SynHAT: A Two-stage Coarse-to-Fine Diffusion Framework for Synthesizing Human Activity Traces

Human activity traces (HATs) play a crucial role in numerous real-world applications such as human mobility modeling, trace prediction, Point-of-Interest (POI) recommendation, and urban planning. However, increasing concerns over data privacy have significantly restricted access to authentic large-scale HATs. Fortunately, recent advances in generative AI open new opportunities to synthesize realistic yet privacy-preserving HATs that can support diverse applications. Despite this promise, two key challenges remain. (i) HATs (e.g., user-level POI check-in traces) are highly irregular and dynamic with long and variable time intervals, which makes it difficult to effectively capture their complex spatio-temporal patterns and intrinsic distributions. (ii) Generative models are typically computationally intensive and resource-demanding, such that generating long-term, fine-grained HATs incurs substantial computational overhead. To address these challenges, we propose SynHAT, a computationally-efficient coarse-to-fine HAT synthesis framework based on a novel spatio-temporal denoising diffusion model. In stage 1, we design a Coarse-grained Human Activity Diffusion model (Coarse-HADiff) to capture the overall spatio-temporal (ST) dependencies of the constructed coarse-grained latent ST traces, which includes a novel Latent Spatio-Temporal UNet (LST-UNet) for denoising through dual Drift-Jitter branches for jointly modeling smooth spatial transitions and temporal variations. In stage 2, we design a three-step pipeline consisting of Behavior Pattern Extraction, Fine-HADiff that shares the same architecture as Coarse-HADiff, and Semantic Alignment to further synthesize fine-grained latent ST traces based on the output from stage 1. We extensively evaluate the proposed SynHAT framework from diverse perspectives, including effectiveness for data fidelity, utility , and privacy , robustness , and scalability. Experimental results on real-world HATs from four cities (Tokyo, New York, Stockholm, and Austin) in three countries show that SynHAT significantly outperforms state-of-the-art baselines by 52% and 33% on spatial and temporal metrics, respectively.