Interpretable and Generalizable Graph Learning via Stochastic Attention Mechanism

Authors: Siqi Miao, Mia Liu, Pan Li

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on eight datasets show that GSAT outperforms the state-of-the-art methods by up to 20% in interpretation AUC and 5% in prediction accuracy.
Researcher Affiliation Academia 1Department of Computer Science, Purdue University, West Lafayette, USA 2Department of Physics and Astronomy, Purdue University, West Lafayette, USA. Correspondence to: Siqi Miao <miao61@purdue.edu>, Pan Li <panli@purdue.edu>.
Pseudocode No The paper does not contain a pseudocode block or a clearly labeled algorithm block.
Open Source Code Yes Our code is available at https: //github.com/Graph-COM/GSAT.
Open Datasets Yes Mutag (Debnath et al., 1991) is a molecular property prediction dataset... OGBG-Molhiv (Wu et al., 2018; Hu et al., 2020) is a molecular property prediction datasets...
Dataset Splits Yes For Ba-2Motifs, we split it randomly into three sets (80%/10%/10%). For Mutag, we split it randomly into 80%/20% to train and validate models... For MNIST-75sp, we use the default splits given by (Knyazev et al., 2019)... For Graph-SST2, Spurious-Motifs and OGBG-Mol, we use the default splits given by (Yuan et al., 2020b) and (Wu et al., 2022).
Hardware Specification Yes IB-subgraph needs 40 hours to train 100 epochs for 1 seed on Spurious-Motif and 150 hours for OGBG-Molhiv on a Quadro RTX 6000.
Software Dependencies No The paper mentions software components like GIN, PNA, and Gumbel-softmax but does not provide specific version numbers for any of them, nor for programming languages or other libraries.
Experiment Setup Yes We use a two-layer GIN (Xu et al., 2019) with 64 hidden dimensions and 0.3 dropout ratio. We use the setting from (Corso et al., 2020) for PNA, which has 4 layers with 80 hidden dimensions, 0.3 dropout ratio, and no scalars are used... All datasets use a batch size of 128; except for MNIST-75sp we use a batch size of 256... GIN uses 0.003 learning rate for Spurious-Motifs and 0.001 for all other datasets... Ba-2Motif and Mutag use r = 0.5, and all other datasets use r = 0.7... β is not tuned and is set to 1 |E| for all datasets... Temperature used in the Gumbel-softmax trick (Jang et al., 2017) is not tuned, and we use 1 for all datasets.