Parallel Selection of Informative Genes for Classification

In this paper, we argue that existing gene selection methods are not effective for selecting important genes when the number of samples and the data dimensions grow sufficiently large. As a solution, we propose two approaches for parallel gene selections, both are based on the well known Relief F feature selection method. In the first design, denoted by P Relie f F-p, the input data are split into non-overlapping subsets assigned to cluster nodes. Each node carries out gene selection by using the Relie f F method on its own subset, without interaction with other clusters. The final ranking of the genes is generated by gathering the weight vectors from all nodes. In the second design, namely P Relief F-g, each node dynamically updates the global weight vectors so the gene selection results in one node can be used to boost the selection of the other nodes. Experimental results from real-world microarray expression data show that P Relief F-p and P Relief F-g achieve a speedup factor nearly equal to the number of nodes. When combined with several popular classification methods, the classifiers built from the genes selected from both methods have the same or even better accuracy than the genes selected from the original ReliefF method.