[5Yin-A] Poster 8

[5Yin-A-28] Forecast as Intervention: Causal Performative Prediction for Congestion Forecast

〇Akira Tanimoto¹, Koh Takeuchi², Hisashi Kashima² (1. Preferred Networks, Inc., 2. Kyoto University)

Keywords:

traffic,performativity,causal inference

Predictive problems are often treated as supervised learning tasks, wherein the prediction results are assumed to have no impact on the underlying phenomenon itself, implicitly assuming that the training and deployment distributions are identical. However, in forecasts of traffic congestion, providing information to users (drivers) can actually affect their behavior and thereby influence the congestion phenomenon itself. In this study, we treat forecast deployment as an intervention and propose a framework that accounts for this impact. Specifically, we propose: modeling forecast effects, optimizing self-consistent forecast values, developing learning methods that are robust to distribution shifts, and improving optimization stability through model constraints. We demonstrate experimentally that these approaches lead to improved forecast accuracy.