A new classifier combination method based on TSK-type fuzzy system

This paper proposes a new classifier combination method based on TSK-type fuzzy system. The new method includes linear combination as a special case and it is potential to outperform linear combination by tuning its parameters. The fuzzy system used to combine the classifiers has a good comprehensibility since the number of fuzzy rules is very small. In the new method, first, the parameters of the TSK-type fuzzy system are initialized based on an efficient linear combination method, multi-response linear regression; then, the gradient descendent algorithm is used to tune the parameters. Experimental results on three data sets form the ELENA project database and the UCI database show that the proposed classifier combination method outperforms the multi-response linear regression and other typical classifier combination methods such as voting and multilayer perceptron networks.