Applying Principal Component Analysis to Optimize Feature Selection in Gene Expression Data: A Case Study on Cancer Classification

The most popular feature-selection method is PCA, and is often used for high-dimensional data such as gene expression data. This article presents the use of PCA to reduce dimensionality and classification of cancer. In this dataset (Cancer gene expression data from TCGA), we perform PCA to simplify the gene expression data but maintain the highest variance. Our models learn from complete feature sets and PCA reduced feature sets respectively with machine learning algorithms such as SVM and RF. We found that PCA yields robust improvements in classification accuracy, precision, recall and F1 score without overfitting or extra computational overhead. We also evaluate PCA with other feature selection approaches (correlation-based selection, Random Forest feature importance, and L1 regularization). These results suggest that PCA is an optimal solution for both performance and computational difficulty, and a very promising feature selection technique for cancer classification.