and successfully applied to tumor classification due to its effectiveness in classification. Different from many existing SRC methods which emphasize on strong representational ability, our proposed method takes the discriminative ability into account, i.e. joint meta-sample extraction and classifier learning, which may be favor to better performance. In addition, a novel decision rule, in relation to both the minimal reconstruction residual and the minimum classification error, is presented and used during the classification stage. Experiments using several publicly available microarray gene expression data demonstrate that the proposed algorithm is efficient and achieves comparable accuracies with the state-of-the-art sparse coding techniques for tumor classification.