摘要
为提高多核学习对大规模及分布不均衡问题的处理能力,提出一种基于神经正切核草图的多核学习方法(neural tangent kernel sketch multiple kernel learning, NS-MKL)。应用神经正切核代替单层核函数作为多核学习基核函数,提高多核学习方法表示能力;使用神经正切核草图算法对神经正切核进行近似,减少神经正切核的特征数量和特征维度,提高多核学习方法计算效率;使用核目标对齐计算每个近似神经正切核的基核权重,根据权重进行多核线性组合,得到多核决策函数。在3个UCI数据集上对神经正线核(neural tangent kernel, NTK)核支持向量机(support vector machine, SVM)与传统核SVM进行比较分析,NTK核SVM比传统核SVM预测准确率最低提高1.9%,精度最低提高2.0%,召回率最低提高2.0%。在3个UCI数据集上对NS-MKL与传统核MKL进行比较分析,NS-MKL比应用传统核MKL预测准确率最低提高2.0%,运行时间最低减少9 s。NS-MKL能提高预测准确率,降低计算速度。
To address large-scale and unbalanced distribution problems,a neural tangent kernel sketch-based multiple kernel learning method(NS-MKL)was proposed.The neural tangent kernel was applied instead of the single-layer kernel function as the base kernel function of the multiple kernel learning to enhance the representation capability of the multiple kernel learning method.The neural tangent kernel was approximated using the neural tangent kernel sketch algorithm,which reduced the number of features and the feature dimension of the neural tangent kernel,thus improved the computational efficiency of the multiple kernel learning method.The kernel target alignment was applied to compute the base kernel weight for each approximate neural tangent kernel,and a linear combination of multiple kernels was performed based on the weight to obtain the final multiple kernel decision function.By compareing the single-layer kernel with the NTK-based SVM in 3 UCI datasets,the NTK-based SVM improved the accuracy at least 1.9%,precision at least 2.0%,and recall rate at least 2.0%.By compareing the NS-MKL with other MKL methods in 3 UCI datasets,the NS-MKL improved the accuracy at least 2.0%,and runtime reduced at least 9 s.The proposed algorithm had higher predictive accuracy and faster computation speed.
作者
王梅
许传海
王伟东
韩非
WANG Mei;XU Chuanhai;WANG Weidong;HAN Fei(School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,Heilongjiang,China;College of Information Engineering,Xinjiang Institute of Technology,Aksu 843100,Xinjiang,China;Artificial Intelligence Energy Research Institute,Northeast Petroleum University,Daqing 163318,Heilongjiang,China)
出处
《山东大学学报（工学版）》
CAS
CSCD
北大核心
2024年第4期13-20,34,共9页
Journal of Shandong University（Engineering Science）
基金
国家自然科学基金资助项目(51774090,62073070)
黑龙江省博士后科研启动金资助项目(LBH-Q20080)
黑龙江省研究生精品课程建设项目(15141220103)。
关键词
多核学习
神经正切核
核目标对齐
反余弦核
草图算法
multi-kernel learning
neural tangent kernel
kernel-target alignment
arc-cosine kernel
sketch algorithm
