The exponential growth in the number of publications and the prevalence of interdisciplinary research in recent years call for new ways to analyze how topics in science are evolving at large. In this paper, we propose a framework for topic discovery and evolution by representing the knowledge structure in a domain as a weighted dynamic network of co-occurring keywords in literature. Our approach finds research topics by detecting communities of keywords in the network. We tracked the evolution of topics in terms of the emergence, growth, shrinkage, survival, merging, splitting, and dissolution of their corresponding keyword communities over time. We then predicted how topics would evolve by treating this task as a supervised classification problem. A set of structural and temporal features were derived from the communities and were used to build classifiers to predict their future states. As a demonstration, this framework was applied to track and predict the evolution of topics from 19 million articles in PubMed that were published from 1980-2014. Our analysis revealed the underlying topic structure of this large scientific corpus at different points in history and identified significant topical developments and shifts in research foci within the field. The resulting predictive model based on the framework also achieved an accuracy of 83% in forecasting how the topics would evolve after five years.
과학문헌의 연구주제 진화에 대한 추적 및 예측 기법 연구
