Improved bio-geography based optimization algorithm for solving feature selection problems

Data dimensionality is the common problem in many pattern recognition applications. Specifically, this happens owing to the infeasibility of populating the feature space (sufficiently) with limited data. To overcome this issue, many real-time practices have adopted a technique called Data Reduction, which can be achieved by either Feature Transformation or by Feature Selection (FS). This FS becomes a tedious task, when performed in high dimensional datasets, due to the presence of large number of redundant features and training data limitations. Redundancies present in these irrelevant features affect the overall classification accuracy. These issues faced in high dimensional datasets can be overcome by introducing meta-heuristics during the FS process. These meta-heuristics are instrumental in solving the FS problem in an efficient manner. But, when FS is performed using these meta-heuristics, they might face certain limitations like premature convergence, low degree of solution accuracy, slow convergence, poor exploration and exploitation ability. These limitations can be overcome by modifying their basic algorithms, merging new operators, improving the existing operators. In this thesis, a newly proposed algorithm called Improved Biogeography Based Optimization (IBBO) algorithm has been introduced to perform an efficient wrapper based feature selection. Sine Cosine Algorithm (SCA) has been used to improve BBO algorithm, i.e. the position update mechanism of SCA is replaced with mutation operation of BBO algorithm. The efficiency of this algorithm is proved by testing with standard benchmark datasets. The RBF kernel of the ELM classifier has been used for feature classification. The experimental results are presented in terms of average classification accuracy, average fitness value, average number of selected features and average reduction rate of the features. A promising classification accuracy has been obtained for all the ten benchmark datasets i.e. 81% for 14_Tumors and 99% for 11_Tumors dataset, 87% for 9_Tumors and 99% for Prostrate Tumors dataset, 100% for Leukemia2 and 93% for Sonar dataset, 94% for LSVT dataset and 88% for LandCover dataset, 84% for Heart dataset, 85% for Lymphography dataset respectively. Moreover, a reduced feature subset has been attained for all these datasets through the IBBO algorithm. The efficiency of this algorithm is further verified through several experiments like Ranking methods, Convergence curves, Statistical analysis in terms of Box-plot analysis, Bonferroni’s Holm Correction and Friedman Test. Alongside the experimental results, the threats to validity have been discussed.