Research map / Architectures and efficiency / Attention and transformers
Interest-Shift-Aware Logical Reasoning for Efficient Long-Sequence Recommendation
Fei Li, Qingyun Gao, Enneng Yang, Jianzhe Zhao, Guibing Guo
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 |
|---|---|---|---|
| add ILR contrastive loss (multi-interest reasoning) to SIE + LIC-S (ablation) | ELECTOR · MoTV (seq len 100) | N@10: 0.400 -> 0.816 | Improved |
| add ILR contrastive loss to recent-sequence reasoning (ablation) | ELECTOR variants · Amazon(500) | N@10: 0.235 -> 0.454 | Improved |
| add subsequence interest extraction (SIE) to NCR (ablation) | NCR · ML-1M (seq len 100) | N@10: 0.207 -> 0.812 | Improved |
| add LIC-S subsequence interest correlations without ILR loss (ablation) | ELECTOR base · ML-1M (seq len 100) | N@10: 0.812 -> 0.693 | Hurt |
| add LIC-S subsequence interest correlations without ILR loss (ablation) | ELECTOR base · Amazon(1000) | N@10: 0.408 -> 0.585 | Improved |
| add recent-item interest correlation (LIC-RI) to Variant 2 (ablation) | ELECTOR · Amazon(500) | N@10: 0.753 -> 0.854 | Improved |
| add LIC-RI recent-interest reasoning alone (no SIE, no ILR) vs NCR (ablation) | NCR · ML-1M (seq len 100) | N@10: 0.207 -> 0.204 | No clear effect |
| reason over limited interest embeddings rather than all items | NCR · Amazon(1000) | training time per epoch (s): 818s -> 241s (LRM), 153s (KFM) | Improved |
| use kernel feature maps (KFM) instead of low-rank matrix approximation (LRM) in SIE | ELECTOR · Amazon(1000) | training time per epoch (s): 241s -> 153s | Improved |
| ELECTOR over strongest baseline on long sequences | ELECTOR · Amazon(1000) | N@5: 288.02% over strongest baseline | Improved |
| increase sequence length from 50 to 100 for NLR/NCR | NCR · ML-1M | N@10: 0.316 -> 0.207 | Hurt |
| increase subsequence length t (ablation) | ELECTOR · sensitivity analysis | accuracy / training time | Improved |
Closest papers by meaning
- UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · similarity 0.87)
- Reinforced Latent Reasoning for LLM-based Recommendation (ICLR 2026 · similarity 0.86)
- LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs (AAAI 2024 · similarity 0.86)
- Iterative Sparse Attention for Long-sequence Recommendation (AAAI 2025 · similarity 0.86)
- R$^2$ec: Towards Large Recommender Models with Reasoning (NeurIPS 2025 · similarity 0.86)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.86)
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.
