Comparative Analysis of Recursive Feature Elimination and Feature Importance for Dimension Reduction in Cancer Prediction

With nearly 10 million deaths in 2020, cancer remains one of the most common causes of death worldwide. In research, microarray techniques have been used to diagnose and predict cancer by simultaneously analyzing gene expression. However, microarray data is characterized by high dimensionality and the presence of many irrelevant genes, necessitating effective feature selection methods. This study explores two major methods—Recursive Feature Elimination (RFE) and Feature Importance (FI)—and evaluates their individual and combined performance through five hybrid strategies: score averaging, intersection, union, RFE→FI, and FI→RFE. Implemented within a Support Vector Machine (SVM) framework, these methods are tested on four publicly available microarray datasets: Lung, Colon, Leukemia, and Ovarian. While hybrid selection has shown advantages over single techniques, the impact of method sequencing has been underexplored. This research finds that the RFE→FI strategy consistently balances relevance and generalizability, leading to superior performance across multiple datasets. To improve generalization and reduce complexity, each dataset was limited to a maximum of 50 selected features. Performance was assessed using accuracy, precision, recall, and F1-score. The results reveal that sequential hybrid approaches—especially RFE→FI—enhance classification performance in complex, high-dimensional genomic data. This study contributes a robust and reproducible hybrid feature selection pipeline, offering practical value for microarray-based cancer prediction systems.