Predicting the citation counts of research papers using neural networks

Abstract

A widely accepted criterion used to measure the scientific impact of a research paper is the citation count. However, for a newly published paper this metric does not become available for several years after the date of publication. Yet, many parties including fellow scholars, research institutes and funding bodies find it important to be able to identify early on, the scientific papers with a higher potential to make a bigger impact. Predicting the future citation counts is an effective solution to overcome this limitation.

However, predicting the future citation count of a scientific paper is a challenging task, particularly due to the highly dynamic nature in the citation accumulation process. Hence this remains an active area of research. A majority of the prior studies that predict future citation counts using features available at the time of publication of a research paper make use of classical machine learning techniques. In this study, the author demonstrates through experiments how artificial neural network models can outperform best performing classical machine learning models discussed in prior studies.

One notable limitation of current approaches to this research problem is that many approaches treat the citation networks as unweighted graphs. In this work, the author demonstrates how treating the citation relationships as weighted relationships could help improve performance of the models. For this, the author introduces a novel feature named Weighted Average Neighboring Citation Score, a value computed by treating the citation network as a weighted graph, and demonstrates through multiple experiments that the newly introduced feature helps improve the performance of multiple models. Moreover, the author experiments with different edge weighting schemes and demonstrates how factoring both the recency of a citation and frequency with which a source has been cited when determining the edge weights help improve the performance of the models.