Adaptation of Graph-Based Semi-Supervised Methods to Large-Scale Text Data

Graph-based semi-supervised learning methods have shown to be efficient and effective on network data by propagating labels along neighboring nodes. These methods can also be applied to general data by constructing a graph where the nodes are the instances and the edges are weighted by the similarity between feature vectors of instances. However, whereas a natural network is often sparse, a network of pairwise similarities between instances is dense, and prohibitively large for even moderately sized text datasets. We show, through using a simple general technique, how these learning methods can be exactly and efficiently applied to text data—using the complete pair-wise similarity manifold—without resorting to sampling or sparsification. This technique also provides a unifying view of prior work on label propagation on text graphs, and we assess its effectiveness applied to two popular graph-based semisupervised methods on several large real datasets.