Research map / Large language models / Alignment and preferences
FacetCRS: Multi-Faceted Preference Learning for Pricking Filter Bubbles in Conversational Recommender System
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 FacetCRS, an end-to-end conversational recommender that models user preference via entity, word, context, and review facets to mitigate filter bubbles, evaluated on REDIAL and TG-REDIAL.
Problem
Filter bubbles in conversational recommender systems, which intensify over time through user-system feedback loops in natural-language interaction, and the limited, sparse preference signals used by prior multi-level preference methods.
Approach
Models user preference as four complementary facets (entity from DBpedia via RGCN, word from ConceptNet via GCN, context and review via Transformers), fuses them with a multi-facet fusion function, and feeds them to both the recommendation head and a cross-attention decoder so the CRS can diversify recommendations.
Key results
- On REDIAL, FacetCRS reaches Recall@1 0.041, Recall@10 0.202, Recall@50 0.386, versus best baseline KGSF at 0.039, 0.183, 0.378.
- On REDIAL, FacetCRS lowers Iso-Index to 0.0856 (KGSF 0.1055) and raises Coverage to 7.6436 (KGSF 6.7920).
- Human evaluation on the conversation task gives FacetCRS Fluency 1.50 and Informativeness 1.39, above the best baseline KGSF (1.48, 1.37).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use multi-faceted preference learning and conversational recommendation | FacetCRS · REDIAL | Recall@10: 0.183 -> 0.202* | Helped |
| use multi-faceted preference learning and conversational recommendation | FacetCRS · TG-REDIAL | Recall@10: 0.032 -> 0.034* | Helped |
| use multi-faceted preference learning and conversational response generation | FacetCRS · REDIAL | Distinct-4: 0.282 -> 0.305* | Helped |
| use multi-faceted preference learning and conversational response generation | FacetCRS · TG-REDIAL | Distinct-3: 0.186 -> 0.228* | Helped |
| use multi-faceted preference learning | FacetCRS · REDIAL | Iso-Index: 0.1055 -> 0.0856 | Helped |
| use multi-faceted preference learning | FacetCRS · TG-REDIAL | Coverage: 9.1373 -> 10.8172 | Helped |
| add entity-facet embeddings Xe (ablation) | FacetCRS · REDIAL | Distinct-2: 0.118 -> 0.126 | Helped |
| add word-facet embeddings Xw (ablation) | FacetCRS · REDIAL | Distinct-2: 0.122 -> 0.126 | Helped |
| add context-facet embeddings Xc (ablation) | FacetCRS · REDIAL | Distinct-2: 0.119 -> 0.126 | Helped |
| add review-facet embeddings Xr (ablation) | FacetCRS · REDIAL | Distinct-2: 0.123 -> 0.126 | Helped |
Surprising findings
- Gains over the strongest baselines in recommendation recall are small (e.g., R@1 0.041 vs 0.039 on REDIAL).
Limitations
- Gains over the strongest baselines are modest in recommendation recall.
- Filter-bubble evaluation relies on offline metrics (Iso-Index, Coverage, item similarity) rather than live user feedback loops, despite the dynamic motivation.
Experimental setup
Datasets: REDIAL TG-REDIAL DBpedia ConceptNet
Benchmarks: REDIAL TG-REDIAL
Metrics: Recall@1 Recall@10 Recall@50 Distinct-2 Distinct-3 Distinct-4 Fluency Informativeness Iso-Index Coverage
Baselines: Popularity TextCNN Trans ReDial KBRD KGSF KECRS
Builds on: KBRD KGSF RevCore RGCN Transformer
Closest papers by meaning
- ConvRec-R1: Training LLM-based Conversational Recommender Systems with Reinforcement Learning (ICLR 2026 · similarity 0.86)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · similarity 0.85)
- FACE: A General Framework for Mapping Collaborative Filtering Embeddings into LLM Tokens (NeurIPS 2025 · similarity 0.85)
- Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation (NeurIPS 2025 · similarity 0.85)
- ESPRESSO: An Effective Approach to Passage Retrieval for High-Quality Conversational Recommender Systems (AAAI 2025 · similarity 0.85)
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.84)
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.