Citation Count Prediction Using Different Time Series Analysis Models

Abstract

The paper helps to predict the future citation value of a fresh dataset of research papers by considering the past values of the citation count of paper using univariate time series analysis models and evaluating their performance through various evaluation metrics. It is important to predict future citation count as it helps to assess researcher's achievements, promotions, fund allocation, etc. This research is in addition to past research where for prediction, different parameters like content of paper, author details, venue impact etc. were considered. The real and original data for the dataset was extracted from the Google Scholar profile of top ranked authors. Three models of time series, Autoregressive Integrated moving average(ARIMA), Simple exponential smoothing (SES), and Holt winter's exponential Smoothing (HWES) are applied to observe the result variations. The models obtained error metric values for the complete dataset. All four-evaluation metrics were calculated. The best results for the predictions for citation count were obtained from the Simple exponential smoothing and Holt winter's exponential Smoothing models, whose values were almost the same for all evaluation metrics because of almost no change in formula. Among all fourerror metrics mentioned in the design, MASE gave sensible results, with almost all values being less than 1. The results showed similar graphs for both Simple exponential smoothing and Holt winter's exponential smoothing models for actual and predicted values of citation count as there is negligible difference in formula.