Template based Graph Neural Network with Optimal Transport Distances
Authors: Cédric Vincent-Cuaz, Rémi Flamary, Marco Corneli, Titouan Vayer, Nicolas Courty
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically validate our claim on several synthetic and real life graph classification datasets, where our method is competitive or surpasses kernel and GNN state-of-the-art approaches. We complete our experiments by an ablation study and a sensitivity analysis to parameters. |
| Researcher Affiliation | Academia | Cédric Vincent-Cuaz Univ. Côte d Azur, INRIA, CNRS, LJAD F-06100 Nice cedric.vincent-cuaz@inria.fr Rémi Flamary IP Paris, CMAP, UMR 7641 F-91120 Palaiseau remi.flamary@polytechnique.edu Marco Corneli Univ. Côte d Azur, INRIA, CNRS, LJAD F-06100 Nice marco.corneli@inria.fr Titouan Vayer Univ. Lyon, INRIA, CNRS, ENS de Lyon LIP UMR 5668, F-69342 Lyon titouan.vayer@inria.fr Nicolas Courty Univ. Bretagne-Sud, CNRS, IRISA F-56000 Vannes nicolas.courty@irisa.fr |
| Pseudocode | No | No explicitly labeled pseudocode or algorithm blocks were found in the paper. |
| Open Source Code | Yes | Code available at https://github.com/cedricvincentcuaz/TFGW. |
| Open Datasets | Yes | We use 8 well-known graph classification datasets [24]: 5 bioinformatics datasets among which 3 have discrete node features (MUTAG, PTC, NCI1 [28, 52]) and 2 have continuous node features (ENZYMES, PROTEINS[6]) and 3 social network datasets (COLLAB, IMDB-B, IDBM-M [69]). |
| Dataset Splits | Yes | We suggest here to quantify the generalization capacities of GNN based models by performing a 10-fold cross validation with a holdout test set never seen during training. For each split, we track the accuracy on the validation fold every 5 epochs, then the model whose parameters maximize that accuracy is retained. Finally, the model used to predict on the holdout test set is the one with maximal validation accuracy averaged across all folds. |
| Hardware Specification | Yes | These measures were taken on CPUs (Intel Core i9-9900K CPU, 3.60 GHz) in order to fairly compare the numerical complexity of these methods, as OT solver used in TFGW and OT-GNN are currently limited to these devices (see the detailed discussion in Section 2.1). |
| Software Dependencies | No | We used Pytorch [45] to implement the model. The FGW distances are computed by adapting the conditional gradient solver implemented in the POT toolbox [18]. |
| Experiment Setup | Yes | For all template based models, we set the size of the templates to the median size of the observed graphs. We validate the number of templates K in {β|Y|}β, with β {2, 4, 6, 8} and |Y| the number of classes. All parameters of our TFGW layers highlighted in red in Figure 1 are learned while φu is a GIN architecture [67] composed of L = 2 layers aggregated using the Jumping Knowledge scheme [68]. For OT-GNN we validate the number of GIN layers in L {2, 4}. Finally for fairness, we validate the number of hidden units within the GNN layers and the application of dropout on the final MLP for predictions, similarly to GIN and Drop GIN. |