Research map / Large language models / Agents, code and math
Fuzzy Collaborative Reasoning
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 FuzzCR, a non-parametric fuzzy-logic collaborative reasoning model that casts sequential recommendation as first-order logic query answering with learning-free fuzzy conjunction, disjunction and negation operators.
Problem
Existing collaborative reasoning recommenders simulate logical operators with parameterized networks that violate classical logic axioms and generalize poorly beyond training patterns.
Approach
Replace learned logical operators with parameter-free product-logic fuzzy operators on fuzzy embeddings in [0,1]^d, cast next-item prediction as an FOL query with negation and disjunction, so that the logic satisfies classical rules and the learned embeddings can be reused for other query structures.
Key results
- FuzzCR achieves best N@10 on ML100K (0.4719 vs 0.4624 for CBox4CR) and HR@10 on Sports (0.5069 vs 0.4988), with reported improvements of 0.43% to 3.65% across metrics.
- Pre-trained FuzzCRp on 1-projection queries is competitive with baselines, e.g., Sports N@10 0.2859 versus 0.2724 for NLR.
- Product logic outperforms Gödel and Łukasiewicz logic in the ablation study.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add disjunction with the user embedding (ablation) | FuzzCR · ML100K, Beauty, Sports | N@10 | Helped |
| Add negation for negative interactions (ablation) | FuzzCR · ML100K, Beauty, Sports | N@10 | Helped |
| Add positional encoding (ablation) | FuzzCR · ML100K, Beauty, Sports | N@10 | Helped |
| Use product logic instead of Gödel or Łukasiewicz logic (ablation) | FuzzCR · ML100K, Beauty, Sports | N@10 | Helped |
| Use FuzzCR instead of CBox4CR | FuzzCR · ML100K | N@5: 3.65% | Helped |
| Use FuzzCR instead of CBox4CR | FuzzCR · ML100K | N@10: 2.05% | Helped |
| Use FuzzCR instead of CBox4CR | FuzzCR · ML100K | HR@5: 3.03% | Helped |
| Use FuzzCR instead of CBox4CR | FuzzCR · ML100K | HR@10: 0.43% | Helped |
| Use FuzzCR instead of CBox4CR | FuzzCR · Beauty | N@5: 2.39% | Helped |
| Use FuzzCR instead of CBox4CR | FuzzCR · Beauty | N@10: 1.86% | Helped |
| Use FuzzCR instead of CBox4CR | FuzzCR · Beauty | HR@5: 2.75% | Helped |
| Use FuzzCR instead of CBox4CR | FuzzCR · Beauty | HR@10: 0.95% | Helped |
Surprising findings
- The pre-trained FuzzCRp does not beat all baselines on every metric; on Beauty it trails NLR and GRU4Rec on N@5 and HR@5.
Limitations
- Only uses user-item interaction sequences; side information is not incorporated.
- Generalization is shown only via pre-training on 1-projection queries and not on broader query structures.
- Results depend on tuned gamma coefficient and sequence length.
Experimental setup
Datasets: MovieLens-100K Amazon Beauty Amazon Sports
Benchmarks: MovieLens-100K Amazon Beauty Amazon Sports
Metrics: NDCG@5 NDCG@10 HR@5 HR@10
Baselines: BPRMF SVD++ NeuMF GRU4Rec NARM STAMP NLR NCR Cbox4CR
Builds on: FuzzQE NCR Cbox4CR BetaE Query2Box
Closest papers by meaning
- Rethinking Complex Queries on Knowledge Graphs with Neural Link Predictors (ICLR 2024 · similarity 0.84)
- Reinforced Latent Reasoning for LLM-based Recommendation (ICLR 2026 · similarity 0.82)
- Think before Recommendation: Autonomous Reasoning-enhanced Recommender (NeurIPS 2025 · similarity 0.81)
- LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs (AAAI 2024 · similarity 0.81)
- R$^2$ec: Towards Large Recommender Models with Reasoning (NeurIPS 2025 · similarity 0.80)
- Neural-Symbolic Collaborative Distillation: Advancing Small Language Models for Complex Reasoning Tasks (AAAI 2025 · similarity 0.80)
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.