Graphon Neural Networks and the Transferability of Graph Neural Networks
Authors: Luana Ruiz, Luiz Chamon, Alejandro Ribeiro
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 6, transferability of GNNs is illustrated in two numerical experiments. |
| Researcher Affiliation | Academia | Luana Ruiz Dept. of Electrical and Systems Eng. University of Pennsylvania Philadelphia, PA 19143 rubruiz@seas.upenn.edu Luiz F. O. Chamon Dept. of Electrical and Systems Eng. University of Pennsylvania Philadelphia, PA 19143 luizf@seas.upenn.edu Alejandro Ribeiro Dept. of Electrical and Systems Eng. University of Pennsylvania Philadelphia, PA 19143 aribeiro@seas.upenn.edu |
| Pseudocode | No | The paper describes algorithms and architectures mathematically but does not include any pseudocode blocks or algorithms labeled as such. |
| Open Source Code | Yes | We use the GNN library available at https://github.com/alelab-upenn/graph-neural-networks and implemented with Py Torch. |
| Open Datasets | Yes | To illustrate Theorem 2 in a graph signal classification setting, we consider the problem of movie recommendation using the Movie Lens 100k dataset (Harper and Konstan, 2016). |
| Dataset Splits | Yes | This data is then split between 90% for training and 10% for testing, with 10% of the training data used for validation. |
| Hardware Specification | No | The paper mentions funding from 'Intel Dev Cloud' in the acknowledgments but does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used for the experiments. |
| Software Dependencies | Yes | We use the GNN library available at https://github.com/alelab-upenn/graph-neural-networks and implemented with Py Torch. |
| Experiment Setup | Yes | This GNN has L = 1 convolutional layer with F = 32 and K = 5, followed by a readout layer at node 405 that maps its features to a one-hot vector of dimension C = 5 (corresponding to ratings 1 through 5). To generate the input data, we pick the movies rated by user 405 and generate the corresponding movie signals by "zero-ing" out the ratings of user 405 while keeping the ratings given by other users. This data is then split between 90% for training and 10% for testing, with 10% of the training data used for validation. Only training data is used to build the user network in each split. To analyze transferability, we start by training GNNs Φ(Hn; Sn; xn) on user subnetworks consisting of random groups of n = 100, 200, . . . , 900 users, including user 405. We optimize the cross-entropy loss using ADAM with learning rate 10 3 and decaying factors β1 = 0.9 and β2 = 0.999, and keep the models with the best validation RMSE over 40 epochs. |