Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Spatio-Temporal Graph Convolutional Network Combined Large Language Model: A Deep Learning Framework for Bike Demand Forecasting

Peisen Li, Yizhe Pang, Junyu Ren

TL;DR

This work introduces STGCN-L, a hybrid framework that fuses spatio-temporal graph convolution with a Large Language Model to forecast bike-sharing demand. It processes POI textual data via GPT-4 embeddings to augment regional features and evaluates on the Philadelphia dataset, comparing against STGCN and AGCRN. STGCN-L achieves the lowest MSE among the models, demonstrating that language-informed node features provide a modest predictive gain, with AGCRN offering competitive MAE. The study highlights the potential of integrating unstructured language data into mobility forecasting and points to weather and additional features as future improvements.

Abstract

This study presents a new deep learning framework, combining Spatio-Temporal Graph Convolutional Network (STGCN) with a Large Language Model (LLM), for bike demand forecasting. Addressing challenges in transforming discrete datasets and integrating unstructured language data, the framework leverages LLMs to extract insights from Points of Interest (POI) text data. The proposed STGCN-L model demonstrates competitive performance compared to existing models, showcasing its potential in predicting bike demand. Experiments using Philadelphia datasets highlight the effectiveness of the hybrid model, emphasizing the need for further exploration and enhancements, such as incorporating additional features like weather data for improved accuracy.

Spatio-Temporal Graph Convolutional Network Combined Large Language Model: A Deep Learning Framework for Bike Demand Forecasting

TL;DR

This work introduces STGCN-L, a hybrid framework that fuses spatio-temporal graph convolution with a Large Language Model to forecast bike-sharing demand. It processes POI textual data via GPT-4 embeddings to augment regional features and evaluates on the Philadelphia dataset, comparing against STGCN and AGCRN. STGCN-L achieves the lowest MSE among the models, demonstrating that language-informed node features provide a modest predictive gain, with AGCRN offering competitive MAE. The study highlights the potential of integrating unstructured language data into mobility forecasting and points to weather and additional features as future improvements.

Abstract

This study presents a new deep learning framework, combining Spatio-Temporal Graph Convolutional Network (STGCN) with a Large Language Model (LLM), for bike demand forecasting. Addressing challenges in transforming discrete datasets and integrating unstructured language data, the framework leverages LLMs to extract insights from Points of Interest (POI) text data. The proposed STGCN-L model demonstrates competitive performance compared to existing models, showcasing its potential in predicting bike demand. Experiments using Philadelphia datasets highlight the effectiveness of the hybrid model, emphasizing the need for further exploration and enhancements, such as incorporating additional features like weather data for improved accuracy.
Paper Structure (18 sections, 5 equations, 8 figures, 1 table)

This paper contains 18 sections, 5 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. Approach
  4. Bike Sharing Demand Prediction on Road Graphs
  5. Embeddings
  6. STGCN
  7. STGCN-L
  8. Experiments
  9. Data sets
  10. Data Processing
  11. Experimental Settings
  12. Results
  13. STGCN
  14. STGCN-L
  15. AGCRN
  16. ...and 3 more sections

Figures (8)

  • Figure 1: Yelp Business POI data example
  • Figure 2: Graph-structured traffic data
  • Figure 3: Example of Embeddings. Embeddings are numerical representations of concepts converted to number sequences, which make it easy for computers to understand the relationships between those concepts.
  • Figure 4: Architecture of spatio-temporal graph convolutional networks combined with Large Language Model. The framework STGCN-L consists of two spatio-temporal convolutional blocks (ST-Conv blocks), a LLM block as an encoder and a fully-connected output layer in the end. Each ST-Conv block contains two temporal gated convolution layers and one spatial graph convolution layer in the middle. The residual connection and bottleneck strategy are applied inside block.
  • Figure 5: Business POI in City Philadelphia
  • ...and 3 more figures