DropGNN: Random Dropouts Increase the Expressiveness of Graph Neural Networks

Authors: Pál András Papp, Karolis Martinkus, Lukas Faber, Roger Wattenhofer

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally validate our theoretical findings on expressiveness. Furthermore, we show that Drop GNNs perform competitively on established GNN benchmarks.
Researcher Affiliation Academia Pál András Papp ETH Zurich apapp@ethz.ch Karolis Martinkus ETH Zurich martinkus@ethz.ch Lukas Faber ETH Zurich lfaber@ethz.ch Roger Wattenhofer ETH Zurich wattenhofer@ethz.ch
Pseudocode No The paper does not include pseudocode or algorithm blocks.
Open Source Code Yes The code is publicly available1.
Open Datasets Yes We use the datasets from Sato et al. [29] that are based on 3 regular graphs. Nodes have to predict whether they are part of a triangle (TRIANGLES) or have to predict their local clustering coefficient (LCC). We test on the two counterexamples LIMITS 1 (Figure 2a) and LIMITS 2 from Garg et al. [11]
Dataset Splits Yes Following previous work [22; 21] the data is split into 80% training, 10% validation, and 10% test sets.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running experiments.
Software Dependencies No The paper mentions software but does not specify version numbers for key components (e.g., Python, PyTorch) that would be needed for replication.
Experiment Setup Yes Unless stated otherwise, we set the number of runs to m and choose the dropout probability to be p = 1/m, where m is the mean number of nodes in the graphs in the dataset.