ETINE: Enhanced Textual Information Network Embedding

Abstract

In many reality networks, nodes contain rich text attribute information that exhibits significant role in describing the properties of them and relationship between them. The integration of structural and textual information is beneficial to downstream network analysis tasks. In this work, we present an Enhanced Textual Information Network Embedding model, called ETINE, for learning network embeddings with not only global structural information but also deep semantic relationship between nodes. Specifically, we formulate the optimization of our proposed structure-based and text-based loss functions as a matrix approximation problem. Moreover, to enhance the efficiency and robustness of the proposed method, we propose to optimize the loss functions with an efficient randomized singular value decomposition (RSVD) method. Extensive experiments on four benchmarks demonstrate that our model outperforms other state-of-the-art baselines in multi-class node classification, network reconstruction and node clustering tasks.