The Hard-Cut EM Algorithm for Mixture of Sparse Gaussian Processes

Abstract

The mixture of Gaussian Processes (MGP) is a powerful and fast developed machine learning framework. In order to make its learning more efficient, certain sparsity constraints have been adopted to form the mixture of sparse Gaussian Processes (MSGP). However, the existing MGP and MSGP models are rather complicated and their learning algorithms involve various approximation schemes. In this paper, we refine the MSGP model and develop the hard-cut EM algorithm for MSGP from its original version for MGP. It is demonstrated by the experiments on both synthetic and real datasets that our refined MSGP model and the hard-cut EM algorithm are feasible and can outperform some typical regression algorithms on prediction. Moreover, with sparse technique, the parameter learning of our proposed MSGP model is much more efficient than that of the MGP model.