基于相对变换距离的半监督分类算法

摘要

针对半监督分类过程中使用欧式距离选择样本的邻节点不能很好适应噪音或稀疏数据,导致算法分类精度下降问题,提出一种基于相对变换的RT-LapRLS算法.该方法利用相对变换距离对样本的近邻点进行选择,构造相对变换邻接图,在相对变换邻接图上构造流形正则项,最后用LapRLS算法得到分类函数.通过人工数据集和真实数据集上的实验验证了该算法的有效性,实验表明相比于欧式距离,相对变换距离可以减少数据稀疏以及噪音对算法的影响,提高算法的鲁棒性.%In semi-supervised classification process,using Euclidean distance to select the neighbour nodes of the sample can not well adapt to the noise or sparse data,and this results in the decline of algorithm classification.Aiming at the problem,a classification RT-LapRLS algorithm is presented,which is based on relative transformation.First,the relative transformation distance is used to select sample neighbour nodes,then the adjacency graph of relative transformation is constructed,and the manifold regularisation is further constructed based on the adjacency graph,finally the classification function is produced by LapRLS algorithm.The effectiveness of the algorithm has been verified on artificial data sets and real data sets,and experiment results show that the relative transformation distance can reduce the impact of sparse data and noise on the algorithm and can improve its robustness as well compared with the Euclidean distance.