Neural Networks and Political Science:Testing the Methodological Frontiers

Abstract

In recent years, a number of significant methodological re-evaluations have taken place in various disciplines of science due to machine learning develop-ments. This is particularly evident in STEM disciplines, while the behavioral and social sciences seem to approach these phenomena with some reserve. A good example is the use of artificial neural networks. Yet, acknowledging their characteristics, it can be safely assumed that they are relatively well designed to solve many problems in political science. This is due to the nature of many social phenomena that are characterized by at least three features: (1) their the-oretical basis is not ultimately determined, (2) they lack fully recognized func-tional relations, and (3) they are described by data that occur in a form that may be cumbersome for traditional modeling. Therefore, the article proceeds with some encouragement for the use of neural networks. At the same time, however, we need to proceed with caution. To mitigate possible opacity, a new political science-informed conceptualization of neural networks categorization scheme is proposed. This aims to help social scientists come to terms with one of the expo-nentially developing methods in the machine learning toolbox.