Deciphering the social interactions that govern collective behavior in animal societies has greatly benefited from advancements in modern computing. Computational models diverge into two kinds of approaches: analytical models and machine learning models. This work introduces a deep learning model for social interactions in the fish species Hemigrammus rhodostomus, and compares its results to experiments and to the results of a state-of-the-art analytical model. To that end, we propose a systematic methodology to assess the faithfulness of a model, based on the introduction of a set of stringent observables. We demonstrate that machine learning models of social interactions can directly compete against their analytical counterparts. Moreover, this work demonstrates the need for consistent validation across different timescales and highlights which design aspects critically enables our deep learning approach to capture both short- and long-term dynamics. We also show that this approach is scalable to other fish species.