Diffusion probabilistic model for bike-sharing demand recovery with factual knowledge fusion

Area of Science:

• Data Science
• Transportation Systems
• Machine Learning

Background:

• Bike-sharing systems generate valuable data for urban mobility analysis.
• Accurate bike flow patterns are crucial but often hindered by missing data.
• Existing methods struggle with the complexities of bike-sharing demand recovery.

Purpose of the Study:

• To develop a robust solution for bike-sharing demand recovery (Biker).
• To address challenges including observation uncertainty, complex dependencies, and environmental factors.
• To introduce a novel diffusion probabilistic approach for accurate data imputation.

Main Methods:

• Proposed DBiker, a diffusion probabilistic model integrating factual knowledge.
• Incorporated a conditional Markov decision-making process for imputation.
• Introduced a Flow Conditioner and Factual Extractor to handle dependencies and environmental data.
• Devised a self-gated fusion layer for adaptive knowledge selection.

Main Results:

• DBiker outperforms several baseline methods in bike-sharing demand recovery.
• The model successfully forecasts missing observations through progressive steps.
• Demonstrated superiority on three real-world bike systems.

Conclusions:

• DBiker offers a significant advancement in addressing missing bike flow data.
• The fusion of diffusion models and factual knowledge proves effective for Biker.
• DBiker enhances the understanding of bike flow patterns despite data gaps.