FlashST: A Simple and Universal Prompt-Tuning Framework for Traffic Prediction

Authors: Zhonghang Li, Lianghao Xia, Yong Xu, Chao Huang

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical evaluations demonstrate the effectiveness of our Flash ST across different spatio-temporal prediction tasks using diverse urban datasets. and We conduct extensive experiments on four distinct types of spatio-temporal data tasks to evaluate the effectiveness of our proposed framework.
Researcher Affiliation Academia Zhonghang Li 1 2 Lianghao Xia 2 Yong Xu 1 3 Chao Huang* 2 1South China University of Technology 2University of Hong Kong 3PAZHOU LAB.
Pseudocode No The paper describes its methodology using textual descriptions and mathematical equations but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Code is available at https://github.com/HKUDS/Flash ST.
Open Datasets Yes In the pre-training phase, we use four datasets (PEMS03, PEMS04, PEMS07, and PEMS08 (Song et al., 2020)) as our training sets. [...] During the subsequent prompt-tuning phase, we concentrate on four distinct target datasets to fine-tune and evaluate our framework: PEMS07(M) (Yu et al., 2018), CA-D5 (Liu et al., 2023b), Cheng Du-DIDI (Lu et al., 2022), and NYC Citi Bike (Ye et al., 2021).
Dataset Splits Yes Each target dataset is partitioned into training, validation, and test sets in a ratio of 6:2:2.
Hardware Specification No The paper discusses training time and computational efficiency, but it does not provide specific details about the hardware used, such as GPU or CPU models.
Software Dependencies No The paper does not provide specific software dependency details, such as programming languages, libraries, or frameworks with their version numbers.
Experiment Setup Yes We configure the model with 32 hidden units for d, dt, and dr, while employing a spatial and temporal encoder with 2 layers. The pre-training phase involves alternating training for 300 epochs on the four pre-training datasets. Subsequently, we conduct fine-tuning for 20 epochs during the prompt-tuning phase. In both the baseline and full fine-tuning setups, the maximum number of training epochs is limited to 100. Furthermore, a batch size of 64 is employed for both phases and all baselines.