Regression Using Multikernel and Semiparametric
Cong-Van M. Nguyen, D.B.H. Tay · 2008
In this letter, we propose a sequential training scheme for multikernel support vector regression (SVR). Unlike the mul- tistage backfitting technique; our method re-tunes, at every stage, all previously trained weights using a semiparametric algorithm in the presence of one more kernel function. In this way, local minima are avoided and any combination of arbitrary kernel functions is acceptable. By experimenting on some synthetic and real data sets, we demonstrate that our method yields a better trade-off between sparsity and accuracy in comparison with the conventional single- kernel SVR and the multikernel backfitting SVR. are zeros. Those input vectors cor- responding to are called support vectors (SVs), all of which define the trained model. We call each component func- tion (w.r.t. )a basis. The kernel function is user-de- fined, and choosing an appropriate one is remaining a chal- lenge. Rather than those general kernels such as Gaussian ra- dial basis functions (GRBFs), some sophisticated kernel func- tions have been proposed to adapt complicated shapes of the data trend. In (2), they build the kernel from a linear combina- tion of some basic kernels, while (3) uses wavelet-based kernel functions. However, all single-kernel SVR methods usually face a problem: the trained model cannot globally fit the system if there are different regions with different data trends. To overcome the above problem, multiscale SVR methods were proposed. They yield a more complex solution in which more than one kernel function are deployed, as follows: This multikernel