摘要
针对现有的动态选择策略局限于寻找待测样本的局部相似样本,未充分考虑样本特征之间的重要性程度,从而对预测精度造成影响的问题,该文提出一种基于近邻样本评估的动态选择性集成预测算法。算法基于误差扰动度量出特征的重要性权值,并在此基础上进行样本近邻的相似性度量。根据不同的待测样本特点自动适应近邻数目,找到最佳近邻。通过最佳近邻对具有不同预测精度的学习器的性能评估,择优筛选出精度较高的学习器进行选择性集成预测。实验结果表明,相比原有集成学习算法和普通选择性集成算法,该算法预测精度得到进一步提升,表现出良好的预测效果和较强的预测性能。
Aiming at the problem that the existing dynamic selection strategy is limited to finding locally similar samples of the test sample and does not fully consider the importance of the sample features,thus affecting the prediction accuracy,a dynamic selective ensemble prediction algorithm based on evaluation of neighborhood sample is proposed.Based on the importance weight of the feature measured by the error perturbation,the similarity of neighborhood sample is measured.According to the characteristics of different samples to be measured,the number of nearest neighbor is automatically adapted to find the best nearest neighbor.By evaluating the performance of learners with different prediction accuracy,the learner subset with high accuracy is selected for selective ensemble prediction.The experimental results show that,compared with the original ensemble learning algorithm and the common selective ensemble algorithm,the prediction accuracy of this algorithm is further improved,showing that it has good effect of prediction and strong performance.
作者
曲文龙
李一漪
陈笑屹
曲嘉一
QU Wenlong;LI Yiyi;CHEN Xiaoyi;QU Jiayi(School of Information Engineering, Hebei GEO University, Shijiazhuang 050031, China;School of Science, Hebei University of Science and Technology, Shijiazhuang 050018, China)
出处
《西北大学学报(自然科学版)》
CAS
CSCD
北大核心
2020年第5期802-810,共9页
Journal of Northwest University(Natural Science Edition)
基金
河北省重点研发计划资助项目(18212005)
河北省自然科学基金资助项目(F2016403055)。
关键词
动态选择性集成
回归预测
近邻样本
相似度量
dynamic selective ensemble
regression prediction
nearest neighbor sample
similarity measure