A novel LASSO-based feature weighting selection method for microarray data classification

The biological data, especially microarray data has become more and more important for medical diagnostics. Microarray data usually has high dimension with small sample size and the positive samples are scarce, which makes the data severely imbalanced. In this paper, we propose a new feature selection model for high dimensions and imbalance data. Firstly, by computing the feature-label correlation we remove the low-score features which we consider are irrelevant. Then, a combined feature selection is employed in which the first stage is to remove the redundant features by using a correlation-based feature selection and then we propose a LASSO-based feature weighting approach to increase the weight of the key data. Finally, considering the imbalance problem, we process the selected data to obtain the balanced one by Synthetic Minority Over-sampling Technique (SMOTE), which will be used before classifier training. We use the Ten-fold cross-validationon the datasets. The experimental results on different classifiers show that the proposed method can achieve a higher accuracy and Area Under Curve (AUC) than the traditional feature selection methods.