Easy Begun Is Half Done: Spatial-Temporal Graph Modeling with ST-Curriculum Dropout

Authors: Hongjun Wang, Jiyuan Chen, Tong Pan, Zipei Fan, Xuan Song, Renhe Jiang, Lingyu Zhang, Yi Xie, Zhongyi Wang, Boyuan Zhang

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our strategy can be applied to any canonical deep learning architecture without extra trainable parameters, and extensive experiments on a wide range of datasets are conducted to illustrate that, by controlling the difficulty level of ST relations as the training progresses, the model is able to capture better representation of the data and thus yields better generalization.
Researcher Affiliation Collaboration 1 SUSTech-UTokyo Joint Research Center on Super Smart City, Southern University of Science and Technology 2 Research Institute of Trustworthy Autonomous Systems, Southern University of Science and Technology 3 Department of Physics, The Chinese University of Hong Kong 4 Center for Spatial Information Science, University of Tokyo 5 Information Technology Center, University of Tokyo 6 Huawei Technologies CO.LTD 7 Didichuxing Inc
Pseudocode No No explicitly labeled pseudocode or algorithm blocks were found in the paper.
Open Source Code No The paper mentions Faiss (https://github.com/facebookresearch/faiss) as a tool to accelerate computing, but it does not provide a link or statement about the availability of their own source code for the proposed methodology.
Open Datasets Yes In this section, we evaluate our proposed STC-Dropout on a wide range of spatial-temporal datasets so as to demonstrate its general applicability: METR-LA (Li et al. 2017), PEMSD7M (Yu, Yin, and Zhu 2017), Covid-19 (Panagopoulos, Nikolentzos, and Vazirgiannis 2021) and NYC-Crime (Xia et al. 2021).
Dataset Splits No Please refer to the appendix A.2-A.9 for more detailed descriptions of the datasets, baselines, experiment setup and extra results.
Hardware Specification No No specific hardware (e.g., GPU models, CPU types, or memory) used for the experiments is mentioned in the paper.
Software Dependencies No The paper does not provide specific version numbers for software dependencies or libraries used in the experiments.
Experiment Setup Yes Last but not least, we explore the hyper-parametersensitivity to ρ, which controls the radius of the d(l)-dimensional ball, and α, which specifies the initial retain rate of nodes. The results of hyperparameter search from ρ α [0.05, 0.10, , 0.95] [0.1, 0.3, 0.9] on ST-GCN using both Pe MSD7M and METR-LA datasets are depicted in Figure 3. As we can see, the model performance is negatively related to the value of α. Thus, it is recommended to set its value to be relatively small to ensure only the easy samples can be retained in the early model training. Concerning ρ, we encourage maintaining ρ [0.25, 0.4] for desirable performances.