AI Paper Atlas

Research map / Graph learning / GNN architectures

The Infinite Contextual Graph Markov Model

Daniele Castellana, Federico Errica, Davide Bacciu, Alessio Micheli

ICML 2022 · Graph learning › GNN architectures · approach: Theory of graph convolutional networks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Replace CGMM's grid-searched layer size with HDP nonparametric states (learned alpha0, gamma)ICGMM · NCI1accuracy: 77.6 vs 76.2 for CGMMNo clear effect
Use Bayesian nonparametric HDP layers instead of CGMMICGMM · D&Daccuracy: +1.4 over CGMM (76.3 vs 74.9)Improved
Use ICGMM in place of CGMM on social datasetsICGMM · IMDB-B, REDDIT-B, COLLABaccuracy: ICGMM better on average than other methods on IMDB-B, REDDIT-B, COLLABImproved
Replace exact Gibbs sampling with batch-of-one-graph approximate sampling (ICGMMf) (ablation)ICGMM · chemical datasets (NCI1, D&D, PROTEINS)accuracy: NCI1 76.4 vs 77.6; D&D 75.1 vs 76.3; PROTEINS 73.2 vs 73.1No clear effect
Fix alpha0 and gamma (ICGMMαγ) instead of learning them with hyper-priors (ablation)ICGMM · REDDIT-Baccuracy: 91.3 vs 91.6 for ICGMMfHurt
Increase depth from 1 to 20 layers (ablation)ICGMM · NCI1 validation accuracyaccuracy: validation accuracy rises with depth then plateaus near 20 layersImproved
Learn number of states adaptively (HDP) instead of fixed C=20 in CGMMICGMM · NCI1 graph embedding sizeembedding size: embedding size approximately 7% of CGMM'sImproved
Learn number of states adaptively (HDP) instead of fixed C=20 in CGMMICGMM · NCI1 unused statesunused states: near-zero unused states vs CGMM's linearly growing unused statesImproved
Use Bayesian nonparametric treatment and fixed hyper-priors, reducing configurations triedICGMM · social and chemical tasksnumber of configurations: 6x reduction in configurations triedImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy