Efficient Traffic Prediction Through Spatio-Temporal Distillation

Authors: Qianru Zhang, Xinyi Gao, Haixin Wang, Siu Ming Yiu, Hongzhi Yin

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments verify that Light ST significantly speeds up traffic flow predictions by 5X to 40X compared to state-of-the-art spatio-temporal GNNs, all while maintaining superior accuracy.
Researcher Affiliation Academia Qianru Zhang 1, Xinyi Gao 2, Haixin Wang 3, Siu-Ming Yiu 1*, Hongzhi Yin 2 1 The University of Hong Kong 2 The University of Queensland 3 University of California, Los Angles
Pseudocode Yes The training process of our Light ST is elaborated in Algorithm 1 in Appendix A.1.
Open Source Code Yes Our codes are available at: https: //github.com/lizzyhku/TP/tree/main.
Open Datasets Yes In this study, we conduct a series of experiments using real-life traffic flow datasets from California, specifically the PEMS3, PEMS4, PEMS7, PEMS8 and Pe MS-Bay datasets released by (Song et al. 2020).
Dataset Splits No The paper mentions datasets used but does not explicitly provide training/validation/test dataset splits. It states 'The traffic data is aggregated into 5-minute time intervals, resulting in 12 points of data per hour.' but no splitting methodology.
Hardware Specification Yes We conduct the experiments on a server with 10 cores of Intel(R) Core(TM) i9-9820X CPU @ 3.30GHz, 64.0GB RAM, and 4 Nvidia Ge Force RTX 3090 GPU.
Software Dependencies No The paper describes the models and architectures (GNNs, TCNs, MLPs) but does not provide specific software dependencies or their version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes We present our results on Pe MSD8 and Pe MSD3 datasets in terms of MAE and RMSE in Figure 4. We summarie our observations as follows: 1) Figure 4 show the effect of the number of MLP layers (ranging from {1, 2, 3, 4, 5}) and varying batch size (ranging from 23, 24, 25, 26, 27 ) on performance. Our framework, Light ST, achieves the best performance on Pe MSD8 and Pe MSD3 when the number of layers is 3 and the batch size is 32. ... 2) λ1, λ2 serve as loss weights to control how strongly our prediction-level and embedding-level restrict the joint model training.