Generative Human Trajectory Recovery via Embedding-Space Conditional Diffusion

Authors: Kaijun Liu, Sijie Ruan, Liang Zhang, Cheng Long, Shuliang Wang, Liang Yu

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments based on two representative real-world mobility datasets are conducted, and the results show significant improvements (an average of 11% in recall) over the best baselines.
Researcher Affiliation Collaboration 1Alibaba-NTU Singapore Joint Research Institute, Singapore 2College of Computing and Data Science, Nanyang Technological University, Singapore 3School of Computer Science and Technology, Beijing Institute of Technology, China 4Alibaba Cloud, Hangzhou, Zhejiang, China.
Pseudocode Yes Algorithm 1 Imputation (Sampling) with Diff Move
Open Source Code No The code of this paper will be released in the link https://github.com/Kaijun L/Diff Move
Open Datasets Yes Foursquare2: This dataset (Yang et al., 2014) was obtained from the Foursquare API... Geolife3: This publicly available dataset is sourced from the Microsoft Research Asia Geolife project (Zheng et al., 2010)...
Dataset Splits Yes The trajectories are split chronologically into training (60%), validation (20%) and test (20%) sets.
Hardware Specification Yes Diff Move is trained using batch gradient descent with the Adam optimizer (Kingma & Ba, 2014), implemented in Python and Py Torch (Paszke et al., 2019), on a Linux server equipped with an NVIDIA RTX A5000.
Software Dependencies No Diff Move is trained using batch gradient descent with the Adam optimizer (Kingma & Ba, 2014), implemented in Python and Py Torch (Paszke et al., 2019), on a Linux server equipped with an NVIDIA RTX A5000.
Experiment Setup Yes We employed a learning rate of 0.001 with a weight decay of 1e-6. We set the location embedding size as 128, the steps (loops) of TGGNN as 2, the number of heads for cross attention as 4, diffusion step embedding dimension and temporal length embedding dimension are 128. We set the number of residual layers as 4, residual channels as 128, and attention heads for the temporal transformer as 8. We set the number of the diffusion step T = 50, the minimum noise level β1 = 0.0001, and the maximum noise level βT = 0.6. Multi-task learning weights λ1 and λ2 are set as 1 after experimental study.