Taming the Long Tail in Human Mobility Prediction
Authors: Xiaohang Xu, Renhe Jiang, Chuang Yang, zipei fan, Kaoru Sezaki
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments with two real-world trajectory datasets demonstrate that Lo TNext significantly surpasses existing state-of-the-art works. and We evaluate our Lo TNext on two publicly available real-world LBSN datasets: Gowalla and Foursquare |
| Researcher Affiliation | Academia | Xiaohang Xu1, Renhe Jiang1 , Chuang Yang1, Zipei Fan1, Kaoru Sezaki1 1The University of Tokyo xhxu@g.ecc.u-tokyo.ac.jp {jiangrh, chuang.yang}@csis.u-tokyo.ac.jp {fanzipei, sezaki}@iis.u-tokyo.ac.jp |
| Pseudocode | Yes | Algorithm 1 Pseudo-code of training Lo TNext |
| Open Source Code | Yes | 2https://github.com/Yukayo/Lo TNext |
| Open Datasets | Yes | We evaluate our Lo TNext on two publicly available real-world LBSN datasets: Gowalla and Foursquare1 2 |
| Dataset Splits | No | We then split each user s check-in records according to temporal order, using the first 80% for training and the remaining 20% for testing. The paper does not explicitly state a validation dataset split for purposes like hyperparameter tuning or early stopping. |
| Hardware Specification | Yes | We implement Lo TNext using Py Torch 1.13.1 on a Linux server equipped with 384GB RAM, 10-core Intel(R) Xeon(R) Silver 4210R CPU @ 2.40GHz, and Nvidia RTX 3090 GPUs. |
| Software Dependencies | Yes | We implement Lo TNext using Py Torch 1.13.1 |
| Experiment Setup | Yes | The embedding dimensions for POIs and users are set to 10, and the time embedding dimension is set to 6. For the Transformer architecture, we incorporate two multi-head attention mechanisms and 2 encoder blocks. For the spatial decay rate β, we follow the settings of Flashback [43]. and The results, shown in Figure 6(a) for Gowalla and Figure 6(c) for Foursquare, indicate that Acc@1 and MRR remain stable across different values, with the optimal threshold identified as δ = 0.5. Next, we vary the logit adjustment weight τ from 1 to 2 in increments of 0.2 to test the model s performance in balancing class imbalances. Figure 6(b) and Figure 6(d) reveal that τ = 1.2 yields the best results on both datasets, suggesting a moderate adjustment weight helps generalize better without overly amplifying rare classes. |