TL;DR
The "15-minute city" concept promises convenience, but accessing relevant local information remains a challenge. This paper introduces AskNearby, an LLM-powered system that bridges this gap by combining advanced retrieval techniques with a personalized "cognitive map" of user preferences. AskNearby significantly outperforms existing methods in providing accurate, context-aware local information for both active searches and passive recommendations.
What Problem Does This Paper Address?
The vision of a "15-minute city," where residents can access daily necessities within a short walk or bike ride, hinges on more than just physical proximity. It requires efficient and reliable access to information about nearby places, services, and events. Current location-based systems often fall short because they struggle to account for the nuanced interplay of spatial, temporal, and cognitive factors that shape how people actually make decisions in their local environment. This gap, termed "Local Life Information Accessibility" (LLIA), means users can't easily find timely, relevant, and personalized information to navigate their neighborhoods effectively. The paper aims to solve this by enabling both explicit information retrieval (answering user queries) and implicit recommendations (proactively suggesting relevant content).
Methods
AskNearby tackles the LLIA problem with a two-pronged approach:
- Multi-Layered Retrieval-Augmented Generation (RAG) for Active Retrieval: This pipeline uses Large Language Models (LLMs) to understand user queries and then employs three specialized retrieval modules:
- GeoRAG: Retrieves geographically proximate candidates using spatial databases and proximity queries.
- GraphRAG: Leverages semantic graphs to find conceptually and relationally relevant information, connecting entities like POIs, categories, and events.
- VectorRAG: Performs semantic similarity searches using embeddings, refining results from GeoRAG and GraphRAG by matching them against the enriched query.
- Cognitive Map Model for Passive Recommendation: This model captures a user's spatial familiarity and preferences. It calculates a recommendation score for an item based on:
- Semantic Relevance: How well the item's functional semantics (derived from TF-IWF) match the user's cognitive profile (built from frequently visited places).
- Spatial Proximity: A distance-decay function that favors closer locations.
- Public Familiarity: The aggregated popularity or visitation frequency of a location, reflecting collective cognitive salience.
Datasets
The primary dataset used for evaluation is sourced from RedNote, a popular Chinese social media platform for local discovery. This dataset consists of approximately 20,000 geotagged posts from Shenzhen, China, covering a diverse range of urban functions. The authors describe a rigorous data collection and cleaning process to ensure representativeness and accuracy. No new benchmarks were introduced; rather, existing platforms and LLMs were used as baselines.
Key Insights and Ideas
- The "15-minute city" needs more than just proximity; it needs information. Physical accessibility is only one piece of the puzzle; digital access to relevant, timely, and personalized local information is crucial for realizing the vision.
- A multi-faceted retrieval approach is key. Combining geospatial, graph-based, and semantic vector retrieval is essential for accurately understanding and fulfilling complex, context-rich local queries. Relying on any single method is insufficient.
- User cognition matters for recommendations. Simply knowing where places are isn't enough; understanding how users perceive and relate to those places (their "cognitive map") is vital for providing truly personalized and useful recommendations.
- LLMs can be grounded for local relevance. While LLMs excel at language understanding, they need specialized retrieval and knowledge grounding mechanisms (like those in AskNearby) to avoid hallucinations and provide accurate, context-specific local information.
Key Results
- Retrieval Performance: AskNearby significantly outperforms LLM-based and map-based baselines in retrieval accuracy and spatiotemporal grounding. It achieved the highest Spatial-Temporal Relevance (STR) score (83.8%) and a strong Precision@4 of 75.6%.
- Recommendation Performance: The full AskNearby system (combining spatial, public, and semantic factors) yielded the best results across metrics like Hit@K, NDCG@K, and MRR.
- Qualitative Evaluation: Human evaluations and a qualitative analysis of a sample query demonstrated AskNearby's ability to provide correct, high-quality answers with minimal hallucinations, unlike baseline LLMs which sometimes generated false addresses.
- Deployed System Performance: In real-world deployments on a university campus and a dense urban community, AskNearby saw rapid adoption and fostered significant hyperlocal engagement, with users initiating queries and leading to offline interactions.
Limitations and Future Work
- Cognitive Map Refinement: The authors plan to enhance the cognitive map model to be more adaptive and capture evolving individual user nuances.
- Data and Evaluation Expansion: Future work includes expanding data sources, geographical coverage, and refining evaluation methodologies for greater rigor and generalizability.
- Practical Utility: Efforts will focus on enriching the knowledge base with more information sources and conducting broader deployments to validate AskNearby's effectiveness across diverse urban contexts.
Summary updated