Citation Count Prediction of Academic Papers (Bilimsel Makalelerin Atıf Sayısı Tahmini)

Even though measuring the impact of scientific papers is not a straightforward process, their citation counts play a significant role in this determination. Citation count of a paper, however, is not available until the paper gets published and a substantial amount of time passes until it spreads through the community. To overcome this issue, we relax the problem by building a deep learning model that predicts whether a paper will receive at least one citation in a one-year interval after its publication. Our model employs Long ShortTerm Memory (LSTM) to capture the relationship between word sequences. In our study, we also analyze the effect of using the abstract versus full-text of papers over performance. We utilize publicly available datasets in our experiments: Kaggle for the full-text of papers, and Microsoft Academic Graph for extracting the abstract, metadata features and the initial year citation counts of papers. Our obtained results show that the use of full-text leads to higher accuracy, yet with an enormous trade-off on training time. Additionally, paper abstracts are easier to access as compared to the full-text. Finally, our model predicts that this paper will receive at least one citation during its initial year of publication.