Energy-based Epistemic Uncertainty for Graph Neural Networks
Authors: Dominik Fuchsgruber, Tom Wollschläger, Stephan Günnemann
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate GEBM over an extensive suite of datasets, distribution shifts, and baselines1. It consistently exhibits state-of-the-art performance in detecting out-of-distribution (o.o.d.) instances, while other approaches are only effective in a subset of settings. |
| Researcher Affiliation | Academia | Dominik Fuchsgruber, Tom Wollschläger, and Stephan Günnemann School of Computation, Information and Technology & Munich Data Science Institute Technical University of Munich, Germany {d.fuchsgruber, tom.wollschlaeger, s.guennemann}@tum.de |
| Pseudocode | No | The paper describes algorithms and methods in text but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We provide our code at cs.cit.tum.de/daml/gebm/ |
| Open Datasets | Yes | We use seven common benchmark datasets for node classification: The five citation datasets Cora ML[4], Cora ML-LLM[4, 78] [4], Citeseer [61, 25], Pub Med [51], Coauthor-Physics and Coauthor-Computers [63], and the co-purchase graphs Amazon Photos and Amazon Computers [44]. All datasets are taken from Py Torch Geometric [19]. They are licensed as C.C.0 1.0 (Cora ML, Cora ML LLM), C.C. Attribution-Non Commercial-Share Alike 3.0 (Citeseer), Odb L 1.0 (Pub Med). |
| Dataset Splits | Yes | All results were averaged over 5 splits and 5 initializations each (for standard deviations, see Appendix C). We evaluate o.o.d. detection metrics on a validation/test set that includes both i.d. and o.o.d. nodes. While the training and validation set are randomized for each split, the test set is shared across all splits to prevent data leakage. |
| Hardware Specification | Yes | We implement our models in Py Torch [58] and Py Torch Geometric [19] and train on two types of machines: (i) Xeon E5-2630 v4 CPU @ 2.20GHz with a NVIDA GTX 1080TI GPU and 128 GB of RAM. (ii) AMD EPYC 7543 CPU @ 2.80GHz with a NVIDA A100 GPU and 128 GB of RAM . |
| Software Dependencies | Yes | All models are trained with the ADAM optimizer [33] with a learning rate of 10-3, weight decay of 10-4, and a cross-entropy objective. We use early stopping on the validation loss with a patience of 50, an absolute improvement threshold of 10-1, and select the model with the best validation loss. We implement our models in Py Torch [58] and Py Torch Geometric [19] |
| Experiment Setup | Yes | At the backbone of all models, we use the same GCN [34] architecture if not specified explicitly otherwise. We use one hidden layer of dimension 64, symmetric normalization Appendix C.6, and add self-loops to the undirected (symmetric) adjacency matrix. We use Re LU nonlinearities, enable the bias term, and use dropout at p = 0.5. All models are trained with the ADAM optimizer [33] with a learning rate of 10-3, weight decay of 10-4, and a cross-entropy objective. We use early stopping on the validation loss with a patience of 50, an absolute improvement threshold of 10-1, and select the model with the best validation loss. |