From reviews to preference profiles: LLM-based transfer for cross-domain recommendation

Cross-domain recommendation under cold start is challenging because users have no history in the target catalog and descriptions differ across domains (e.g., Movies vs. Books). We propose LIPA ( Language-Induced Profile Abstraction ) language-based profile abstraction layer that transfers preferences without any training on the target domain. A language model condenses multiple reviews into one neutral paragraph per user (stable preferences) and per item (salient themes), yielding a shared semantic representation that is independent of domain labels and rating scales. These profiles are then encoded and matched by cosine similarity to generate top- k recommendations. We evaluate on Amazon Movies to Books with two protocols: (i) full-catalog, item-level ranking and (ii) a genre-filtered setting that restricts candidates to broad genres surfaced by the user profile. Compared to applying the same retrieval directly on raw, unsummarized text, LIPA improves ranking quality by ∼ 20–23% on Normalized Discounted Cumulative Gain @ 10 and Mean Average Precision @ 10, increases Recall @ 10 (0.487 vs. 0.387), achieves near-ceiling hit rates ( ≥ 0 . 98 ), and broadens catalog coverage (0.72–0.84), indicating stronger exposure to long-tail items. A sensitivity study shows that using more textual evidence per item and a lower generation temperature produces clearer profiles and better rankings. By summarizing heterogeneous reviews into concise, readable profiles before retrieval, LIPA reduces cross-domain mismatch, maintains interpretability, and delivers scalable, training-free, cross-domain recommendations suitable for cold-start scenarios.