Scalable Spatiotemporal Graph Neural Networks

Authors: Andrea Cini, Ivan Marisca, Filippo Maria Bianchi, Cesare Alippi

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

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical results on relevant datasets show that our approach achieves results competitive with the state of the art, while dramatically reducing the computational burden.
Researcher Affiliation Academia The Swiss AI Lab IDSIA, Universit a della Svizzera italiana 2 Ui T the Arctic University of Norway 3 NORCE Norwegian Research Centre 4 Politecnico di Milano
Pseudocode No The paper describes the methods using mathematical equations and prose but does not include a clearly labeled pseudocode or algorithm block.
Open Source Code Yes We provide an efficient open-source implementation of SGP together with the code to reproduce all the experiments3. 3https://github.com/Graph-Machine-Learning-Group/sgp
Open Datasets Yes In the first experiment we consider the METR-LA and PEMS-BAY datasets (Li et al. 2018), which are popular medium-sized benchmarks... The first dataset contains data coming from the Irish Commission for Energy Regulation Smart Metering Project (CER-E; Commission for Energy Regulation 2016)... The second large-scale dataset is obtained from the synthetic PV-US4 dataset (Hummon et al. 2012)... 4https://www.nrel.gov/grid/solar-power-data.html
Dataset Splits Yes We use the same preprocessing steps of previous works to extract a graph and obtain train, validation and test data splits (Wu et al. 2019). ...for both datasets, we consider the first 6 months of data (4 for months for training, 1 month for validation, and 1 month for testing).
Hardware Specification Yes The time required to encode the datasets with SGP s encoder ranges from tens of seconds to 4 minutes on an AMD EPYC 7513 processor with 32 parallel processes. ...we measure the time required for the update step of each model on an NVIDIA RTX A5000 GPU... Nvidia Corporation for the donation of two GPUs.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x, CUDA 11.x).
Experiment Setup Yes In particular, each model is trained to predict the 12-step-ahead observations. ...for both datasets, we consider the first 6 months of data (4 for months for training, 1 month for validation, and 1 month for testing). ...we fix a maximum GPU memory budget of 12 GB and select the batch size accordingly.