AI Paper Atlas

Research map / Graph learning / Graph self-supervised learning

Diverse and Sparse Mixture-of-Experts for Causal Subgraph–Based Out-of-Distribution Graph Learning

Jerry Sun, Mohamed Abubakr Hassan, Yaoyu Zhang, Wanying Zhang, Chi-Guhn Lee

ICLR 2026 · Graph learning › Graph self-supervised learning · approach: Graph out-of-distribution generalization

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
add diversity loss among experts (ablation)DiSCO · GOOD-Motif Sizeaccuracy: 70.07 -> 75.52Improved
add diversity loss among experts (ablation)DiSCO · GOOD-HIV ScaffoldROC-AUC: 65.95 -> 71.55Improved
add gate loss (teacher-student, sparsity, balance) (ablation)DiSCO · GOOD-HIV SizeROC-AUC: 61.91 -> 66.98Improved
add keep-rate regularization loss (ablation)DiSCO · GOOD-Motif Basisaccuracy: 67.48 -> 92.80Improved
add keep-rate regularization loss (ablation)DiSCO · GOOD-SST2 Lengthaccuracy: 83.46 -> 83.73No clear effect
add diversity loss (ablation)DiSCO · GOOD-Twitter Lengthaccuracy: 60.10 -> 61.13No clear effect
add semantic diversity objective (ablation)DiSCO · Twitter, SST2, Motif-Basisper-batch loss gap: Twitter 0.13 -> 0.19; SST2 0.07 -> 0.22; Motif-Basis 0.076 -> 0.12Improved
increase experts from 1 to 4 (ablation)DiSCO · GOOD benchmark (6 splits, avg)average accuracy/ROC-AUC: 70.84 -> 74.68Improved
increase experts from 4 to 8 (ablation)DiSCO · GOOD benchmark (6 splits, avg)average accuracy/ROC-AUC: 74.68 -> 75.29Mixed
perturb keep-rate prior by +/-0.1 or 0.2 (ablation)DiSCO · Twitter, SST2, Motif-Basisaccuracy: Motif-Basis 92.80 -> 90.13 at -0.2No clear effect
use single-expert causal subgraph extractor with only sparsity priorDiSCO (1 expert) · GOOD benchmark (6 splits, avg)average accuracy/ROC-AUC: 64.24 (ERM) -> 70.84Improved
use MoE of diverse causal-subgraph experts instead of single-invariant-subgraph methodsDiSCO vs LECI · CFP-Motif (covariate, FIIF, PIIF)accuracy: 83.20 -> 90.83; 77.73 -> 84.17; 69.40 -> 77.19Improved

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