论文部分内容阅读
在ε-不敏感支持向量回归(ε-insensitive support vector regression,ε-SVR)正则化路径的基础上,提出基于输入K-近邻的三步式SVR模型组合方法。在整个样本集上进行训练,求得ε-SVR的正则化路径。由SVR正则化路径的分段线性性质确定初始模型集合,并应用平均贝叶斯信息准则(Bayesian Information Criterion,BIC)策略对初始模型集合进行修剪以获得候选模型集合。该修剪策略可减小候选模型集合的规模,提高模型组合的计算效率和预测性能。在预测或测试阶段,根据样本输入向量采用K-近邻法确定最终组合模型集合,并实现贝叶斯组合预测。证明了ε-SVR模型组合的Lε-风险一致性,给出了SVR模型组合基于样本的合理性解释。试验结果验证了正则化路径上基于输入K-近邻的ε-SVR模型组合的有效性。
Based on the ε-insensitive support vector regression (ε-SVR) regularization path, a three-step SVR model combination method based on input K-nearest neighbor is proposed. Training is performed on the entire sample set to obtain the regularization path of ε-SVR. The initial model set is determined by the piecewise linear properties of the SVR regularization path. The initial Bayesian Information Criterion (BIC) strategy is used to trim the initial model set to obtain the candidate set. This pruning strategy can reduce the size of the candidate model set and improve the computational efficiency and predictive performance of the model combination. In the prediction or test phase, the K-nearest neighbor method is used to determine the final set of combined models according to the sample input vector, and Bayesian combination prediction is implemented. The Lε-risk consistency of ε-SVR model portfolio is proved, and the sample-based rationality of SVR model portfolio is given. The experimental results verify the validity of the combination of ε-SVR models based on input K-nearest neighbors on the regularization path.