AI Paper Atlas

Research map / Graph learning / Graph self-supervised learning

Negative Sampling From the Ground Up: A Redesign for Recommendation

Yanbang Wang, Jon Kleinberg, Yanhong Wu

ICML 2026 · Graph learning › Graph self-supervised learning · approach: Heterophily and imbalance on graphs

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 adaptive reweighting by model probability pf(v|u) (adaptive hard negatives) (ablation)Our negative sampling (Node Embedding) · MovieLensRecall@20 (%): 7.60 -> 12.77Improved
add adaptive reweighting by model probability pf(v|u) (ablation)Our negative sampling (Node Embedding) · PinterestRecall@20 (%): 6.54 -> 12.59Improved
learn propensity score with graph features instead of a constant (ablation)Our negative sampling (Node Embedding) · PinterestRecall@20 (%): 9.38 -> 12.59Improved
add regularization (first moment, consistency) to propensity estimator (ablation)Our negative sampling (Node Embedding) · PinterestRecall@20 (%): 8.99 -> 12.59Improved
tune positive class prior instead of setting it to 0.5 (ablation)Our negative sampling (Node Embedding) · PinterestRecall@20 (%): 9.39 -> 12.59Improved
model observation bias instead of using unbiased-case sampling (ablation)Our negative sampling (Node Embedding) · PinterestRecall@20 (%): 8.01 -> 12.59Improved
model observation bias instead of using unbiased-case sampling (ablation)Our negative sampling (Node Embedding) · MovieLensRecall@20 (%): 10.35 -> 12.77Improved
use proposed sampling instead of AHNS (strongest baseline), LightGCNLightGCN · PinterestRecall@20 (%): 11.35 -> 17.61Improved
use proposed sampling on unbiased LastFM splitNode Embedding · LastFMRecall@20 (%): 7.37 -> 7.38No clear effect
use proposed sampling instead of full contrastive lossGRACE · Cora, CiteSeeraccuracy: Cora 83.3 -> 83.6; CiteSeer 72.1 -> 71.6Mixed
use proposed sampling instead of full contrastive lossGCE · Wiki-CS, Amazon-Photoaccuracy: 78.30 -> 77.41; 92.49 -> 91.94Hurt
use proposed sampling instead of RNSLightGCN · MovieLensRecall@20 (%): 8.94 -> 17.41Improved

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