Unsupervised Neighborhood Propagation Kernel Layers for Semi-supervised Node Classification

Authors: Sonny Achten, Francesco Tonin, Panagiotis Patrinos, Johan A.K. Suykens

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate the effectiveness of the proposed framework. Experiments Datasets and Main Setting As datasets, we use four homophilious graphs: Cora, Cite Seer, Pub Med (Sen et al. 2008; Yang, Cohen, and Salakhudinov 2016), and OGBArxiv (Hu et al. 2020), which are citation graphs, as well as two heterophilious graphs: Chameleon and Squirrel (Wikipedia graphs (Rozemberczki, Allen, and Sarkar 2021)).
Researcher Affiliation Academia Sonny Achten, Francesco Tonin, Panagiotis Patrinos, Johan A. K. Suykens KU Leuven, Department of Electrical Engineering (ESAT), STADIUS Center for Dynamical Systems, Signal Processing and Data Analytics {sonny.achten, francesco.tonin, panos.patrinos, johan.suykens}@esat.kuleuven.be
Pseudocode Yes Algorithm 1 Optimization algorithm of GCKM. 1: Initialize {H(1) 0 , H(2) 0 , H(3) 0 } 2: for k 0, 1, . . . , T do 3: Compute K(1) c from aggregated X 4: Compute K(2) c from aggregated H(1) k 5: Update {H(1) k+1, H(2) k+1} Cayley Adam(JGCKM) 6: Compute K(3) from H(2) k+1 7: Update {H(3) k+1} Solve (11) with K(3)
Open Source Code Yes The reported results can be reproduced using our code on Git Hub1 and the Appendix is available at Ar Xiv2. 1https://github.com/sonnyachten/GCKM
Open Datasets Yes As datasets, we use four homophilious graphs: Cora, Cite Seer, Pub Med (Sen et al. 2008; Yang, Cohen, and Salakhudinov 2016), and OGBArxiv (Hu et al. 2020), which are citation graphs, as well as two heterophilious graphs: Chameleon and Squirrel (Wikipedia graphs (Rozemberczki, Allen, and Sarkar 2021)).
Dataset Splits Yes For Cora, Cite Seer, and Pub Med, there are 4 training labels per class, 100 labels for validation, and 1000 labels for testing. For Chameleon and Squirrel, a 0.5%/0.5%/99% train/validation/test-split is used. For OGB-Arxiv, we use a 2.5%/2.5%/95% random split
Hardware Specification No We also thank the Flemish Supercomputer (VSC).
Software Dependencies No We therefore employ the Cayley Adam optimizer (Li, Li, and Todorovic 2019) to update H(1), H(2) with H(3) fixed.
Experiment Setup Yes We used GCN aggregation for the citation networks and sum aggregation for the heterophilious graphs. We used a deep GCKM with only two unsupervised layers, with RBF bandwidth σ2 RBF = mσ2 with m the input dimension and σ2 the variance of the inputs, and clustering obtained by k-means on H(2).