Boosting the Cycle Counting Power of Graph Neural Networks with I$^2$-GNNs
Authors: Yinan Huang, Xingang Peng, Jianzhu Ma, Muhan Zhang
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This section aims to validate our theoretical results and study I2-GNNs empirical performance (https://github.com/Graph PKU/I2GNN.). Particularly, we focus on the following questions: Q1: Does the discriminative power of I2-GNNs increase compared to Subgraph MPNNs? Q2: Can I2-GNNs reach their theoretical counting power? Q3: How do I2-GNNs perform compared to MPNNs, Subgraph MPNNs and other state-of-the-art GNN models on open benchmarks for graphs? |
| Researcher Affiliation | Academia | Yinan Huang1, Xingang Peng1,2, Jianzhu Ma3, Muhan Zhang1 1Institute for Artificial Intelligence, Peking University 2School of Intelligence Science and Techology, Peking University 3Institute for AI Industry Research, Tsinghua University |
| Pseudocode | No | The paper includes mathematical equations for message passing but no clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | This section aims to validate our theoretical results and study I2-GNNs empirical performance (https://github.com/Graph PKU/I2GNN.). |
| Open Datasets | Yes | Datasets. To answer Q2, we adopt the synthetic dataset from Zhao et al. (2021) and a bioactive molecules dataset named Ch EMBL (Gaulton et al., 2012) to perform node-level counting tasks. ... For QM9, we adopt baselines including 1-GNN, 1-2-3-GNN (Morris et al., 2019), DTNN (Wu et al., 2018), Deep LRP (Chen et al., 2020), PPGN and NGNN. ... Dataset. We use the dataset provided by Dwivedi et al. (2020). It contains ZINC-12k and a ZINC-250k. ... Dataset. The ogbg-molhiv dataset is provided by Open Graph Benchmark (OGB). |
| Dataset Splits | Yes | The training/validation/test spliting is 0.3/0.2/0.5. ... The training/validation/test splitting ratio is 0.3/0.2/0.5 on synthetic dataset, and 0.6/0.2/0.2 on Ch EMBL dataset. ... The training/validation/test splitting ratio is 0.8/0.1/0.1. ... The training/validation/test splitting is given by the dataset. |
| Hardware Specification | No | No specific hardware details (GPU models, CPU models, etc.) were mentioned for running experiments. |
| Software Dependencies | No | The paper mentions software like GIN, Gated GNN, NNConv, GINE, but without specific version numbers. For example, 'Adopting GIN (Xu et al., 2018) as the base GNN' or 'We adopt 5-layer GINE (Hu et al., 2019) as base GNNs'. |
| Experiment Setup | Yes | Training details. The training/validation/test splitting ratio is 0.3/0.2/0.5 on synthetic dataset, and 0.6/0.2/0.2 on Ch EMBL dataset. We uniformly use Mean Absolute Error as loss function except 3-cycle. In the 3-cycle case we use Mean Square Error instead, as some training loss get stuck. We use Adam optimizer with initial learning rate 0.001, and use plateau scheduler with patience 10 and decay factor 0.9. We train 2,000 epochs for each models. The batch size is 256. ... We use Adam optimizer with initial learning rate 0.001, and use plateau scheduler with patience 10 and decay factor 0.95. We train 400 epochs with batch size 64 for each target separately. ... We use Adam optimizer with initial learning rate 0.001, and use plateau scheduler with patience 10 and decay factor 0.95. We train 1,000 epochs with batch size 256 on Zinc-12k, and train 800 epochs with batch size 256 on Zinc-250k. ... We use Adam optimizer with initial learning rate 0.001, and use step scheduler with step size 20 and decay factor 0.5. We train 50 epochs with batch size 64. |