Causal Abstraction Via Emergence for Predicting Bilateral Trade
Proceedings of The 2021 Causal Analysis Workshop Series, PMLR 160:39-51, 2021.
Causal abstraction is key in finding efficient representations of noisy and complex systems, for decision-making and prediction of future system states. Hand-crafted causal abstractions, although accurate and interpretable, can be costly to construct and cannot generalize to large, novel datasets. In this paper, we explore the information-theoretic concept of causal emergence, its correspondence to recent definitions of causal abstraction, and the properties of emergent representations that enable more accurate state predictions and semantic interpretations. Using the bilateral trade network as a case study, we enumerate the conditions under which trade agreements exhibit causal emergence properties, and show that causally emergent representations are indeed able to provide better prediction capability than original trade network representations in a variety of cases.