AI Paper Atlas

Research map / Label-efficient and robust learning / Knowledge distillation

Rejuvenating Cross-Entropy Loss in Knowledge Distillation for Recommender Systems

Zhangchi Zhu, Wei Zhang

ICLR 2026 · Label-efficient and robust learning › Knowledge distillation · approach: Rethinking distillation objectives · 2 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
Compute CE only on teacher's top items (vanilla CE) vs RCE-KD's split-and-sample schemeCE-based KD for recommenders · CiteULike/Gowalla/Yelp2018Recall/NDCG: RCE-KD best in all 12 metrics per setting; e.g. Gowalla MF->MF R@20 0.1427 -> 0.1525 vs best baselineImproved
Add separate losses for the two teacher-top-item subsets (ablation)RCE-KD · Gowalla, Yelp2018Recall/NDCGImproved
Use adaptive gamma weight instead of constant (ablation)RCE-KD · Gowalla, Yelp2018Recall/NDCGImproved
Include student's top items (not only teacher's top items) in the CE item set (ablation)RCE-KD · Gowalla, Yelp2018Recall/NDCGImproved
RCE-KD sampled subset vs full-item CE (ablation)RCE-KD · Gowalla, Yelp2018Recall/NDCGNo clear effect
Use RCE-KD instead of other KD methods, training costStudent MF on Gowalla · Gowallatraining time per epoch (s): RCE-KD 99.30 vs RRD 132.37, HetComp 137.62, Student 27.33Mixed

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