Research map / Label-efficient and robust learning / Recognition and long-tail
Temporally and Distributionally Robust Optimization for Cold-Start Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Temporal DRO (TDRO), a group-DRO variant for cold-start recommendation that adds a worst-case factor and a gradient-based shifting factor capturing temporal item feature shifts, and plugs into CLCRec and GAR.
Problem
Cold-start item recommendation degrades because item features of cold items temporally shift away from those of warm items, so a feature extractor trained with ERM or standard DRO generalizes poorly to cold items.
Approach
Extends group DRO with two group-selection factors: a worst-case factor (larger-loss groups prioritized) and a shifting factor that scores each group's gradient against a period-weighted sum of per-period group gradients (the shifting trend), emphasizing groups likely to become popular in later periods; made efficient via first-order Taylor approximation and smoothed with KL-regularized group weig
Key results
- On Amazon, CLCRec+TDRO improves all-item Recall@20 from 0.0106 to 0.0130 and cold Recall@20 from 0.0135 to 0.0166.
- On Micro-video, GAR+TDRO improves cold Recall@20 from 0.0840 to 0.0947 and all-item Recall@20 from 0.0644 to 0.0711.
- TDRO outperforms S-DRO on the backbones across datasets, with improvements marked statistically significant (p<0.01, one-sample t-tests) in most settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TDRO | CLCRec · Amazon cold | Recall@20: 0.0135 -> 0.0166 | Helped |
| add TDRO | CLCRec · Micro-video cold | Recall@20: 0.0623 -> 0.0761 | Helped |
| add TDRO | CLCRec · Kwai cold | Recall@20: 0.0160 -> 0.0186 | Helped |
| add TDRO | CLCRec · Amazon all | Recall@20: 0.0106 -> 0.0130 | Helped |
| add TDRO | CLCRec · Micro-video all | Recall@20: 0.0583 -> 0.0703 | Helped |
| add TDRO | CLCRec · Kwai all | Recall@20: 0.0743 -> 0.0841 | Helped |
| add worst-case factor (ablation) | CLCRec + TDRO · Amazon all | Recall@20: 0.0121 -> 0.0130 | Helped |
| add shifting factor (ablation) | CLCRec + TDRO · Amazon all | Recall@20: 0.0126 -> 0.0130 | Helped |
| add worst-case factor (ablation) | CLCRec + TDRO · Micro-video all | Recall@20: 0.0648 -> 0.0703 | Helped |
| add shifting factor (ablation) | CLCRec + TDRO · Micro-video all | Recall@20: 0.0643 -> 0.0703 | Helped |
| add worst-case factor (ablation) | CLCRec + TDRO · Kwai cold | Recall@20: 0.0145 -> 0.0186 | Helped |
| increase the shifting-factor strength (ablation) | TDRO · cold items | Recall@20 | Helped |
Surprising findings
- Standard S-DRO gives only inconsistent gains for cold items, and the authors attribute this to an inconsistency between overemphasized minority warm groups and the shifting cold-item distribution.
- InvRL improves cold performance over its CLCRec backbone but sacrifices warm-item accuracy.
Limitations
- Group division relies on pre-defined K-means clustering of item features, which the authors note may be suboptimal.
- Kwai uses random interaction splitting, so temporal shifts are less pronounced there and the worst-case factor ablation behaves differently.
- The number of time periods and shifting strength are sensitive hyperparameters; overly large values hurt performance.
Experimental setup
Datasets: Amazon Micro-video Kwai
Benchmarks: Amazon Micro-video Kwai
Metrics: Recall@20 NDCG@20
Baselines: DUIF DropoutNet M2TRec MTPR Heater CB2CF CCFCRec CLCRec GAR InvRL S-DRO
Builds on: Group DRO CLCRec GAR S-DRO DropoutNet
Closest papers by meaning
- Preference Aware Dual Contrastive Learning for Item Cold-Start Recommendation (AAAI 2024 · similarity 0.89)
- Counterfactual Task-augmented Meta-learning for Cold-start Sequential Recommendation (AAAI 2025 · similarity 0.87)
- Feature-Structure Adaptive Completion Graph Neural Network for Cold-start Recommendation (AAAI 2025 · similarity 0.86)
- Addressing Cold-Start Problem in Click-Through Rate Prediction via Supervised Diffusion Modeling (AAAI 2025 · similarity 0.85)
- On Softmax Direct Preference Optimization for Recommendation (NeurIPS 2024 · similarity 0.85)
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.85)
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.