Accurate prediction of scientific impact is important for scientists,academic recommender systems, and granting organizations alike.Existing approaches rely on many years of leading citation valuesto predict a scientific paper’s citations (a proxy for impact), eventhough most papers make their largest contributions in the firstfew years after they are published. In this paper, we tackle a newproblem that predicting a newly published paper’s citation timeseries from the date of publication (i.e., without leading values). WeproposeHINTS, a novel end-to-end deep learning framework thatturns citation signals from dynamic heterogeneous informationnetworks (DHIN) to citation time series. HINTS imputes pseudo-leading values for a paper in the years before it is published fromDHIN embeddings, and then transforms these embeddings into theparameters of a model that can predict citation counts immediatelyafter publication. Empirical analysis on two real-world datasetsfrom Computer Science and Physics show that HINTS is competi-tive with baseline citation prediction models. While we focus oncitations, our approach generalizes to other “cold start” time seriesprediction tasks where relational data is available and accurateprediction in early timestamps is crucial.

The Web Conference is announcing latest news and developments biweekly or on a monthly basis. We respect The General Data Protection Regulation 2016/679.