Ensemble Neural Networks with Novel Gene-Subsets for Multiclass Cancer Classification

Multiclass gene selection and classification of cancer are rapidly gaining attention in recent years, while conventional rank-based gene selection methods depend on predefined ideal marker genes that basically devised for binary classification. In this paper, we propose a novel gene selection method based on a gene’s local class discriminability, which does not require any ideal marker genes for multiclass classification. An ensemble classifier with multiple NNs is trained with the gene subsets. The Global Cancer Map (GCM) cancer dataset is used to verify the proposed method for comparisons with the conventional approaches.