A community-driven recommender system for speed safety camera policy using large language models and knowledge graphs

Speeding is a persistent contributor to roadway fatalities, accounting for over 30% of traffic deaths in the U.S., prompting the widespread use of Speed Safety Cameras (SSCs) as a proven countermeasure. However, despite their effectiveness, public skepticism, which is driven by concerns over fairness, transparency, and revenue motives, has limited the scalability and sustainability of SSC programs. Existing research lacks systematic integration of public perception into policy frameworks and fails to fully leverage advancements in artificial intelligence for context-aware recommendation. Addressing this gap, this study develops a scalable policy recommender system for SSC deployment in Oregon by fusing public sentiment with authoritative guidance using a novel hybrid framework that integrates Large Language Models (LLMs), Knowledge Graphs (KGs), and Retrieval-Augmented Generation (RAG). A statewide survey of 1,000 residents captured public opinion across six policy dimensions, which, along with state legislation (HB4109), FHWA guidelines, and best practices, formed a multi-source knowledge base stored in a Neo4j graph. The system retrieves both structured and unstructured context to support LLM-based generation of interpretable policy recommendations. Through Monte Carlo simulations, sentence similarity benchmarking, and RAGAS metrics, the proposed framework achieved the highest performance in accuracy, faithfulness, and context recall compared to baseline LLM and RAG pipelines. The results underscore the effectiveness of combining symbolic reasoning with semantic retrieval for generating transparent and community-aligned SSC policy guidance, offering a transferable blueprint for jurisdictions aiming to align traffic safety interventions with public trust and domain expertise.