Learning a Joint Representation for Classification of Networked Documents

Abstract

Recently, several researchers have incorporated network information to enhance document classification. However, these methods are tied to some specific network representations and are unable to exploit different representations to take advantage of data specific properties. Moreover, they do not utilize the complementary information from one source to the other, and do not fully leverage the label information. In this paper, we propose CrossTL, a novel representation model, to find better representations for classification. CrossTL improves the learning at three levels: (1) at the input level, it is a general framework which can accommodate any useful text or graph embeddings, (2) at the structure level, it learns a text-to-link and link-to-text representation to comprehensively describe the data; (3) at the objective level, it bounds the error rate by incorporating four types of losses, i.e., text, link, and the combination and disagreement of text and link, into the loss function. Extensive experimental results demonstrate that CrossTL significantly outperforms the state-of-the-art representations on datasets with either rich or poor texts and links.