summary: The authors propose an improved neural network architecture for predicting time series representing pairwise interacting entities. The architecture is based on a soft attention mechanism over pairwise relations among entities. pros: - addresses important challenge of relational reasoning - demonstrates improved empirical performance over relevant baselines on a quite wide range of datasets - ablation demonstrating the importance of individual model parts cons: - somewhat opaque presentation meta review: Paper with good empirical results on a relevant topic. Accept.