LLM agents plus knowledge graphs boost urban socioeconomic prediction accuracy

Socioeconomic prediction—estimating indicators like population and commercial activity in urban regions—is critical for city planning and policy-making. Traditional approaches rely on knowledge graphs (KG) to model diverse urban data, then apply graph representation learning. However, these methods depend heavily on manual heuristics to extract task-relevant features and often ignore correlations between different indicators, limiting accuracy. In a new arXiv paper (2411.00028), researchers from Tsinghua University tackle this by combining the reasoning power of large language models (LLMs) with the structured knowledge of graphs.

Their framework first builds an urban knowledge graph (UrbanKG) from multi-sourced data and fine-tunes an embedding language model to generate semantic entity embeddings. To avoid heuristic feature extraction, they use LLM agents to automatically identify task-specific meta-paths—meaningful connection patterns in the graph that matter for each prediction task. A semantic-guided attention module then fuses knowledge along these paths. The key innovation is cross-task communication: LLM agents for different tasks (e.g., population vs. commercial activity) collaborate to generate more diverse meta-paths, and embeddings from different tasks are adaptively merged at the KG level. Experiments on two datasets demonstrate significant accuracy gains over existing graph-based baselines, highlighting the advantage of synergizing LLM reasoning with graph representation learning. This approach offers a scalable, data-driven path to understanding urban regions and supporting decision-making.