Time-Frequency Disentanglement Boosted Pre-Training: A Universal Spatio-Temporal Modeling Framework

Authors: Yudong Zhang, Zhaoyang Sun, Xu Wang, Xuan Yu, Kai Wang, Yang Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments conducted on realworld datasets demonstrate that USTC significantly outperforms the advanced baselines in forecasting, imputation, and extrapolation across cities. We conduct extensive experiments on four real-world datasets, evaluating USTC on spatio-temporal forecasting, imputation, and extrapolation tasks.
Researcher Affiliation Academia 1 University of Science and Technology of China (USTC), Hefei, China 2 Suzhou Institute of Advanced Research, USTC, Suzhou, China 3 State Key Laboratory of Precision and Intelligent Chemistry, USTC, Hefei, China {zyd2020@mail., sunzhaoyang@mail., wx309@, yx2024@mail., zaizwk@mail., angyan@}ustc.edu.cn
Pseudocode No The paper describes the methodology in text and through architectural diagrams (Figure 1 and Figure 2), but it does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain any explicit statements about releasing source code, nor does it provide a link to a code repository or mention code in supplementary materials.
Open Datasets Yes Four real-world widely used datasets are employed to evaluate our proposed framework, including PEMS-BAY, METR-LA [Li et al., 2018], Chengdu, and Shenzhen. These datasets comprise several months of traffic flow information, with the statistics listed in Table 1.
Dataset Splits Yes The dataset is divided into three parts: pre-training data from three cities, few-shot fine-tuning data, and testing data from the other city. We use the comprehensive data from three cities for pre-training and select one city s data for both few-shot fine-tuning and testing. For instance, if Shenzhen is the city chosen for fine-tuning, the complete datasets from PEMS-BAY, METR-LA, and Chengdu are used for pre-training. A three-day dataset from Shenzhen is allocated for few-shot fine-tuning, while the rest of the data in Shenzhen is reserved for testing. We use 1-day historical data to predict future 1-hour data.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU models, CPU types, or memory) used for running the experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers.
Experiment Setup No The paper specifies task details like prediction horizons and missing data ratios (e.g., 'predicting the future 1-hour data based on 1-day historical data', 'randomly masking observed data with a ratio of 30%'), and metrics (MAE, RMSE). However, it does not provide specific hyperparameter values (e.g., learning rate, batch size, number of epochs, optimizer settings) in the main text.