Graph Neural Networks with Adaptive Readouts
Authors: David Buterez, Jon Paul Janet, Steven J Kiddle, Dino Oglic, Pietro Liò
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our empirical results demonstrate the effectiveness of neural readouts on more than 40 datasets spanning different domains and graph characteristics. Moreover, we observe a consistent improvement over standard readouts (i.e., sum, max, and mean) relative to the number of neighborhood aggregation iterations and different convolutional operators. We perform a series of experiments1 to evaluate the effectiveness of the adaptive and differentiable neural readouts presented in Section 2.1 relative to the standard pooling functions (i.e., sum, max, mean) used for the aggregation of node features into a graph-level representation. |
| Researcher Affiliation | Collaboration | David Buterez 1 Jon Paul Janet 2 Steven J. Kiddle 3 Dino Oglic 3 Pietro Liò 1 1 Department of Computer Science and Technology, University of Cambridge, UK 2 CVRM, Bio Pharmaceuticals R&D, Astra Zeneca, Sweden 3 DS&AI, Bio Pharmaceuticals R&D, Astra Zeneca, UK |
| Pseudocode | No | The paper does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code is available at https://github.com/davidbuterez/gnn-neural-readouts. |
| Open Datasets | Yes | To showcase the potential for learning effective representation models over graph structured data, we use in excess of 40 datasets originating from different domains such as quantum mechanics, biophysics, bioinformatics, computer vision, social networks, synthetic graphs, and function call graphs. QM9 dataset and ENZYMES dataset are frequently mentioned and QM9 is cited as "Zhenqin Wu, Bharath Ramsundar, Evan N. Feinberg, Joseph Gomes, Caleb Geniesse, Aneesh S. Pappu, Karl Leswing, and Vijay Pande. Moleculenet: a benchmark for molecular machine learning. Chem. Sci., 9:513 530, 2018." |
| Dataset Splits | Yes | computed by averaging over five random splits of the data. Data splits, hyperparameters, how they were chosen? [Yes] In Appendices B, C and F and the associated code. |
| Hardware Specification | No | The paper states in its checklist 'Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] Please see Appendix P.' However, Appendix P is not provided in the given text, so specific hardware details cannot be extracted. |
| Software Dependencies | No | The paper mentions 'RDKit' and discusses software in general terms but does not provide specific version numbers for software dependencies (e.g., 'Python 3.8, PyTorch 1.9, and CUDA 11.1'). |
| Experiment Setup | Yes | A detailed experimental setup (including loss functions, reporting metrics, and other details relevant for reproducibility) has been provided in Appendices B and F. For the adaptive readouts, the dimensionality of the graph representation is a hyperparameter of the model, which was fixed to 64. In our experiments, we also apply Bernoulli dropout with rate p = 0.4 as the last operation within MLP. We trained both non-variational and guided variational graph autoencoder [19, 20] models (denoted by GNN and VGAE, respectively) for 200 epochs. |