Agent-based Graph Neural Networks

Authors: Karolis Martinkus, Pál András Papp, Benedikt Schesch, Roger Wattenhofer

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide an extensive theoretical analysis of Agent Net: We show that the agents can learn to systematically explore their neighborhood and that Agent Net can distinguish some structures that are even indistinguishable by 2-WL. Moreover, Agent Net is able to separate any two graphs which are sufficiently different in terms of subgraphs. We confirm these theoretical results with synthetic experiments on hard-to-distinguish graphs and real-world graph classification tasks. In both cases, we compare favorably not only to standard GNNs but also to computationally more expensive GNN extensions.
Researcher Affiliation Collaboration Karolis Martinkus1, P al Andr as Papp2, Benedikt Schesch1, Roger Wattenhofer1 1ETH Zurich 2Computing Systems Lab, Huawei Zurich Research Center
Pseudocode No No pseudocode or clearly labeled algorithm block found. The model details are described in paragraph text within Section 3 and Appendix C.
Open Source Code Yes Code is available at https://github.com/Karolis Mart/Agent Net
Open Datasets Yes following Papp et al. [59] we use three bioinformatics datasets (MUTAG, PTC, PROTEINS) and two social network datasets (IMDB-BINARY and IMDB-MULTI) [76]. (...) We test this extension on the OGB-Mol HIV molecule classification dataset, which uses edge features [38]. (...) we test our Agent Net on the molecule property regression task on the QM9 dataset [63]. (...) We also test our model on a popular ZINC-12K molecular graph regression dataset [24; 40].
Dataset Splits Yes We follow the evaluation setup by Xu et al. [75], as was done by all of the baseline models. We perform a 10-fold cross-validation and report the mean and the standard deviation. (...) We follow the standard evaluation setup proposed by Hu et al. [38]. Similarly to the previous graph classification tasks, we set k to mean number of nodes (n 26), considered number of steps ℓ {8, 16}, batch size {32, 64} and hidden units {64, 128}. We train the model with 10 random seeds for 100 epochs, select the model with the best validation ROC-AUC and report the mean and the standard deviation.
Hardware Specification No The higher-order methods cannot even train on these large graphs on a 24GB GPU. This is the only mention of hardware, but it lacks specific model or processor details required for a 'Yes' classification.
Software Dependencies No The model is implemented using Py Torch [60] and Py Torch Geometric [29]. Specific version numbers for these software components are not provided.
Experiment Setup Yes For Agent Net in all of the tasks, we use Adam W optimizer [47] with a weight decay of 0.1. We set the initial learning rate to 10 4 and decay it over the whole training to 10 11 using a cosine schedule. We also clip the global gradient norm to 1. In all of the cases, Gumbel-Softmax temperature is set to 2 3. (...) we perform a grid search over the batch size {32, 128}, hidden units {32, 64, 128} and number of steps ℓ {8, 16}.