I’m Me, We’re Us, and I’m Us: Tri-directional Contrastive Learning on Hypergraphs

Authors: Dongjin Lee, Kijung Shin

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our extensive experiments using 14 baseline approaches, 10 datasets, and two tasks demonstrate the effectiveness of Tri CL, and most noticeably, Tri CL almost consistently outperforms not just unsupervised competitors but also (semi-)supervised competitors mostly by significant margins for node classification. The code and datasets are available at https://github.com/wooner49/Tri CL.
Researcher Affiliation Academia Dongjin Lee1 and Kijung Shin1,2 1School of Electrical Engineering, KAIST, South Korea 2Kim Jaechul Graduate School of AI, KAIST, South Korea {dongjin.lee, kijungs}@kaist.ac.kr
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes The code and datasets are available at https://github.com/wooner49/Tri CL.
Open Datasets Yes The code and datasets are available at https://github.com/wooner49/Tri CL. We assess the performance of Tri CL on 10 commonly used benchmark datasets; these datasets are categorized into (1) co-citation datasets (Cora, Citeseer, and Pubmed) (Sen et al. 2008), (2) co-authorship datasets (Cora and DBLP (Rossi and Ahmed 2015)), (3) computer vision and graphics datasets (NTU2012 (Chen et al. 2003) and Model Net40 (Wu et al. 2015)), and (4) datasets from the UCI Categorical Machine Learning Repository1 (Zoo, 20Newsgroups, and Mushroom).
Dataset Splits Yes For all the datasets, we randomly split them, where 10%, 10%, and 80% of nodes are chosen for the training, validation, and test set, respectively, and evaluate the model with 20 dataset splits.
Hardware Specification Yes All experiments were conducted on a single 24GB NVIDIA GeForce RTX 3090 GPU.
Software Dependencies No The paper states that the method is 'implemented in Python with PyTorch and PyTorch Geometric' but does not provide specific version numbers for these software components.
Experiment Setup Yes The encoder is firstly trained in a fully unsupervised manner and computes node representations; then, a simple linear classifier is trained on top of these frozen representations through a ℓ2-regularized logistic regression loss, without flowing any gradients back to the encoder. For all the datasets, we randomly split them, where 10%, 10%, and 80% of nodes are chosen for the training, validation, and test set, respectively, and evaluate the model with 20 dataset splits. We employ a one-layer mean pooling hypergraph encoder described in (3) and PRe LU (He et al. 2015) activation for non-linearlity. A further description of the model, optimizer, and model hyperparameters are provided in Appendix C.