Swarming network inference with importance clustering of relative interactions

Swarming is central to many problems in physics, biology, and engineering where collective motion and cooperation emerge through interactions of many agents. As a result, inferring network connections underlying swarms from time series data is an important problem with broad application. In this work, we propose a method based on phase-space regression of the observed dynamics with random forest models, from which relative interactions are clustered according to their Gini importance with respect to a given agent. Network connections are inferred by comparing how the statistics of the strongly and weakly important clusters overlap. Because the method entails fitting the dynamics and finding the most important interactions for each agent individually with general swarming assumptions, high accuracy can be maintained with relatively modest amounts of observation data and utilizing only a small number of generalizable hyperparameters across a variety of behavioral patterns, diverse physical ingredients, and heterogeneous topologies.