Implicit GPS-based bicycle route choice model using clustering methods and a LSTM network.

Area of Science:

• Urban planning and transportation science
• Machine learning applications in mobility
• Geographic Information Systems (GIS) and spatial analysis

Background:

• Increasing global popularity of cycling for health and environmental benefits.
• Urban policies promote cycling through infrastructure and bike-sharing systems (BSS).
• Need for optimized urban policies necessitates understanding and predicting cyclist behavior.

Purpose of the Study:

• To develop a prediction-centered bicycle route choice model.
• To move beyond classical methods relying on external factors and choice sets.
• To leverage deep and machine learning algorithms for enhanced predictive capacity.

Main Methods:

• Utilized deep and machine learning algorithms on GPS tracks, replacing explicit factors with learned representations.
• Employed DBSCAN clustering to identify preferred road segments from GPS data.
• Developed a path generation method weighting road graph weights and trained a Long Short-Term Memory (LSTM) network for cluster retrieval.

Main Results:

• The developed model generates bicycle route tracks that are more similar to original GPS tracks.
• Outperformed traditional shortest path algorithms and a prominent path computation service in track similarity.
• Effectively learns cyclist preferences directly from GPS data without external variables.

Conclusions:

• The prediction-centered model offers a more accurate approach to understanding and predicting bicycle route choices.
• Deep learning on GPS data provides a powerful alternative to traditional discrete choice models in transportation.
• This method can aid in optimizing urban cycling infrastructure and policies for better cyclist experience.