Feature Selection for High-Dimensional Gene Expression Data: A Review

In recent decades, dealing with high-dimensional data has become an undeniable challenge in most data mining applications. In certain domains, such as bio-informatics, and specifically in microarray data analysis, exploring gene expression data often involves the use of tens of thousands of features (genes) measured across just a few dozen samples. Such scenarios, make the use of classical data mining tools a real challenge due to the involvement of a significant number of irrelevant or redundant genes. In response to this challenge, several approaches-based feature selection have been proposed, each with its advantages and disadvantages. This work introduces a classification of feature selection methods and also reviews the state-of-the-art approaches developed over the past five years. Our review has revealed a notable trend towards hybrid approaches, approximately 50% of the surveyed studies propose hybrid feature selection techniques, most frequently combining filter with wrapper methods. Additionally, the 10-fold cross-validation technique stands out as the dominant evaluation method, employed by 61.6% of surveyed approaches. Support Vector Machines emerge as the most favored classification algorithm, demonstrating optimal performance in 77.78% of cases. These findings contribute to the advancement of feature selection approaches, particularly in reducing the dimensionality of gene expression data, thereby enhancing cancer classification methodologies.