[arxiv]
Human mobility prediction faces two key challenges: diverse trajectory complexity impedes efficient training, and focusing solely on next-location prediction ignores auxiliary mobility characteristics. This paper proposes an entropy-driven curriculum learning framework combined with multi-task training. The approach uses Lempel-Ziv compression to quantify trajectory predictability and organizes training from simple to complex patterns for faster convergence. Multi-task learning simultaneously predicts locations, movement distances, and directions to capture realistic mobility patterns. Experiments on the HuMob Challenge dataset show a GEO-BLEU score of 0.354 and a DTW distance of 26.15, with up to 2.92x faster convergence compared to conventional training.
git clone --recursive https://github.com/tum-bgd/2025-ICML-1DPathConv.git
The required environment is quite easy. We only need a PyTorch
environment (tested in v2.6) with pandas==2.3.2
.
tba
Licensed under Apache-2.0 license (LICENSE or https://opensource.org/licenses/Apache-2.0)