Mapping the physics research space: a machine learning approach

Abstract

Abstract Scientific discoveries do not occur in vacuum but rather by connecting existing pieces of knowledge in new and creative ways. Mapping the relation and structure of scientific knowledge is therefore central to our understanding of the dynamics of scientific production. Here we introduce a new approach to generate scientific knowledge maps based on a machine learning approach that, starting from the observed publication patterns of authors, generates an N-dimensional space where it is possible to measure the similarity or distance between different research topics and knowledge domains. We provide an implementation of the proposed approach that considers the American Physical Society publications database and generates a map of the research space in Physics that characterizes the relation among research topics over time. We use this map to measure two indicators, the research capacity fingerprint and the knowledge density, to profile the research activity in physical sciences of more than 400 urban areas across the world. We show that these indicators can be used to analyze and predict the evolution over time of the research capacity and specialization of specific geographical areas. Furthermore we provide an extensive analysis of the relation between socio-economic development indicators and the ability to produce new knowledge for 67 countries, as measured by our approach, highlighting some key correlates of scientific production capacity. The proposed approach is scalable to very large datasets and can be extended to study other disciplines and research areas without having to rely on ad-hoc science classification schemes.

Keywords

• Embeddings
• Research space
• Principle of relatedness
• Fingerprinting scientific production
• Revealed comparative advantage
• Predicting research specialization