Zero-shot Node Classification with Graph Contrastive Embedding Network

Authors: Wei Ju, Yifang Qin, Siyu Yi, Zhengyang Mao, Kangjie Zheng, Luchen Liu, Xiao Luo, Ming Zhang

TMLR 2023 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments indicate that our Graph CEN significantly outperforms the state-of-the-art approaches on multiple challenging benchmark datasets. ... 4 Experiments
Researcher Affiliation Academia 1School of Computer Science, National Key Laboratory for Multimedia Information Processing, Peking University 2School of Statistics and Data Science, Nankai University 3Department of Computer Science, University of California, Los Angeles
Pseudocode Yes Algorithm 1: Optimization Algorithm of Graph CEN Input : Graph data G = (V, E, X, A), classes semantic descriptions matrix S, total classe set C, seen classes Cs, trainable weight matrix W in GNN, parameter matrix Θ = (Θ 1 , Θ 2 ) in two projection heads, temperature parameter τ and contrastive parameter β Output: Class assignments for unlabeled nodes from unseen class Cu Construct the class affinity graph GA by Eq. 2; Initializing trainable weight matrix W and parameter matrix Θ; while not done do Compute the joint information matrix G by Eq. 4; Sample a mini-batch of nodes from G; Sample one augmentation from Section 2; Compute node-level contrastive loss LNCL by Eq. 6; Compute class-level contrastive loss LCCL by Eq. 8; Compute the supervised loss LCE by Eq. 10; Update parameters W and Θ by gradient descent to minimize L by Eq. 11 end
Open Source Code No The paper does not provide an explicit statement or link for open-source code for the described methodology.
Open Datasets Yes For comprehensive comparisons, we conduct extensive experiments on three real-world citation datasets, which are Cora (Mc Callum et al., 2000), Citeseer (Giles et al., 1998), C-M10M (Pan et al., 2016) and a large-scale dataset ogbn-arxiv (Wang et al., 2020).
Dataset Splits Yes The specific class labels of the datasets are displayed in Table 1. For the zero-shot node classification, we follow the same seen/unseen class split settings introduced in (Wang et al., 2021b) to make a fair comparison as shown in Table 1. ... Table 1: The data split and class information of three citation datasets. Dataset Classes Class Split I Class Split II Class Labels [Train/Val/Test] [Train/Val/Test] Cora 7 [3/0/4] [2/2/3] Neural Network, Rule Learning, Reinforcement Learning, Probabilistic Methods, Theory, Genetic Algorithms, Cased based
Hardware Specification No The paper does not provide specific details about the hardware used for running its experiments.
Software Dependencies No Besides, the framework is trained using Adam optimizer (Kingma & Ba, 2014). This mentions an optimizer but lacks a specific version number for it or other key software components/libraries.
Experiment Setup Yes For our Graph CEN, we adopt the grid search for the parameters under the class split I, and determine the optimal hyper-parameters by the validation classes under the class split II. The space of hyper-parameters is carefully selected as follows: the learning rate {0.0005, 0.001, 0.005, 0.01}, the number of hidden units {32, 64, 128, 256}, the temperature parameter τ {0.05, 0.1, 0.2, 0.5, 1.0, 2.0, 5.0, 10.0}, the contrastive parameter β {0.05, 0.1, 0.2, 0.5, 1.0, 2.0, 5.0, 10.0}. Besides, the framework is trained using Adam optimizer (Kingma & Ba, 2014).