Increasing the Accuracy of NEWFM using a Geometric Graph-Based Gene Selection Algorithm

摘要

In a microarray dataset with thousands of genes, the majority of the genes are irrelevant to the occurrence of disease. One of the challenges in analyzing microarray datasets is selecting a suitable number of the most relevant genes with maximum classification accuracy. A neural network with weighted fuzzy membership functions (NEWFM) is a supervised neuro-fuzzy system that reduces the size of available data to improve classification accuracy and computational efficiency. This paper proposes a new gene selection algorithm based on a geometric graph using the bounded sum of weighted fuzzy membership functions (BSWFM) that improves the accuracy of NEWFM. The proposed gene selection algorithm reduces computational load and improves accuracy by removing irrelevant genes using the Euclidean distances of the centers of gravity multiplied by the non-overlap area between two BSWFMs. Further, the results of comparative experiments conducted using the colon cancer and the leukemia microarray problem datasets indicate that NEWFM with the proposed gene selection algorithm is more accurate than NEWFM without the proposed algorithm. More specifically, 2000 genes from the colon cancer dataset and 7129 genes from the leukemia dataset used as inputs to NEWFM without the proposed algorithm resulted in accuracies of 90.3% and 58.8%, respectively. In contrast, inputs of seven minimum genes from the colon cancer dataset and nine minimum genes from the leukemia dataset to NEWFM with the proposed algorithm resulted in accuracies of 96.8% and 100%, respectively.