Classification of cancer patterns from gene expression data is a difficult task in computational biology and artificial intelligence due to the sufficient number of training samples is often difficult, expensive, and hard to gather. Although, the classification results obtained by the conventional classifiers trained with insufficient training samples are generally low. However, unlabeled samples are relatively low-cost and easy to gather, whereas conventional classifiers do not utilize these unlabeled samples to train the model. In this context, a self-training-based model semi-supervised ordered weighted average fuzzy-rough nearest neighbour classifier for cancer pattern classification from gene expression data is proposed. The experiments are carried out on eight publicly available real-life gene expression cancer datasets. The performance of the proposed method is compared with four other methods (two supervised and two semi-supervised) in terms of percentage accuracy, precision, recall, macro averaged F1 measure, micro averaged F1 measure and kappa. The dominance of the proposed method is justified by the experimental results.