Adaptive sampling using self-paced learning for imbalanced cancer data pre-diagnosis

The early diagnosis of cancer diseases is an indispensable part in the cancer research. It urges people to develop many new machine learning approaches to assist the diseases identification based on the gene expression data. However, the race occurrence of malignant tumors creates a challenge due to the potential over-fitting risk in the current model training. Typically, people use various sampling methods (e.g., random oversampling and undersampling) to address this challenge to provide a balanced data distribution. However, these methods might discard potentially useful samples. In this paper, we proposed an imbalanced sampling approach via self-paced learning (ISPL) to effectively select high-quality samples to improve the robustness. The experimental results showed that our proposed ISPL method increased the classification accuracy by approximately 16% compared with the average performance obtained by other sampling methods. In addition, the new method successfully selected some important genes for further investigation.