摘要
相关向量机(relevance vector machine,RVM)是一种基于稀疏贝叶斯原理的分类和回归建模方法,具有泛化能力强、有效刻画数据不确定性以及参数设置简单等优点。然而,RVM假定权重矩阵和数据噪声均服从高斯分布,降低了RVM的鲁棒性和泛化性能。为此,提出一种变分Bayesian推理的鲁棒稀疏相关向量机建模方法,继承了RVM的优点,同时具有更好的鲁棒性和泛化性。该方法新颖之处在于:通过对权重矩阵分布施加Laplace分布以保证权重矩阵的稀疏性;通过对建模噪声施加学生分布约束以及自适应调节学生分布的自由度参数,较好地描述数据的不确定性,增强所提方法对复杂数据建模能力;引入变分Bayesian推理方法求取最优RVM模型参数和超参数。仿真结果证明所提算法具有良好的鲁棒性和稀疏性,优于现有的变形RVM算法。
Relevance vector machine(RVM)is a sparse Bayesian principle-based classification or regression method with merits of good generalization,capturing uncertainty of data and simple selection of model parameters.However,weight matrix of RVM and modeling noises are assumed to follow Gaussian distribution,limiting the generalization performance and robustness of RVM.Therefore,a variational Bayesian inference-based robust and sparse relevance vector machine is developed in this paper to deal with the issues,inheriting the merits of RVM and achieving better generalization performance and robustness.The novelties of this proposed paper are as following:the sparsity of weight matrix can be ensured in theory by imposing Laplace distribution on weight matrix;uncertainties of the data can be effectively captured by imposing student distribution on model noises,and then the robustness and modeling accuracy of the proposed method for complex data can be enhanced greatly by tuning the freedom parameters of student distribution;variational Bayesian inference is introduced to yield the optimal RVM model parameters and hyperparameters.The simulation results show that the proposed algorithm has good robustness and sparsity,and is superior to the existing variants of RVM.
作者
任世锦
吴晓轩
朱艳冉
叶雨晴
胡晓双
柯源鑫
REN Shijin;WU Xiaoxuan;ZHU Yanran;YE Yuqing;HU Xiaoshuang;KE Yuanxin(School of Smart Education,Jiangsu Normal University,Xuzhou 221116,China)
出处
《江苏海洋大学学报（自然科学版）》
CAS
2023年第2期79-87,共9页
Journal of Jiangsu Ocean University:Natural Science Edition
