Towards Robust Trajectory Representations: Isolating Environmental Confounders with Causal Learning
Kang Luo, Yuanshao Zhu, Wei Chen, Kun Wang, Zhengyang Zhou, Sijie Ruan, Yuxuan Liang
TL;DR
TrajCL addresses confounding by geospatial context in trajectory representation learning by modeling an SCM and applying backdoor adjustment at the representation level, i.e., $P(H|do(X)) = \sum_{e\in \mathcal{E}} P(H|X,e) P(e)$. It introduces an Environment Alignment Module with an environment codebook and a Cross-Attention mechanism to disentangle causal and confounding factors, followed by a Causal Learning Module that performs disentangled learning and causal interventions. Experiments on GeoLife and Grab-Posisi show that TrajCL improves trajectory classification and generalization, with ablations and interpretability analyses demonstrating the contribution of each component. The approach is plug-and-play with existing trajectory models and offers interpretable insights into environmental contexts affecting movement.
Abstract
Trajectory modeling refers to characterizing human movement behavior, serving as a pivotal step in understanding mobility patterns. Nevertheless, existing studies typically ignore the confounding effects of geospatial context, leading to the acquisition of spurious correlations and limited generalization capabilities. To bridge this gap, we initially formulate a Structural Causal Model (SCM) to decipher the trajectory representation learning process from a causal perspective. Building upon the SCM, we further present a Trajectory modeling framework (TrajCL) based on Causal Learning, which leverages the backdoor adjustment theory as an intervention tool to eliminate the spurious correlations between geospatial context and trajectories. Extensive experiments on two real-world datasets verify that TrajCL markedly enhances performance in trajectory classification tasks while showcasing superior generalization and interpretability.