AI Paper Atlas

Research map / Graph learning / Graph self-supervised learning

Self-Supervised Representation Learning via Latent Graph Prediction

Yaochen Xie, Zhao Xu, Shuiwang Ji

ICML 2022 · Graph learning › Graph self-supervised learning · approach: Heterophily and imbalance on graphs · 8 citations (OpenAlex)

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
Use latent-graph prediction with reconstruction and representation invariance lossesLaGraph · Graph-level classification datasetsaccuracy: NCI1: 77.9±0.4 -> 79.9±0.5; PROTEINS: 74.4±0.3 -> 75.2±0.4; MUTAG: 89.7±1.1 -> 90.2±1.1; COLLAB: 71.4±1.2 -> 77.6±0.2; RImproved
Use latent-graph prediction with reconstruction and representation invariance lossesLaGraph · DD and IMDB-Baccuracy: DD: 78.6±0.4 -> 78.1±0.4; IMDB-B: 74.2±0.7 -> 73.7±0.9Mixed
Use LaGraph rather than the strongest listed SSL baselinenode-level transductive classification · Amazon Photos, Coauthor CS, Coauthor Physicsaccuracy: Amazon Photos: 92.9±0.3 -> 93.5±0.4; Coauthor CS: 93.2±0.2 -> 93.3±0.2; Coauthor Physics: 95.6±0.1 -> 95.8±0.1Improved
Use LaGraph rather than BGRLAmazon Computers · Amazon Computersaccuracy: 89.7±0.3 -> 88.0±0.3Hurt
Use LaGraph rather than the strongest listed SSL baselinenode-level inductive classification · PPI and Flickrmicro-F1: PPI: 70.5±0.1 -> 74.6±0.0; Flickr: 50.0±0.3* -> 51.3±0.1Improved
Use LaGraph rather than SUBG-CONReddit · Redditmicro-F1: 95.2±0.0 -> 95.2±0.0No clear effect
Reduce sampled training nodes while training LaGraph (ablation)LaGraph · Flickrmicro-F1: 100 sampled: 6.07; 1,000: 51.12; 2,500: 51.12; 5,000: 51.27; 10,000: 51.29; all: 51.26Mixed
Reduce sampled training nodes while training LaGraph (ablation)LaGraph · Redditmicro-F1: 100 sampled: 5.76; 1,000: 95.05; 2,500: 95.06; 5,000: 95.08; 10,000: 95.09; all: 95.22Mixed
Replace embedding-level invariance with output-level invariance (ablation)LaGraph · Graph-level classification datasetsaccuracy: MUTAG: 90.2±1.3 -> 89.2±2.1; RDT-B: 90.4±0.9 -> 89.8±0.9; RDT-M5K: 56.4±0.2 -> 56.0±0.4; COLLAB: 77.6±0.1 -> 77.7±0.1Mixed
Replace MSE reconstruction/invariance with cross-entropy/KL versions (ablation)LaGraph · Graph-level classification datasetsaccuracy: NCI1: 79.9±0.5 -> 79.9±0.5; PROTEINS: 75.2±0.3 -> 75.2±0.3; DD: 78.1±0.3 -> 78.1±0.4No clear effect
Remove concatenation of original node features with encoder output (ablation)LaGraph · Amazon Computers and PPIaccuracy / micro-F1: Amazon Computers: 88.0±0.3 -> 88.8±0.3; PPI: 74.6±0.0 -> 75.2±0.0Mixed
Remove concatenation of original node features with encoder output (ablation)LaGraph · Amazon Photos, Coauthor CS, Coauthor Physics, Redditaccuracy / micro-F1: Amazon Photos: 92.7±0.4 -> 93.5±0.4; Coauthor CS: 92.6±0.2 -> 93.3±0.2; Coauthor Physics: 95.3±0.1 -> 95.8±0.1; Reddit: Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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