ABSTRACT
In soil chemistry, the nutrients exhibit non-linear and complex relationships owing to their stochastic nature but mostly their similarity is a function of the distance between the data points. The similarity assessment using distance metrics is a popular technique employed by the regression, classification and feature selection algorithms. To enhance the precision of distance metric, the kernel trick is performed on the input space and the similarity is ascertained in the new high dimensional feature space. In Kernel Distance Metric Learning (KDML), the relevance of distance metrics is intensified to capture the precise similarity measure. In Hierarchical Kernel Learning (HKL) and Additive Gaussian Process (AGP) models, several orders of interactions among the subsets of predictors are emphasized while learning the kernel. In this paper a novel method, Restricted Additive Model (RAM) embedded in Additive Gaussian Process (AGP), to compute the distance in input space by adding selective weighted distances from the subset of predictors is proposed. RAM focuses on reusing the information content obtained while preprocessing the data and incorporate it while learning with the kernel. This can save a good amount of computational resources for high dimensional datasets. The proposed model is compared with HKL, AGP and a normal Gaussian Process (GP). The adjusted R2 and the Mean Absolute Error values showed that the proposed model showcased good accuracy reducing the computational time and resources. Further, the comparison of RAM with Automatic Relevance Determination of GP testified that the reusability of the information content turned to be effective in building a parsimonious model.
