A Novel Regressive Algorithm Based on Relevance Vector Machine
Errui Ding, Ping Zeng, Yong Fang Yao · 2007
To improve the prediction accuracy and running efficiency, a novel sparse Bayesian learning algorithm for regression is proposed. Based on relevance vector machine, the algorithm firstly increases the prediction accuracy by adopting multiple kernels which are constructed from the angles of complete and over-complete bases. To lessen the training time caused by multiple kernels, the algorithm has two reduced steps involving a preliminary model and an eventual model. The improved locality preserving projections is used to reduce the column dimension of the input matrix, which forms the preliminary model. To further relieve the time pressure for a larger training sample set, the eventual model generates a pruned training sample set by pruning the old sample set with the preliminary model based on the cluster centers. Experimental results indicate the proposed algorithm is superior, in both prediction accuracy and robustness, to relevance vector machine while having less training time.