Interpretable Urban Metro Flow Prediction via Spatio-Temporal Graph Learning and Large Language Models

This repository contains the official implementation of MFP-LLM, a novel framework for urban metro passenger flow prediction that integrates spatio-temporal graph learning with instruction-driven large language models (LLMs).

🚇 The goal is to achieve accurate, robust, and interpretable metro flow prediction under full, few-shot, and zero-shot settings.

---

✨ Key Features

• Spatio-Temporal Encoder: Attention-based encoder that jointly captures spatial and temporal dependencies in metro networks.
• Representation Alignment Module: Transforms structured spatiotemporal features into token sequences compatible with LLMs.
• LLM-Enhanced Prediction: Leverages reasoning and generalization capabilities of pre-trained LLMs with minimal task-specific fine-tuning.
• Interpretability: Generates natural language explanations of prediction results to improve transparency and decision support.
• Strong Generalization: Robust performance across datasets and unseen conditions.

---

📊 Datasets

We use two large-scale real-world metro datasets:

• HZMetro

  • Hangzhou Metro system
  • Duration: Jan 1 – Jan 25, 2019 (25 days)
  • 80 stations, aggregated traffic statistics
  • Time resolution: 15-minute intervals
  • Records both inflow and outflow of passengers

• SHMetro

  • Shanghai Metro system
  • Duration: Jul 1 – Sep 30, 2016
  • 288 stations, high spatiotemporal resolution
  • Time resolution: 15-minute intervals
  • Provides fine-grained passenger flow dynamics

Dataset Download

Baidu Netdisk: https://pan.baidu.com/s/1lesAk4WOfBQtg0a0XgDfvA
Extraction code: np5p

For more implementation details, refer to run_MFP-LLM.py.

Installation

Requirements

torch==2.2.2

accelerate==0.28.0

matplotlib==3.7.0

numpy==1.23.5

pandas==1.5.3

scikit_learn==1.2.2

tqdm==4.65.0

transformers==4.31.0

deepspeed==0.14.0

🚀 Quick Start

Clone the repo

git clone https://github.com/your-username/MFP-LLM.git
cd MFP-LLM