查询结果:   李斌,王卫星.NCA降维和贝叶斯优化调参对分类模型的改进[J].计算机应用与软件,2019,36(8):281 - 287,299.
中文标题
NCA降维和贝叶斯优化调参对分类模型的改进
发表栏目
算法
摘要点击数
446
英文标题
IMPROVEMENT OF CLASSIFICATION MODEL BY NCA DIMENSION REDUCTION AND BAYESIAN OPTIMIZATION PARAMETER ADJUSTMENT
作 者
李斌 王卫星 Li Bin Wang Weixing
作者单位
河南科技大学应用工程学院现代教育技术中心 河南 三门峡 472000     
英文单位
Modern Education Technology Center,College of Applied Engineering, Henan University of Science and Technology, Sanmenxia 472000, Henan,China     
关键词
分类算法 领域分量分析 贝叶斯调优 MATLAB 贫困生判别
Keywords
Classification algorithm Neighborhood component analysis(NCA) Bayesian tuning MATLAB Poor student discriminant
基金项目
河南省2017年高等教育教学改革研究与实践项目(2017SJGLX636)
作者资料
李斌,高工,主研领域:数据分析,网络规划。王卫星,副教授。 。
文章摘要
高校贫困生的贫困程度判定可以归属于构建分类模型对样本数据进行训练。但单个分类模型的精准度要取决于处理样本数据的大小和类型复杂度,在模型速度和准确性之间不易取舍。集成多个分类算法可以避免单个分类算法的过拟合。通过邻域分量分析(Neighborhood Component Analysis, NCA)进行特征降维降低初始分类模型的计算成本,对误判损失引入一个成本函数进行惩罚的同时采用贝叶斯优化进行超参数调优。结果表明,改进后的分类模型泛化能力得到明显提升。计算时间成本降低的同时,误判率由初始的8%下降到5%,模型的准确率提升了近4%。
Abstract
Poverty levels of poor students in the university can be attributed to build a classification model of training sample data. But the model of a single classification accuracy depends on the size of the sample data and types of complexity, and it is difficult to choose between the speed and accuracy of the model. Integrating multiple classification algorithm can avoid a single classification algorithm of fitting. Through the neighborhood component analysis (NCA) for feature dimension reduction, we reduced initial classification model of calculating cost. For misjudgment loss, we introduced a cost function to punish and used bayesian optimization to super parameter tuning simultaneously. The results show that the generalization ability of improved classification model is improved significantly. At the same time, the computation time cost decreases, misjudgment rate decreases from 8% to 5%, and the accuracy of the model increases by nearly 4%. 
下载PDF全文