Predicting the Future Impact of Academic Publications

Predicting the future impact of academic publications has many important applications. In this paper, we propose methods for predicting future article impact, leveraging digital libraries of academic publications containing citation information. Using a set of successive past impact scores, obtained through graph-ranking algorithms such as PageRank, we study the evolution of the publications in terms of their yearly impact scores, learning regression models to predict the future PageRank scores, or to predict the future number of downloads. Results obtained over a DBLP citation dataset, covering papers published up to the year of 2011, show that the impact predictions are highly accurate for all experimental setups. A model based on regression trees, using features relative to PageRank scores, PageRank change rates, author PageRank scores, and term occurrence frequencies in the abstracts and titles of the publications, computed over citation graphs from the three previous years, obtained the best results.