How Gaode’s AI Generates Personalized Travel Routes from Billions of Trips

Overview

Gaode (AutoNavi) presents an AI‑powered Interest Route (LOI) generation system that expands the "what to eat" and "where to go" decision‑making of its "Street‑Scanning Ranking" product into full, executable travel itineraries. By leveraging billions of real travel records and crowd‑sourced experience data, the system produces both classic, crowd‑approved routes and niche, theme‑based itineraries.

Unified Data Foundation

A high‑quality, multi‑dimensional data base is essential for both classic and thematic route generation. Gaode builds a processing pipeline that integrates:

Proprietary navigation, POI heat, and interaction data from Gaode Maps.

External travel experience data such as user comments, notes, and publicly available guides.

These sources are cross‑validated to ensure factual accuracy and to fill gaps in the core data.

Technical Implementation Steps

Core Data Insight & Experience Validation

Core data: Massive real‑world travel logs (navigation behavior, POI popularity, interaction metrics) serve as the factual backbone.

Fusion validation: User‑generated experience data is matched against core data to verify authenticity and uncover missing details.

Multi‑source Retrieval & Query Rewriting

Searches multiple content platforms for rich information.

Applies a query rewrite strategy that expands POI names into aliases and combines them with activity keywords (e.g., "one‑day tour", "route", "guide") to improve recall.

Multi‑dimensional Relevance Filtering

Uses freshness weighting to prioritize recent content.

Builds a quality model based on interaction signals (likes, saves, shares) to discard low‑engagement items.

Applies a relevance model to remove ads, geographically mismatched text, and non‑extractable passages.

Schema Definition & Information Extraction

Defines a concise Schema covering city, POI, activity type, transport mode, and trip length.

Leverages large language models (LLM) to extract structured fields from unstructured travel notes and populate the schema for downstream processing.

Dual‑Track Generation Strategy

The system runs two parallel pipelines to satisfy different user needs.

Track 1 – Classic Routes (Maximum Commonality)

Targets first‑time visitors by delivering the most popular, low‑risk itineraries.

Data Application

Heat quantification: Uses real navigation and click heatmaps to compute a "national popularity" score for each POI.

Trip analysis: Anonymized travel logs reveal frequent POI connections, forming the factual basis for route construction.

POI Normalization Clusters synonymous POI names (e.g., different spellings of the same landmark) to aggregate popularity.

Route Representation & Clustering

Encodes each route with features such as days, visited locations, entry/exit points, and key transport hubs.

Combines two heat signals – community‑derived popularity and actual foot‑traffic – to weight POIs in the feature vector.

Scoring & Selection A scoring model prefers routes with comprehensive popular POI coverage, inclusion of unique spots, and high overall itinerary richness, yielding a "golden route" for each cluster.

Track 2 – Thematic Routes (Personalized Niches)

Serves users with specific interests (e.g., motorbike mountain rides, seasonal flower viewing).

Feature Verification Matches experience‑derived signals (e.g., "motorbike mountain") against Gaode’s road attribute data (slope, curvature) and navigation traffic to confirm relevance.

Demand Matching Uses anonymized demand tags ("self‑drive enthusiast", "outdoor lover") to align POI preferences with route generation.

Theme Extraction Combines demand tags, keyword trends, and road attributes to generate specialized itineraries such as "motorbike mountain" or "autumn foliage" routes.

Unified Output: Entity Linking & AIGC Content Generation

Both classic and thematic pipelines produce a route skeleton that is then enriched with AI‑generated titles, highlights, and images.

High‑Precision Entity Linking Employs a multi‑recall strategy to fetch candidate POIs, then uses LLM semantic understanding to match textual mentions to exact map entities.

Controlled Text Generation Generates user‑facing copy with a balance of authenticity, readability, and appeal, enforced by JSON constraints and citation mechanisms. A "persona" approach (e.g., food critic, historian) adds depth.

AI‑Driven Visual Selection Filters user‑uploaded images with compliance models, vectorizes remaining images, and removes near‑duplicates to present a diverse, high‑quality visual set.

Technical Outlook

The AI‑driven LOI system is now a core engine for Gaode’s "Street‑Scanning Ranking" products. Future work includes scaling thematic discovery without prior knowledge, improving LLM extraction of unstructured experience data, and expanding content formats to video and voice for immersive travel experiences.

Conclusion

By tightly integrating massive real‑world travel data with AI‑generated content, Gaode transforms static point‑of‑interest lists into dynamic, personalized travel plans that guide users from "where to go" to "how to explore" efficiently and enjoyably.