Conformal Inductive Graph Neural Networks
Authors: Soroush H. Zargarbashi, Aleksandar Bojchevski
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6 EXPERIMENTAL EVALUATION |
| Researcher Affiliation | Academia | Soroush H. Zargarbashi CISPA Helmholtz Center for Information Security zargarbashi@cs.uni-koeln.de Aleksandar Bojchevski University of Cologne bojchevski@cs.uni-koeln.de |
| Pseudocode | Yes | Algorithm 1: Node Ex and Edge Ex CP for inductive node classification |
| Open Source Code | Yes | The code is accessible at github.com/soroushzargar/conformal-node-classification. |
| Open Datasets | Yes | We consider 9 different datasets and 4 models: GCN Kipf & Welling (2017), GAT Veliˇckovi c et al. (2018), and APPNPKlicpera et al. (2019) as structure-aware and MLP as a structure-independent model. We evaluate our Node Ex, and Edge Ex CP on the common citation graphs Cora ML Mc Callum et al. (2004), Cite Seer Sen et al. (2008), Pub Med Namata et al. (2012), Coauthor Physics and Coauthor CS Shchur et al. (2018), and co-purchase graphs Amazon Photos and Computers Mc Auley et al. (2015); Shchur et al. (2018) (details in E). |
| Dataset Splits | Yes | For any of the mentioned datasets, we sample 20 nodes per class for training and 20 nodes for validation with stratified sampling. |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., CPU/GPU models, memory, or cloud instance types) used for running the experiments. |
| Software Dependencies | No | The paper mentions software components like "Adam optimizer" and "categorical cross-entropy loss" but does not provide specific version numbers for these or other libraries/frameworks (e.g., PyTorch, TensorFlow, Python version). |
| Experiment Setup | Yes | For all architectures, we built one hidden layer of 64 units and one output layer. We applied dropout on the hidden layer with probability 0.6 for GCN, and GAT, 0.5 for APPNPNet, and 0.8 for MLP. For GAT we used 8 heads. We trained all models with categorical cross-entropy loss, and Adam optimizer with L2 regularization. |