TermBall: Tracking and Predicting Evolution Types of Research Topics by Using Knowledge Structures in Scholarly Big Data

The exponential growth in the number of publications and the prevalence of interdisciplinary research in recent years call for new approaches for analyzing how topics in science are evolving at large. This paper proposes TermBall, a framework that tracks and predicts fine-grained topic evolution in terms of the evolution types: emergence, growth, shrinkage, survival, merging, splitting, and dissolution. TermBall builds the knowledge structure, which is a weighted dynamic network of co-occurring keywords in the literature, and then discovers key topic structures that consist of keywords and their relationships by performing community detection methods. Based on the topic structures, TermBall provides two applications: (1) Retrospective application to identify topic evolution in the past and (2) Predictive application to forecast upcoming topic evolution type based on the structural and temporal features of the topic structures. For the evaluation, we built the knowledge structure by applying TermBall to 19 million articles in PubMed that were published from 1980 to 2014. We conducted qualitative analysis on the derived topic evolution types and quantitative analysis on the prediction results. As a result, our qualitative analysis reveals that TermBall is able to find various topic evolution types from the knowledge structure and also can predict how topics will evolve after five years with an accuracy of 83& x0025;.