Gray Forest: A Modified Hybrid Cost-sensitive Feature Selection in DNA Microarrays

DNA Microarray has paved the way for cancer, tumor, and other diseases classification methods. Major challenges with DNA microarrays are the “curse of dimensionality” and “curse of sparsity” which lead to computational instability. Feature selection techniques are widely employed by researchers to tackle against those challenges and identify disease biomarkers. However, available feature selection techniques ignore (1) the rare variance embedded in DNA microarrays which could result in biased technique and (2) features' costs (i.e., gene cost) which could result in high-cost biomarkers. This study proposes a new algorithm, called Gray Forest; a modified hybrid cost-sensitive algorithm for feature selection and was inspired by the G-Forest algorithm [1]. Gray Forest produces a population of features for Random Forest and performs a simultaneous feature selection in a cost-sensitive manner by assigning a selection probability that is inversely related to the feature cost. To advance the search efficiency of the Gray Forest, high-level stochastic operators of (1) randomness (2) local search and (3) competition among offsprings were included. The Gray Forest outperformed multiple state-of-the-art algorithms and demonstrated efficacy and robustness in selecting the best features subset. The Gray-Forest improved accuracy up to 12% and decreased costs up to 15% when compared with the other approaches.