Forecasting research influence: a recurrent neural network approach to citation prediction

Abstract

As the volume of scientific publications continues to proliferate, effective evaluation tools to determine the impact and quality of research articles are increasingly necessary. Citations serve as a widely utilized metric for gauging scientific impact. However, accurately prognosticating the long-term citation impact of nascent published research presents a formidable challenge due to the intricacy and unpredictability innate to the scientific ecosystem. Sophisticated machine learning methodologies, particularly recurrent neural networks (RNNs), have recently demonstrated promising potential in addressing this task. This research proposes an RNN architecture leveraging encoder-decoder sequence modeling capabilities to ingest historical chronicles and predict succeeding evolution via latent temporal dynamics learning. Comparative analysis between the RNN approach and baselines, including random forest, support vector regression, and multi-layer perceptron, demonstrate superior performance on unseen test data and rigorous k-fold cross-validation. On a corpus from Petra University, the RNN methodology attained the lowest errors (root mean squared error (RMSE) 1.84) and highest accuracy (0.91), area under the curve (AUC) (0.96), and F1-score (0.92). Statistical tests further verify significant improvements. The findings validate our deep learning solution's efficacy, robustness, and real-world viability for long-term scientific impact quantification to aid stakeholders in research evaluation. The findings intimate that RNN-based predictive modeling constitutes a potent technology for citation-driven scientific impact quantification.