Popularity Forecasting for Emerging Research Topics at Its Early Stage of Evolution.

The accurate modelling and forecasting of the popularity of emerging topics can benefit researchers by allocating resources and efforts on promising research directions. While existing forecasting approaches enjoy various levels of success, most suffer from at least one of the following three limitations: a limited scope due to having to mine topic terms from only a few documents, low generalizability due to assigning arbitrary binary classifications on topics to be either "emerging" or not, or using an emerging topic or field of study's historical features as inputs to forecast its future popularity while disregarding the existing effect of a "cold start". In this paper we propose a forecasting algorithm that address all three limitations in three steps. Firstly, we leverage the field of study taxonomy present in most academic databases to obtain a neighborhood of trending fields within the discipline of the field of study of interest. Then, dynamic time warping is used to measure the similarity of each neighbour's trending pattern compared to the trending pattern of the field of study of interest. Lastly, we conduct multivariate forecasting using a LSTM model while utilizing the historical popularity scores of similar trending neighbours as input. Experimental results on 5 emerging fields of study showcases the "cold start" phenomenon as well as the proposed algorithm reducing RMSE, MAE, and MAPE by half for 4 emerging topics. This validates the claim of the limitations for existing methods and provides insight on the dependency structure of emerging topics with their historical features.