Poi reranking balances relevance proximity and quality with two stage model
H2CE: Modeling Geo-Semantic Interactions for POI Reranking with Heterogeneous Two-Stage Cross-Encoders
Information Retrieval
Summary
When searching for places nearby, it's tricky to decide whether a closer spot that's only somewhat relevant is better than a farther one that's a better match and has higher ratings. The authors introduce a method called H2CE that uses two steps to rank places faster and more accurately. It combines words describing reviews and exact numbers like ratings in smart ways, comparing top choices directly to fine-tune the results. Their tests show noticeable improvements over previous ranking methods.
What this means in practice
- •For local search engineers: Improve search result ranking by integrating semantics, location, and quality scores in a scalable two-stage model.
- •For recommendation system developers: Build rerankers that combine text-based and numerical signals for better user-relevant suggestions under real-time constraints.
Authors
Zhengwei Bai, Moreno D'Incà, Danielle Class, Alessandro Moschitti
Abstract
Point-of-Interest (POI) reranking in local search must model query-conditioned tradeoffs among lexical semantics, geospatial proximity, and numerical quality signals such as rating and review count, while remaining practical under real-time serving constraints. A close POI may only partially satisfy the query intent, while a farther one may offer stronger semantic and quality evidence. We present H2CE, a Heterogeneous Two-stage Cross-Encoder for latency-bounded POI reranking. H2CE represents numerical attributes in two complementary ways: bucketized natural-language descriptors are inserted into the cross-encoder input to support semantic--numeric attention, while exact scalar values are processed by dedicated MLPs to preserve magnitude information. The resulting semantic and numerical embeddings are fused through latent-space aggregation, enabling nonlinear interactions beyond scalar weighted sums. H2CE then applies a two-stage architecture: Stage 1 scores all candidates pointwise for scalable filtering, and Stage 2 performs head-to-head pairwise comparison among the top-$K$ candidates with Copeland aggregation, making fine-grained relative tradeoffs explicit while reducing pairwise cost from O(N^2) to O(N+K(K-1)). On a 5,743-query local search test set, H2CE achieves 67.48% NDCG@5, improving over XGBoost LTR by +22.82% absolute and over a zero-shot LLM reranker by +35.89%. The pairwise stage adds +1.98% NDCG@5 over the pointwise model alone. Ablations confirm the value of numerical features, latent aggregation, top-K pairwise reranking, and aligned training.