2018/12/17 by Elif Vural, Vural, Elif
Computer Science · #Domain Adaptation and Few-Shot Learning #Machine Learning and ELM
paper · pdf · doi:10.48550/arxiv.1812.06944
Traditional machine learning algorithms assume that the training and test\ndata have the same distribution, while this assumption does not necessarily\nhold in real applications. Domain adaptation methods take into account the\ndeviations in the data distribution. In this work, we study the problem of\ndomain adaptation on graphs. We consider a source graph and a target graph\nconstructed with samples drawn from data manifolds. We study the problem of\nestimating the unknown class labels on the target graph using the label\ninformation on the source graph and the similarity between the two graphs. We\nparticularly focus on a setting where the target label function is learnt such\nthat its spectrum is similar to that of the source label function. We first\npropose a theoretical analysis of domain adaptation on graphs and present\nperformance bounds that characterize the target classification error in terms\nof the properties of the graphs and the data manifolds. We show that the\nclassification performance improves as the topologies of the graphs get more\nbalanced, i.e., as the numbers of neighbors of different graph nodes become\nmore proportionate, and weak edges with small weights are avoided. Our results\nalso suggest that graph edges between too distant data samples should be\navoided for good generalization performance. We then propose a graph domain\nadaptation algorithm inspired by our theoretical findings, which estimates the\nlabel functions while learning the source and target graph topologies at the\nsame time. The joint graph learning and label estimation problem is formulated\nthrough an objective function relying on our performance bounds, which is\nminimized with an alternating optimization scheme. Experiments on synthetic and\nreal data sets suggest that the proposed method outperforms baseline\napproaches.\n