Fusing similarities and kernels for classification
Yihua Chen, Maya R. Gupta · International Conference on Information Fusion · 2009
The problem of fusing indefinite similarity information and positive semidefinite similarity information together for classification is considered. The proposed solution jointly (i) learns a spectrum modification to make the indefinite similarity positive semidefinite, (ii) learns a conic combination of multiple given positive semidefinite kernels, and (iii) learns the parameters of a discriminative classifier. We show that the proposed fusion method can be formulated as a convex optimization problem. This work extends previous work in multiple kernel learning. Though applicable to other kernel methods, the focus is on the support vector machine. Experiments with four real data sets show that the proposed method is consistently among the best performers.