Tree decomposition for large scale semi-supervised classification

This paper proposes an algorithm framework (CFTD-SSC) for large scale semi-supervised classification based on Clustering Feature (CF) tree decomposition and local learning. The method firstly applies the CF tree to organize unlabeled data points and decomposes the unlabeled dataset into a series of subsets. Secondly on each subset, CFTD-SSC can classify the unlabeled data points through some widely used semi-supervised classification methods. In this paper, Gaussian Fields and Harmonic Functions (GFHF) and Local and Global Consistency (LGC) are adopted. In addition, this paper improves LGC algorithm into Local Graph Transduction (LGT). Thus this paper designs three semisupervised classification algorithms including CFTD-GFHF, CFTD-LGC and CFTD-LGT. Thirdly, this paper analyzes the influence of the parameters in CFTD-SSC framework. The experimental results show that compared to current excellent large scale semi-supervised classification algorithms, such as Prototype Vector Machine (PVM) and AnchorGraphReg (AGR), the three algorithms under CFTD-SSC framework have the advantages of shorter learning time and higher classification accuracy. Among the three algorithms under CFTD-SSC framework, CFTD-LGC is better in learning time, but improved CFTD-LGT is superior in classification accuracy. Copyright © 2013 Binary Information Press.