Genetic fuzzy classification fusion of multiple SVMs for biomedical data

Classification of biomedical data faces a special challenge because of the characteristics of the data: too few data examples with too many features. How to improve the classification performance or the generalization ability of a classifier in the biomedical domain becomes one of the active research areas. One approach is to build a fusion model to combine multiple classifiers together and result in a combined classifier which can achieve a better performance than any of its composing individual classifiers. In this paper, we propose a SVM classifier fusion model to combine multiple SVMs by applying the knowledge of fuzzy logic and genetic algorithms. The fuzzy logic system (FLS) is constructed based on SVM accuracies and distances of data examples to SVM hyperplanes in SVM feature spaces. A genetic algorithm (GA) is used to tune the fuzzy membership functions (MFs) in the FLS and determine the optimal fuzzy fusion model. We have applied the proposed model to two biomedical data: colon tumor data and ovarian cancer data. Our experiment shows that multiple SVM classifiers complement each other well in the proposed fusion model and the ensemble achieves a better, more robust and more reliable performance than individual composing SVMs.