LanczosNet: Multi-Scale Deep Graph Convolutional Networks

Authors: Renjie Liao, Zhizhen Zhao, Raquel Urtasun, Richard Zemel

ICLR 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We benchmark our model against several recent deep graph networks on citation networks and QM8 quantum chemistry dataset. Experimental results show that our model achieves the state-of-the-art performance in most tasks.
Researcher Affiliation Collaboration University of Toronto1, Uber ATG Toronto2, Vector Institute3, University of Illinois at Urbana-Champaign4, Canadian Institute for Advanced Research5
Pseudocode Yes Algorithm 1 : Lanczos Algorithm
Open Source Code Yes We implement all methods using Py Torch [65] and release the code at https://github.com/lrjconan/Lanczos Network.
Open Datasets Yes We test them on two sets of tasks: (1) semi-supervised document classification on 3 citation networks [63], (2) supervised regression of molecule property on QM8 quantum chemistry dataset [64].
Dataset Splits Yes We use the split provided by Deep Chem 2 which have 17428, 2179 and 2179 graphs for training, validation and testing respectively.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No The paper states 'We implement all methods using Py Torch [65]', but it does not specify the version number of PyTorch or any other software dependencies.
Experiment Setup Yes All methods are trained with Adam with learning rate 1.0e 2 and weight decay 5.0e 4. The maximum number of epoch is set to 200. Early stop with window size 10 is also adopted. We tune hyperparameters using Cora alone and fix them for citeseer and pubmed. For convolution based methods, we found 2 layers work the best. In GCN-FP, we set the hidden dimension to 64 and dropout to 0.5.