Towards Robust Travel Time Estimation: An Out-of-Distribution Generalization Approach

Travel is becoming increasingly convenient with the development of the Internet. Travel Time Estimation (TTE) serves as a fundamental task for online traffic services. However, it faces a pressing challenge due to the volatile nature of traffic: the out-of-distribution (OOD) problem. In this paper, we investigate the OOD generalization problem, with a specific focus on the TTE task. We analyze the underlying generative process of traffic data by constructing a relational Structural Causal Model (SCM). We reveal that the complex causal relations can be simplified through a technique we term selective blocking. Based on this simplification, we propose a two-step deconfounding procedure to eliminate the spurious correlation of the environment on the invariant representation. Specifically, we design an invariant representation model, OOD model for Travel Time Estimation (OOD4TTE). First, our model infers potential environments for data generation through an environmental inference module. Then, we implement the two-step deconfounding procedure by contrastive learning and an invariant gating network. Finally, we generate robust travel time estimates based on the invariant representations. We demonstrate the superior generalization capability of OOD4TTE through comprehensive experiments on two real-world datasets.