Research map / Label-efficient and robust learning / Recognition and long-tail
Everyone's Preference Changes Differently: A Weighted Multi-Interest Model For Retrieval
Hui Shi, Yupeng Gu, Yitong Zhou, Bo Zhao, Sicun Gao, Jishen Zhao
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 changed | Where | Measured | Effect |
|---|---|---|---|
| two-stage training (freeze cluster weights at 1, then unfreeze) instead of joint training (ablation) | MIP · Amazon | AUC: 0.576 -> 0.806 | Improved |
| two-stage training instead of joint training (ablation) | MIP · Taobao | AUC: 0.803 -> 0.885 | Improved |
| two-stage training instead of joint training (ablation) | MIP · MovieLens | AUC: 0.924 -> 0.930 | Mixed |
| learned personalized cluster weights instead of equal weights or exponential decay (ablation) | MIP · Amazon, Taobao, MovieLens, Pinterest | AUC / recall | Improved |
| NLL loss instead of triplet loss (alpha 0.2, 0.5, 0.8) (ablation) | MIP · MovieLens | AUC: 0.885 -> 0.930 | Improved |
| add positional and temporal encoding to item embedding (ablation) | MIP · Amazon | AUC: 76.34 -> 79.31 | Improved |
| add positional or temporal encoding (ablation) | MIP · Taobao | AUC: 82.06 -> 86.59 | Improved |
| add positional and temporal encoding (ablation) | MIP · MovieLens | AUC: 95.26 -> 94.61 | Hurt |
| add sinusoid temporal and positional encoding (ablation) | MIP · Pinterest | AUC: 0.8923 -> 0.8926 | No clear effect |
| replace sinusoid temporal encoding with one-hot or two-hot (ablation) | MIP · Pinterest | AUC: 0.8921 -> 0.8850 | Hurt |
| multi-interest MIP versus single user embedding TiSASRec | MIP · Pinterest | AUC: 0.850 -> 0.893 | Improved |
| self-attention multi-interest (MIP) instead of global-query attention (ComiRec) | MIP · Pinterest | AUC: 0.875 -> 0.893 | Improved |
Closest papers by meaning
- Density-based User Representation using Gaussian Process Regression for Multi-interest Personalized Retrieval (NeurIPS 2024 · similarity 0.86)
- On the Embedding Collapse when Scaling up Recommendation Models (ICML 2024 · similarity 0.86)
- Auto Encoding Neural Process for Multi-interest Recommendation (AAAI 2025 · similarity 0.86)
- UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · similarity 0.85)
- Long-Sequence Recommendation Models Need Decoupled Embeddings (ICLR 2025 · similarity 0.85)
- Learning Multiple User Distributions for Recommendation via Guided Conditional Diffusion (AAAI 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.
