An End-to-End Deep Learning Architecture for Graph Classification

Authors: Muhan Zhang, Zhicheng Cui, Marion Neumann, Yixin Chen

AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on benchmark graph classification datasets demonstrate that the proposed architecture achieves highly competitive performance with state-of-the-art graph kernels and other graph neural network methods.
Researcher Affiliation Academia Department of Computer Science and Engineering, Washington University in St. Louis
Pseudocode No The paper provides mathematical formulations and descriptions of its architectural components (e.g., graph convolution layer), but it does not include any clearly labeled 'Pseudocode' or 'Algorithm' blocks with structured, step-by-step procedures.
Open Source Code Yes The code and data are available at https://github.com/muhanzhang/DGCNN.
Open Datasets Yes We use five benchmark bioinformatics datasets to compare the graph classification accuracy of DGCNN with graph kernels. The datasets are: MUTAG, PTC, NCI1, PROTEINS, D&D. We include detailed dataset information in the supplementary material.
Dataset Splits Yes Following the conventional settings, we performed 10-fold cross validation with LIBSVM (Chang and Lin 2011) (9 folds for training and 1 fold for testing) using one training fold for hyperparameter searching, and repeated the experiments for 10 times (thus 100 runs per dataset).
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as CPU or GPU models, or memory specifications. It does not mention any cloud or cluster resources with their specifications.
Software Dependencies No The paper mentions software like LIBSVM, Torch, and the ADAM updating rule, but it does not specify any version numbers for these software components, which is necessary for reproducibility.
Experiment Setup Yes The network has four graph convolution layers with 32, 32, 32, 1 output channels, respectively... The remaining layers consist of two 1-D convolutional layers and one dense layer. The first 1-D convolutional layer has 16 output channels followed by a Max Pooling layer with filter size 2 and step size 2. The second 1-D convolutional layer has 32 output channels, filter size 5 and step size 1. The dense layer has 128 hidden units followed by a softmax layer as the output layer. A dropout layer with dropout rate 0.5 is used after the dense layer. We used the hyperbolic tangent function (tanh) as the nonlinear function in graph convolution layers, and rectified linear units (Re LU) in other layers. Stochastic gradient descent (SGD) with the ADAM updating rule (Kingma and Ba 2014) was used for optimization. The only hyperparameters we optimized are the learning rate and the number of training epochs (details in the supplementary material).