Convolutional Kernel Networks for Graph-Structured Data

Authors: Dexiong Chen, Laurent Jacob, Julien Mairal

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate GCKN and compare its variants to state-of-the-art methods, including GNNs and graph kernels, on several real-world graph classification datasets, involving either discrete or continuous attributes.
Researcher Affiliation Academia 1Univ. Grenoble Alpes, Inria, CNRS, Grenoble INP, LJK, 38000 Grenoble, France 2Univ. Lyon, Universit e Lyon 1, CNRS, Laboratoire de Biom etrie et Biologie Evolutive UMR 5558, 69000 Lyon, France.
Pseudocode Yes Algorithm 1 Forward pass for multilayer GCKN
Open Source Code Yes Our code is freely available at https://github.com/claying/GCKN.
Open Datasets Yes We use the same benchmark datasets as in Du et al. (2019), including 4 biochemical datasets MUTAG, PROTEINS, PTC and NCI1 and 3 social network datasets IMDB-B, IMDB-MULTI and COLLAB. [...] All datasets and size information about the graphs can be found in Kersting et al. (2016). http://graphkernels.cs.tu-dortmund.de.
Dataset Splits Yes We follow the same protocols as (Du et al., 2019; Xu et al., 2019), and report the average accuracy and standard deviation over a 10-fold cross validation on each dataset. We use the same data splits as Xu et al. (2019), using their code.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper mentions using "the SVM implementation of the Cyanure toolbox (Mairal, 2019)" and "an Adam optimizer (Kingma & Ba, 2015)", but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes We use an Adam optimizer (Kingma & Ba, 2015) with the initial learning rate equal to 0.01 and halved every 50 epochs, and fix the batch size to 32. [...] We tune the bandwidth of the Gaussian kernel (identical for all layers), pooling operation (local (13) or global (14)), path size k1 at the first layer, number of filters (identical for all layers) and regularization parameter λ in (11).