Predicting Scientific Impact via Heterogeneous Academic Network Embedding

Abstract

Predicting the scientific impact of literatures and researchers has been widely studied for a long time, and prevailing graph-based methods mainly rely on the global structure of the academic network. However, systematically integrating the local structural information of academic network and the text information of articles into a unified model to jointly predict the future impact of published articles and future potential of researchers is relatively unexplored. In this paper, we focus on how to effectively leverage these aforementioned information to predict the future impact of articles and researchers. Specifically, we first design a novel network embedding model which can simultaneously learn the local structural information of heterogeneous academic network and the text information of articles into a low-dimensional vector. Then a multivariate random-walk model is proposed to mutually rank the future impact of articles and researchers by leveraging such learned embeddings. We conduct extensive experiments on the AMiner dataset and the ACL Anthology Network, and the results demonstrate that the proposed method significantly outperforms existing state-of-the-art ranking approaches.