LLMs Generate Candidates for Long-Tail Vacation Rentals.

Vacation rental marketplaces, like Vrbo, face a common challenge: a small percentage of popular listings receive most user engagement, leaving a vast "long tail" of new, niche, or seasonal properties with insufficient behavioral data for effective recommendation systems. Traditional collaborative filtering methods, such as item-based k-nearest neighbors (IBKNN), struggle to generate candidates for these sparsely interacted listings. This paper introduces a novel, training-free candidate generation pipeline powered by Large Language Models (LLMs). The system leverages only static property metadata. An off-the-shelf LLM synthesizes diverse semantic queries for each property, which are then embedded by a pre-trained text encoder. An approximate nearest-neighbor index retrieves relevant candidates from a massive catalog of 11.7 million properties. A "Union fusion" strategy merges these LLM-generated candidates with those from IBKNN, ensuring that the behavioral channel's ordering is preserved and no degradation occurs for well-served properties. The combined pool is then re-scored by a downstream learning-to-rank model. The system successfully extended candidate coverage to tens of thousands of properties previously unreachable by IBKNN, delivering its largest gains in the long-tail segment. Notably, the Union fusion strategy also significantly closed the recall gap between smaller open-weight LLMs and frontier API models, supporting the feasibility of self-hosted, smaller model deployments at scale.