Improving Graph Neural Networks by Learning Continuous Edge Directions
Authors: Seong Ho Pahng, Sahand Hormoz
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate through extensive experiments on both synthetic and real graph ensemble datasets that learning continuous edge directions significantly improves performance both for undirected and directed graphs compared with existing methods. |
| Researcher Affiliation | Academia | 1Department of Chemistry and Chemical Biology, Harvard University 2Department of Data Science, Dana-Farber Cancer Institute 3Department of Systems Biology, Harvard Medical School 4Broad Institute of MIT and Harvard EMAIL, sahand EMAIL |
| Pseudocode | No | The paper describes the model architecture and message-passing steps using mathematical equations (e.g., Equation 4 for feature update) and descriptive text, but it does not contain any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code is available on Git Hub. |
| Open Datasets | Yes | We benchmarked our method on eleven standard datasets including both undirected and directed graphs... We used the Replogle-gwps dataset (Replogle et al., 2022; Peidli et al., 2024)... We also modeled the traffic flow between Wikipedia articles using the Wiki Math dataset, which is classified as a static graph with temporal signals in the Py Torch Geometric Temporal library (Rozemberczki et al., 2021)... We applied Co ED to the optimal power flow (OPF) problem using the OPFData (Lovett et al., 2024) from the Py Torch Geometric library. |
| Dataset Splits | Yes | Texas, Wisconsin, Citeseer, and Cora datasets were downloaded using Py Torch Geometric library (Fey & Lenssen, 2019) with split= geom-gcn argument to use the 10 fixed 48%/32%/20% training/validation/test splits provided by (Pei et al., 2020). We downloaded AM-Computers and AMPhoto datasets from the same library but used the 60%/20%/20%-split file provided in the repository of Deng et al. (2024). We divided these 500 instances of feature-target pairs using 60%/20%/20% random training/validation/test split. |
| Hardware Specification | Yes | All experiments were performed on two NVIDIA RTX 6000 Ada Generation GPUs with 48GB of memory and one NVIDIA A100 Tensor Core GPU with 80GB, and it took roughly three weeks of training to produce the results. |
| Software Dependencies | No | The paper mentions using 'Py Torch Geometric library (Fey & Lenssen, 2019)', 'Scanpy software (Wolf et al., 2018)', and 'Py Torch Geometric Temporal library (Rozemberczki et al., 2021)', but it does not specify explicit version numbers for these or other crucial software dependencies like Python or PyTorch itself. |
| Experiment Setup | Yes | Training was early-stopped when the counter reached a patience of 200. Unless otherwise mentioned, we used the default hyperparameter settings of the respective models. We used the Re LU activation function and the ADAM optimizer in all experiments. Across all models, we searched over the following hyperparmeters: hidden dimension [16, 256], learning rate [5e-4, 2e-2], weight decay [0, 1e-2], dropout rate [0, 0.7], and the number of layers [2, 5]. |