Fuzzy rough based regularization in Generalized Multiple Kernel Learning
Yamuna Prasad, K. K. Biswas · 2012
Recent advances in kernel methods have positioned it as an attractive tool for many research areas. To reveal precise data similarity, learning of good kernel representation is essential. GMKL formulation based on gradient descent optimization with various regularizations has been well established in the literature. GMKL learns linear, product and exponential combinations of given base kernels which makes it more robust and efficient than traditional Multiple Kernel Learning (MKL). GMKL also has been proven a good tool for feature selection as well. The time taken for convergence of MKL depends upon the initialization of kernel weights. Several optimizations initialize kernel weights randomly which produces variability in convergence time. To tackle this issue, we propose fuzzy rough based kernel weight initialization unlike random initialization in GMKL, which makes GMKL converge faster. The proposed fuzzy rough GMKL (FR-GMKL) is tested on benchmark UCI and microarray databases. Our results show the faster and stable convergence of FR-GMKL as compared to GMKL.