Combining labeled and unlabeled data with graph embedding

Abstract Learning the manifold structure of the data is a fundamental problem for pattern analysis. Utilizing labeled and unlabeled data, this paper presents a novel manifold learning algorithm, called semi-supervised aggregative graph embedding (SSAGE). In SSAGE, the graph of the original data is constructed and preserved according to a certain kind of similarity, which takes special consideration of both the local geometry information (of both labeled and unlabeled data) and the class information (of labeled data). The similarity has several good properties which help to discover the true intrinsic structure of the data, and make SSAGE a robust technique for inductive inference. Experimental results suggest that the proposed SSAGE approach provides a better representation of the data and achieves much higher recognition accuracies than Zhou's algorithm [D. Zhou, O. Bousquet, T.N. Lal, J. Weston, B. Scholkopf, Learning with local and global consistency, Advances in Neural Information Processing Systems, vol. 16, MIT Press, Cambridge, MA, 2003] and PCA.