🤖 AI Summary
This work addresses the failure of collaborative filtering in vacation rental platforms due to sparse interactions with long-tail listings. The authors propose a training-free LLM-based candidate generation method that leverages only static listing metadata: it uses a general-purpose large language model to generate semantic queries, which are then encoded by a pretrained text encoder and retrieved via approximate nearest neighbor (ANN) search. These candidates are fused with those from an existing item-based k-nearest neighbors (IBKNN) channel using a union strategy. This approach is the first to effectively integrate LLMs with static metadata without fine-tuning, significantly improving long-tail coverage while preserving recommendation quality for popular listings. It achieves the largest gains on tens of thousands of listings previously unserved by IBKNN, consistently matching or outperforming baselines across all recall@K metrics, with further improvements after Learning-to-Rank reranking, and narrows the recall gap between a 3B open-source model and state-of-the-art API models to within 1%.
📝 Abstract
Vacation rental marketplaces face a structural imbalance on the supply side: a small fraction of properties receive most user interactions, while the long tail of new, niche, and seasonal listings generates too little behavioral signal for collaborative filtering to serve effectively. At Vrbo, item-based k-nearest neighbors (IBKNN) is a core candidate generation channel, but leaves tens of thousands of properties with no candidates and produces weak neighborhoods for sparsely interacted ones. We present a training-free, LLM-based candidate generation pipeline that complements IBKNN using static property metadata alone. An off-the-shelf LLM synthesizes diverse semantic queries per property, a pre-trained text encoder embeds them, and an approximate nearest-neighbor index retrieves candidates from an 11.7M-property catalog. A Union fusion strategy merges these with IBKNN while preserving the behavioral channel's ordering, guaranteeing no degradation on well-served properties, and a downstream learning-to-rank model re-scores the fused pool. Evaluated on 1.6M focal properties, the system extends candidate coverage to tens of thousands of properties IBKNN cannot reach, delivers its largest gains on the long-tail segment where behavioral methods are weakest, and matches or beats IBKNN at every K on shared properties. A downstream learning-to-rank stage further lifts the fused pool, yielding a complete candidate generation and re-ranking stack that serves the long tail without regressing well-served properties. We additionally show that Union fusion collapses the recall gap between a 3B open-weights LLM and frontier API-based models from 27-46% to under 1%, supporting self-hosted small-model deployment at marketplace catalog scale.