Path Integral Based Convolution and Pooling for Graph Neural Networks

Authors: Zheng Ma, Junyu Xuan, Yu Guang Wang, Ming Li, Pietro Liò

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section, we present the test results of PAN on various datasets in graph classification tasks. We show a performance comparison of PAN with some existing GNN methods. All the experiments were performed using Py Torch Geometric [22] and run on a server with Intel(R) Core(TM) i9-9820X CPU 3.30GHz, NVIDIA Ge Force RTX 2080 Ti and NVIDIA TITAN V GV100. The codes can be downloaded at https://github.com/Yu Guang Wang/PAN.
Researcher Affiliation Academia Zheng Ma Department of Physics Princeton University mazhengparnassum@gmail.com Junyu Xuan Centre for Artificial Intelligence Faculty of Engineering and Information Technology University of Technology Sydney junyu.xuan@uts.edu.au Yu Guang Wang Max Planck Institute for Mathematics in the Sciences & School of Mathematics and Statistics University of New South Wales yuguang.wang@mis.mpg.de Ming Li Department of Educational Technology Zhejiang Normal University mingli@zjnu.edu.cn Pietro Liò Department of Computer Science and Technology University of Cambridge Pietro.Lio@cl.cam.ac.uk
Pseudocode No The paper describes the proposed methods in narrative text but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes The codes can be downloaded at https://github.com/Yu Guang Wang/PAN.
Open Datasets Yes We test the performance of PAN on five widely used benchmark datasets for graph classification tasks [33], including two protein graph datasets PROTEINS and PROTEINS_full [10, 19]; one mutagen dataset MUTAGEN [50, 32] (full name Mutagenicity); and one dataset that consists of chemical compounds screened for activity against non-small cell lung cancer and ovarian cancer cell lines NCI1 [56]; one dataset that consists of molecular compounds for activity against HIV or not AIDS [50]. We name the dataset Point Pattern, which can be downloaded from the links contained in the supplementary material.
Dataset Splits Yes In each experiment, we split 80%, 10%, and 10% of each dataset for training, validation, and testing. We split the data into training, validation, and test sets of size 12,000, 1,500, and 1,500.
Hardware Specification Yes All the experiments were performed using Py Torch Geometric [22] and run on a server with Intel(R) Core(TM) i9-9820X CPU 3.30GHz, NVIDIA Ge Force RTX 2080 Ti and NVIDIA TITAN V GV100.
Software Dependencies No All the experiments were performed using Py Torch Geometric [22]. The paper mentions a software library but does not provide specific version numbers for it or other ancillary software.
Experiment Setup Yes We use batch size of 128 and 200 epochs. The learning rate and weightdecay are set to 0.01 and 5e-4, respectively. We fix the number of neurons in the convolutional layers to 64, the learning rate and weight decay are set to 0.001 and 0.0005.