Multi layer Kernel Learning for time series forecasting

Multiple Kernel Learning (MKL) is one of recent approaches to choose suitable kernels from a given pool of kernels by exploring the combinations of multiple kernels. For linear kernel, the target kernel is a linear combination some base kernels. However, some literatures suggest that a linear combination of kernels cannot consistently outperform either the uniform combination of base kernels or simply the best single kernel. Hence, some researchers attempt to study the non-linear combination of kernels, such as polynomial combination of kernels or two-layer MKL. This paper extends the previous work on two-layer MKL into three-layer MKL especially in the field of regression to forecast future values of time series. Our experiment on several time series dataset demonstrates that our proposed method generally outperforms the linear combination of kernels.