ALF-Score++, a novel approach to transfer knowledge and predict network-based walkability scores across cities

Area of Science:

• Urban Planning
• Geographic Information Systems (GIS)
• Machine Learning

Background:

• Walkability is crucial for public health but challenging to assess comprehensively.
• Existing methods for calculating walkability are time-consuming and resource-intensive.
• Dynamic urban environments necessitate adaptable walkability assessment tools.

Purpose of the Study:

• To develop and evaluate a transfer-learning framework for predicting city-level walkability scores.
• To address limitations in current walkability assessment methods by leveraging prior knowledge.
• To reduce the computational burden and data requirements for walkability prediction in new urban areas.

Main Methods:

• Application of transfer-learning techniques to predict walkability scores.
• Development of ALF-Score++, a model that reuses trained predictive models.
• Training and testing models using data from St. John's, NL, and Montréal, QC, with predictions for Kingston, ON, and Vancouver, BC.

Main Results:

• Transfer-learned models using Multilayer Perceptron (MLP) achieved a Mean Absolute Error (MAE) of 13.87 units.
• Direct training using Random Forest on personalized clusters yielded a lower MAE of 4.56 units.
• Demonstrated the feasibility of reusing trained models for predicting walkability in unseen cities.

Conclusions:

• Transfer-learning offers a viable strategy to accelerate walkability assessment across diverse urban settings.
• While direct training on specific clusters shows higher accuracy, transfer-learning significantly reduces training time and resources.
• ALF-Score++ provides a scalable solution for estimating walkability, supporting urban health and planning initiatives.