Data-driven, kernel-based learning methods have been of increasing interest lately, and have shown great impact in areas such as machine learning, computer vision and pattern recognition. However, the rapidly increasing amount of available data poses new challenges that traditional methods struggle with. In this paper we present computationally tractable method for kernel regression on large data sets. The size of the kernel matrix grows rapidly as the number of data points increases, and computing and storing it may even become intractable. Hence, we have proposed a criteria based on linear independence between feature vectors forselecting the relevant input data points for building the kernel matrix. The proposed method do not require storing the full kernel matrix in memory and will result in a sparse kernel approximation. The proposed method has been applied in a case study for estimating the recovered waste-heat energy in ships based on engine load data.
|Publication status||Published - 2016|