PathGPT: A New Paradigm for Integrating Large Language Models into Personalized Path Recommendation

PathGPT is an innovative path recommendation system whose core lies in redefining path recommendation as a natural language generation task. By combining Retrieval-Augmented Generation (RAG) technology and leveraging the comprehension capabilities of large language models (e.g., the quantized version of Qwen2.5:14b-instruct), it provides personalized routes. It significantly outperforms basic LLMs on real-world datasets from multiple cities, offering a flexible and scalable new framework for the path recommendation field.

Traditional path recommendation algorithms (e.g., Dijkstra) struggle to capture users' personalized preferences. Although data-driven machine learning models have made progress, once trained, they can only generate paths that conform to the distribution of training data. When facing new scenarios, retraining is required, which is costly and lacks flexibility.

Innovative Ideas

Redefine path recommendation as a natural language generation task. Core advantages include a unified model architecture, zero-shot adaptation capability, and RAG-enhanced generation.

System Architecture

Adopts a RAG architecture, including modules for vector database, context generation, prompt generation, and subgraph construction.

Supported Path Types

most_used (most frequently used), fastest (quickest), shortest (shortest distance), touristic (scenic), highway_free (highway-free).

Base Model

Uses the 4-bit quantized version of Qwen2.5:14b-instruct, deployed locally via Ollama, requiring only about 10GB of VRAM.

Datasets

Validated on real-world data from three cities: Beijing, Chengdu, and Harbin, including map-matched trajectories, OSM maps, and POI data.

Key Results

• Scenic route recommendation: The precision/recall of PathGPT@3 is significantly higher than that of basic LLMs (e.g., precision in Chengdu increased from 30.14% to 88.34%);
• Highway-free route recommendation: PathGPT@3 also leads by a large margin in performance;
• Impact of retrieval quantity: Performance improves steadily with increasing top_k, but marginal benefits diminish. Experiments show that PathGPT outperforms basic LLMs in all scenarios, with recall rates increasing by more than 3 times at maximum.

Application Scenarios

Intelligent navigation (personalized routes), travel planning (scenic routes), logistics delivery (highway-free routes), urban planning (travel preference analysis).

Future Directions

Expand to fields such as indoor navigation, drone path planning, and abstract concept path recommendation; continue to optimize the RAG architecture and local deployment solutions.

PathGPT is an important breakthrough in the path recommendation field. By combining large language models with geographic information systems, it improves recommendation quality through task redefinition and RAG technology, provides a flexible and scalable framework, and opens up new directions for the future development of personalized path recommendation.