Early Prediction of Scientific Impact Based on Multi-Bibliographic Features and Convolutional Neural Network

The increasingly available large-scale bibliographic data that generate a heterogeneous network provide opportunities to detect, track, and predict the evolution of science. Recently, many efforts have been devoted to quantifying the impact of scientific papers within different citation time windows. However, the complex patterns of the citation network make it difficult to predict future citations on the basis of a short time window. Accordingly, we present a data-centric methodology to predict long-term scientific impact by combining numerous bibliographic features and convolutional neural network. More specifically, we first expand the input features from the annual citation records to the features of the whole heterogeneous bibliographic information network that completely represents the topology structure of academic activities. Then, a convolutional neural network model is designed to capture the complex nonlinear relationships between the early network features and the final cumulative citation count. Finally, we conduct an experiment on papers of Markov Chain from 1980 to 1985. The result shows that the prediction performance can be improved by 5% to baseline models under the same problem definition and with the same dataset. Meanwhile, the long-term scientific impacts are strongly correlated with its recognition by authoritative authors or venues in the early stage.