An improved learning with local and global consistency
Ming Li, Xiaoli Zhang, Xuesong Wang · 2010
Learning with local and global consistency (LLGC) algorithm can effectively label a data, but it is helpless for noise data. The reason is that the LLGC algorithm will predict a label for each unlabelled data without taking into account whether a data has noise or not. Aiming at the deficiency of the LLGC algorithm, an improved version for semi-supervised learning algorithm with local and global consistency is proposed in this paper. At first, we compute the similarity of each data to all classes. And then the data can be ascribed to one class according to its similarities. The improved LLGC algorithm not only can label data as the conventional LLGC, but also can identify noise existed in data set effectively. Simulation results show that the improved LLGC algorithm can effectively avoid noise data being viewed as normal data.