Título: Preventing Error Propagation In Semi-Supervised Learning Using Teams Of Walking Particles
Autores: Breve, Fabricio; Zhao, Liang
Resumo: Semi-supervised learning algorithms are applied to classification problems where only a small portion of the data points is labeled. In these cases, the reliability of the labels in the labeled subset is very important, because mislabeled samples may propagate their wrong labels to a large portion of the data set. This paper presents a novel and efficient semi-supervised learning graph-based method specifically designed to handle data sets with mislabeled samples. It uses walking particles with cooperative and competitive behavior in order to propagate labels. The proposed model also incorporates some features to make it robust to large amounts of mislabeled samples. Computer simulations show the performance of the method in the presence of different amounts of mislabeled data, in networks of different sizes and mixtures. These simulations identify critical points of mislabeled subset size, below which the network is free of wrong label contamination, but above which the mislabeled samples start to propagate their labels to the rest of the network. Moreover, the proposed method is compared to other representative semi-supervised learning graph-based methods and its performance in real-world data sets is increasingly better than the others as the amount of mislabeled samples in the data set increases.
Palavras-chave: Semi-supervised learning; Learning from imperfect data; Particle competition and cooperation; Error propagation analysis
Código DOI: 10.21528/CBIC2011-39.3
Artigo em pdf: st_39.3.pdf
Arquivo BibTex: st_39.3.bib