Fast and Deep Graph Neural Networks

Authors: Claudio Gallicchio, Alessio Micheli3898-3905

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

Reproducibility Variable Result LLM Response
Research Type Experimental Through experimental results, we show that even without training of the recurrent connections, the architecture of small deep GNN is surprisingly able to achieve or improve the state-of-the-art performance on a significant set of tasks in the field of graphs classification.
Researcher Affiliation Academia Claudio Gallicchio, Alessio Micheli Department of Computer Science, University of Pisa Largo B. Pontecorvo, 3. 56127 Pisa, Italy {gallicch, micheli}@di.unipi.it
Pseudocode No The paper describes the model and process steps in text and equations, but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is made available online2. 2https://sites.google.com/site/cgallicch/FDGNN/code
Open Datasets Yes All the used datasets are publicly available online (Kersting et al. 2016).
Dataset Splits Yes The performance on the graph classification tasks was assessed in terms of accuracy, and it was evaluated through a process of stratified 10-fold cross validation. For each fold, the FDGNN hyper-parameter configuration was chosen by model selection, by means of a nested level of stratified 10-fold cross validation applied on the corresponding training samples.
Hardware Specification Yes a MATLAB implementation of FDGNN, running on a system with Intel Xeon Processor E5-263L v3 with 168 GB of RAM.
Software Dependencies No The paper states 'a MATLAB implementation of FDGNN' but does not provide specific version numbers for MATLAB or any other software dependencies.
Experiment Setup Yes We adopted a simple general setting, where all the hidden layers of the architecture in the graph embedding component shared the same values of the hyper-parameters, i.e., for i 1 we set: number of neurons H(i) = H and effective spectral radius ρ(i) = ρ; for i > 1 we set the inter-layer scaling ω(i) = ω. In particular, we set the number of neurons in each hidden layer to H = 50 for all datasets except for NCI1 and COLLAB, for which we used H = 500. We implemented very sparse weight matrices with C = 1... Values of ρ, ω(1) and ω were explored in the same range (0, 1). The Tikhonov regularizer for training the readout was searched in a log-scale grid in the range 10 8 103. For the graph embedding convergence process in each layer we used a threshold of ϵ = 10 3, and maximum number of iterations ν = 50. The projection dimension for the readout was set to twice the number of neurons in the last hidden layer, i.e., P = 2 H.