A Reduced Variable Neighborhood Search Approach for Feature Selection in Cancer Classification

In this work we propose a Reduced Variable Neighborhood Search (RVNS) algorithm, to handle the gene selection problem in cancer classification. RVNS is utilized as the search method and gene subsets obtained are evaluated by three learning algorithms, namely support vector machine, k-nearest neighbors, and random forest. Experiments are conducted on five publicly available cancer related datasets, all characterized by a small sample size to dimensionality ratio. Since RVNS seeks gene subsets that yield accurate predictions for all three aforementioned classifiers, the obtained results can be considered more reliable. To the best of our knowledge, the proposed methodology is innovative due to the fact that, it combines the Recursive Feature Elimination (RFE) heuristic with a RVNS algorithm. Despite the large size of the problem instances, the suggested feature selection scheme converges within reasonably short time, when compared to similar methods. Results indicate high performance for RVNS that, is further improved when the RFE method is applied as a pre-processing step.