Simultaneous Informative Gene Extraction and Cancer Classification Using ACO-AntMiner and ACO-Random Forests

Microarray cancer gene expression datasets consist of high dimensional data. Gene selection helps in the removal of irrelevant genes. The reduced dimensions of the datasets help in improving the overall classification performance. We present two hybrid techniques, Ant Colony Optimization-AntMiner (ACO-AM) and ACO-RandomForests (ACO-RF) with weighted gene ranking as heuristics. The heuristic information is obtained by a weighted sum of the Information Gain, Chi-Square, Correlation based Feature Selection (CFS) and Gini Index scores for each gene. The ACO algorithm selects a small subset of relevant genes from this ranking. The fitness’s of these subsets are then assessed by the cAnt-Miner and the Random Forest classifiers. The performances of the algorithms are tested using two cancer gene expression datasets retrieved from the Kent Ridge Bio-medical Dataset Repository. We demonstrate that genes selected by the suggested algorithms yield better classification accuracies.