一种新的回归型约简多分辨率相关向量机

A novel regression algorithm of reduced multi-resolution relevance vector machine

提出一种新的稀疏贝叶斯回归算法.基于相关向量机,首先通过尺度核和小波核构造完备基以提高预测精度;然后利用保局投影对输入矩阵的列进行主成分提取以减少训练时间,从而形成算法的初步模型.为进一步减小较大规模训练数据集的回归时间压力,算法对训练数据集的分层采样建立了初步模型,进而产生实际较小规模的训练数据集.实验结果表明,算法在预测精度和鲁棒性上优于传统支持向量机和相关向量机,且其训练时间较相关向量机少.

To further improve the prediction accuracy and running efficiency, a novel sparse Bayesian algorithm for regression is proposed. Based on relevance vector machine, a set of complete basises are constructed by combing scaling and wavelet kernels to increase the prediction accuracy, and then the principal components of input matrix columns are extracted by using locality preserving projections to reduce the training time, which forms a primary model. To further reduce the time pressure for a larger training data set, the algorithm creates a smaller training data set via the primary model based on a stratified sample. Experimental results of artificial and real data show that the proposed algorithm is superior to traditional support vector machine and relevance vector machine in both prediction accuracy and robustness and has less training time than relevance vector machine.