How an End-to-End Geolocation Search System Bridges Recall and Ranking

Background and Challenges

Geolocation search must understand user intent, filter and rank the best POI results from tens of millions within hundreds of milliseconds. Unique challenges include rigid spatial constraints (e.g., a coffee shop 30 km away cannot satisfy a "nearby" query), coupling of intent with geographic context, multi‑factor decision making (relevance, distance, quality, price), and the need for massive candidate generation with real‑time latency.

Core Idea of End‑to‑End Optimization

Traditional cascaded architectures separate recall (maximizing recall rate) and ranking (optimizing quality), causing misaligned objectives. The proposed end‑to‑end approach aligns the whole pipeline by sharing a pre‑trained language model between recall and ranking and introducing three layers:

Unified Representation : Both stages use the same pre‑trained model to ensure consistent semantic understanding.

Knowledge Transfer : A reverse knowledge‑distillation channel lets the recall stage learn ranking preferences.

Spatial‑Semantic Integration : Geographic information is encoded as learnable representations that flow through pre‑training, recall, and ranking.

System Architecture

The system follows a "pre‑training → recall → ranking" three‑layer design.

Pre‑training Layer : Large‑scale search logs and POI corpora are combined with S2 spatial encoding for domain‑adapted pre‑training, producing geo‑aware semantic embeddings.

Recall Layer : A multi‑path parallel recall combines semantic ANN retrieval via a dual‑tower model with S2 spatial encoding and traditional inverted indexing for coverage.

Ranking Layer : A cascade of coarse, fine, and re‑ranking stages integrates relevance, distance, quality, and demand‑satisfaction signals.

Domain Data Construction

Training data comes from three sources:

Search logs containing user queries and POI interactions, encoding the core "query‑to‑POI" knowledge.

POI textual information (name, address, tags, category, business scope), with address details providing rich geographic cues.

User behavior signals such as clicks, navigation, favorites, and dwell time, offering high‑quality positive and contrastive samples.

S2 Spatial Encoding

S2 Geometry partitions the earth into hierarchical cells with unique 64‑bit IDs. Different levels (e.g., Level 12 ≈ 3.3 km², Level 17 ≈ 0.003 km²) allow flexible granularity. The hierarchy preserves spatial locality, enabling models to learn proximity from ID relationships.

During pre‑training, each POI’s latitude/longitude is converted to a chosen S2 cell ID, embedded, and concatenated with text tokens before feeding the Transformer encoder. This yields three concrete benefits:

Associating spatial cells with categories (e.g., CBD cells with office and business‑dining POIs).

Linking regions with query types (e.g., tourist‑area cells with travel‑related queries).

Learning adjacency between cells.

Semantic Recall with Spatial Fusion

Pure semantic similarity ignores distance; a semantically perfect POI 30 km away may outrank a closer one. The system injects S2 embeddings into the POI tower, allowing ANN retrieval to naturally cluster semantically similar and spatially adjacent POIs.

Training also incorporates distance as a dynamic signal: the final score combines semantic and spatial components, with spatial weight adaptively increased for strong geographic intents (e.g., "nearby") and decreased for brand searches.

Data augmentation creates weighted near‑positive samples and adversarial hard negatives that are semantically similar but spatially distant, forcing the model to respect distance.

Experiments across multiple S2 levels show that overly fine granularity leads to sparse embeddings, while overly coarse granularity loses intra‑city discrimination. The selected level balances spatial resolution with sufficient training data, yielding significant recall gains.

Multi‑Queue Recall by Distance Segment

Separate queues handle different distance ranges: a near‑distance queue focuses on high‑precision matching, a far‑distance queue emphasizes semantic relevance, and a full‑distance queue ensures long‑tail coverage. Priority‑based merging and diversity controls improve recall rates across all segments.

Co‑Optimizing Recall and Ranking

Ranking signals are distilled back to the recall model: the fine‑ranking scores become additional supervision targets for recall, aligning their "taste." Hard negatives identified by the ranking model (semantically similar but low‑scoring) are fed back to recall training, reducing semantic drift.

Ranking Model

The ranking stage uses a three‑stage cascade (coarse‑fine‑re‑ranking) and multi‑objective training combining Pairwise, Pointwise, and Listwise losses. Large language models (LLMs) assist in two ways:

Labeling: Structured prompts drive LLMs to produce multi‑dimensional relevance, distance, quality, and demand‑satisfaction annotations, filtered by confidence and verified manually (>85% consistency).

Model Distillation: LLM‑generated Listwise preferences are distilled into the online ranking model, and domain‑specific fine‑tuning aligns LLM judgments with search scenarios.

Feature fusion is dynamic: weights for relevance, distance, quality, and demand‑satisfaction adapt to the query intent, learned from massive online experiments.

Spatial Features in Ranking

Beyond recall, the ranking model incorporates multi‑level spatial features derived from S2 grids, precise and binned distances, and category‑specific distance distributions, enabling nuanced reasoning such as:

"Nearby coffee shop" queries penalize distant POIs despite high semantic similarity.

Brand searches like "Quanjude roast duck" down‑weight distance.

Region‑specific queries (e.g., "Sanlitun restaurants") anchor the spatial anchor to the named area rather than the user's current location.

Data‑Model Co‑Evolution

A closed‑loop pipeline continuously generates high‑quality labeled data, which fuels both ranking model fine‑tuning and domain‑specific LLM instruction tuning, creating a positive feedback flywheel for ongoing improvement.

Future Directions

Generative search using LLMs to produce structured results directly from intent.

Multimodal fusion of images, video, and other media into POI representations.

Real‑time personalization based on instantaneous user context (time, weather, transport mode).

Finer spatial intelligence such as road‑network distances, traffic‑aware reachability, and commercial‑area heatmaps.