Neuroevolution based multi-objective algorithm for gene selection and microarray classification.

Microarrays allow the expression level analysis of thousands of genes simultaneously; thus, it is a common technique used for cancer detection and diagnosis. However, existing microarray datasets have huge data dimension and class imbalance, therefore, it is important to find relevant genes that accurately set classes apart and allow building more reliable classification models. A multi-objective algorithm is proposed to evolve artificial neural networks' topology and connection weights for microarray classification by minimizing the number of selected genes and the cross-entropy loss. The algorithm is based on the evolutionary multi-objective algorithm SMS-EMOA along with the genetic encoding and the crossover and mutation operators from the neuroevolution algorithms NEAT/N3O. Moreover, a speciation algorithm was implemented to protect the diversity of the selected features within the solutions. To test the algorithm performance, open datasets were used, most of them were gathered from the Curated Microarray Database (CuMiDa). The algorithm performance was measured by the geometric mean, the number of selected features, and the population hypervolume, and it was compared against N3O on microarray binary classification problems with competitive results. Furthermore, the results were investigated for statistical significance and show that the novel algorithm has a promising performance on the problem at hand.