Beyond Low-frequency Information in Graph Convolutional Networks
Authors: Deyu Bo, Xiao Wang, Chuan Shi, Huawei Shen3950-3957
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on six real-world networks validate that FAGCN not only alleviates the over-smoothing problem, but also has advantages over the state-of-the-arts. We compare FAGCN with two types of representative GNNs: Spectral-based methods, i.e., SGC (Wu et al. 2019), GCN (Kipf and Welling 2017), Cheb Net (Defferrard, Bresson, and Vandergheynst 2016) and GWNN (Xu et al. 2019b); Spatialbased methods, i.e., GIN (Xu et al. 2019c), GAT (Velickovic et al. 2018), Mo Net (Monti et al. 2017), Graph SAGE (Hamilton, Ying, and Leskovec 2017) and APPNP (Klicpera, Bojchevski, and G unnemann 2019). |
| Researcher Affiliation | Academia | Deyu Bo1, Xiao Wang1, Chuan Shi1 , Huawei Shen2 1Beijing University of Posts and Telecommunications 2CAS Key Lab of Network Data Science and Technology, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China {bodeyu, xiaowang, shichuan}@bupt.edu.cn, shenhuawei@ict.ac.cn |
| Pseudocode | No | The paper describes mathematical formulas for the model (Eq. 1-7) but does not provide structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement about the release of its source code or a link to a code repository. |
| Open Datasets | Yes | Assortative datasets. We choose the commonly used citation networks, e.g., Cora, Citeseer and Pubmed for assortative datasets... In each network, we use 20 labeled nodes per class for training, 500 nodes for validation and 1000 nodes for testing. Details can be found in (Kipf and Welling 2017). Disassortative datasets. We consider the Wikipedia networks1 and Actor co-occurrence network (Tang et al. 2009) for disassortative datasets. Chameleon and Squirrel are two Wikipedia networks... we use 20% for validation, 20% for testing and change the training ratio from 10% to 60%. More detailed characteristics of the datasets can be found in Table 1. Note that a higher value of the second column represents a more obvious assortativity (Newman 2003). |
| Dataset Splits | Yes | In each network, we use 20 labeled nodes per class for training, 500 nodes for validation and 1000 nodes for testing. In disassortative networks...we use 20% for validation, 20% for testing and change the training ratio from 10% to 60%. |
| Hardware Specification | No | The paper states that 'All methods were implemented in Pytorch with Adam optimizer (Kingma and Ba 2015)' but does not specify any hardware details such as CPU, GPU, or memory. |
| Software Dependencies | No | All methods were implemented in Pytorch with Adam optimizer (Kingma and Ba 2015). The paper mentions 'Pytorch' and 'Adam optimizer' but does not specify their version numbers. |
| Experiment Setup | Yes | The hidden unit is fixed at 16 in assortative networks and 32 in disassortative networks. The hyper-parameter search space is: learning rate in {0.01, 0.005}, dropout in {0.4, 0.5, 0.6}, weight decay in {1E-3, 5E-4, 5E-5}, number of layers in {1, 2, , 8}, ε in {0.1, , 1.0}. For FAGCN, the hyper-parameter setting is: learning rate = 0.01, dropout = 0.6, weight decay = 1E-3, layers = 4. ε = 0.2, 0.3, 0.3 for Cora, Citeseer and Pubmed. In disassortative datasets, the hyper-parameter for FAGCN is: learning rate = 0.01, dropout = 0.5, weight decay = 5E-5, layers = 2. ε = 0.4, 0.3, 0.5 for Chameleon, Squirrel and Actor, respectively. Besides, we run 500 epochs and choose the model with highest validation accuracy for testing. |