Improved Filter Ranking Incorporated Binary Black Hole Algorithm for Feature Selection

Feature selection is a very important preprocessing step in all machine learning tasks. Selection of the most informative attributes increases algorithm performance, provides valuable domain information, reduces noise and reduces computational requirements. Recently, the Black Hole metaheuristic algorithm inspired by the real-life behavior of stars and the black hole has been developed in the literature. This algorithm is increasingly employed for solving several real-life problems. We have modified the existing binary Black Hole algorithm by incorporating filter ranking for improving algorithm performance. We have introduced two different combinations of filters: these are Pearson correlation and mean decrease in Gini combination and, Pearson correlation and mutual information combination. In our implementation we probabilistically switched between the standalone BH algorithm fitness criteria and the filter ranking enabled BH algorithm fitness criteria during the progress of the algorithm. We ran simulations with benchmark datasets covering different fields of science and engineering. We have compared our results with the standalone Black Hole algorithm. Our new algorithms exhibit improved performance in terms of selected subset size and accuracy. We also compared our algorithm with the filter ranking-incorporated ant colony algorithm available in the literature. We found that our algorithm compares very well with this algorithm. Our results indicate that the probabilistic incorporation of filter combinations enhances performance considerably. Various other synergistic filter combinations can also be incorporated in future for performance improvement.