FairTraj: Density-Aware Generative Data Augmentation for Fairness in Downstream Trajectory Learning Tasks

Large-scale trajectory data underpin location-based services, yet their collection often exhibits pronounced spatial sampling bias, leading to performance gaps between dense and sparse regions and raising fairness concerns in downstream tasks. Existing approaches typically rely on training-time constraints or post-hoc adjustments; however, they are often model-specific and deliver unstable gains, motivating a model-agnostic pre-processing solution that debiases data at the source. Generative data synthesis is a natural direction, but trajectory fairness remains challenging because geographic space is continuous and current generators lack fine-grained spatial density awareness, limiting region-faithful generation. To bridge these gaps, we propose FairTraj, a density-aware generative data augmentation framework. To the best of our knowledge, FairTraj is the first model-agnostic pre-processing approach for improving fairness in downstream trajectory tasks. Specifically, FairTraj (i) leverages a hierarchical Quadtree and a novel Restricted Kernel Density Estimation (RKDE) to quantify spatial sampling bias as normalized trajectory-level density probabilities, enabling effective dense--sparse group separation, and (ii) designs a Density-Weighted Graph Attention Network (DW-GAT) and integrates its density-aware representations into a diffusion model via dual-level conditioning, thereby synthesizing high-fidelity trajectories for underrepresented regions. Extensive experiments on three real-world datasets across trajectory interpolation, generation, and travel time estimation demonstrate that FairTraj improves the inter-group disparity reduction ratio (IDRR) by 50% on average while preserving overall utility.