摘要： 通过采用稀疏贝叶斯推理方法,设计出可同时进行学习最优分类器与选取最优特征子集的特征选择概率分类向量机算法。该算法是对概率分类向量机特征选择的扩展,可提高其在高维数据集上的性能。通过选取零均值的高斯分布作为先验,在模型中起到正则项的作用,同时在核函数和特征中引入稀疏,得到泛化性更好的分类模型。在高维度和低维度数据集中的实验结果表明,该算法同时具有较好的分类和特征选择能力。

关键词: 机器学习, 核函数, 稀疏贝叶斯, 特征选择, 概率分类向量机, 自动相关性检测

Abstract: Through using sparse Bayesian inference thought,a Feature Selection Probabilistic Classification Vector Machine (FPCVM) is designed which can learn optimal classifier and automatically select the most relevant feature subset.FPCVM is an extension of Probabilistic Classification Vector Machine(PCVM),which improves the performance of PCVM on high dimension datasets.It uses zero-mean Gaussian distribution as priori to introduce sparseness both in kernel functions and feature space;these priors are preformed as regularization items in the likelihood function to acquire more generalized model.Experimental results on high dimension datasets and low dimension datasets show that the algorithm has better classification and feature selection.

Key words: machine learning, kernel function, sparse Bayesian, feature selection, Probabilistic Classification Vector Machine(PCVM), automatic relevance determination

中图分类号: 

• TP18