Comparative analysis for Machine Learning vs Traditional Gravity Model
This repository contains the code for a research paper on a comprehensive transport model for trip predictions. The research investigates the data requirements for trips at the city level with considerable diversity and compares simulations based on SIM and Machine Learning models.
The full research paper can be found here (https://www.sciencedirect.com/science/article/pii/S2772586323000175). The research focuses on developing a machine learning model for predicting trip patterns in urban areas based on various factors such as time of day, day of the week, weather conditions, and more. The model uses data from trip logs provided by Telia company, which has been anonymized to protect privacy.The research also includes network analysis to calculate the shortest distance between all urban districts, which is used for further analysis in the spatial interaction model (SIM) as well as the machine learning model for the parameter calculations of cost of transport. The spatial interaction model is developed based on Wilson's ideology of Gravity theory.
The interactive dashboard is available over the hosted platform pythonanywhere, powered specifically for Python web applications. The programming in Python is based on Dash and uses IBM Watson (Cloud Technology) as a backend where the machine learning model is deployed. The SHAP analysis for the XGBoost regression model is included in the Google Colab notebook.
The interactive dashboard is available over the hosted platform 'pythonanywhere', powered specifically for Python web applications. The programming in Python is based on Dash and uses IBM Watson (Cloud Technology) as a backend where the ML model is deployed. The files and code are available over the web interface here: [InteractiveApp](http://transportmodelapp.pythonanywhere.com/). (Sign in Username: 'Omkar' Password: '12345') PS. The backend over IBM Watson has been disconnected due to a security breach. We are shifting to AWS to upgrade to a better, safer backend.
The machine learning model is deployed in the backend over IBM Watson. The deployed model consumes the configuration at an average of 2.5 Capacity Unit Hour (CUH) for every run of the analysis. Even with a low processing capacity, the model has relatively fast processing for analysis.
It is possible to develop the model by providing the ability to upload the parameters and invoke the road network over OSM or Mapbox GL JS (for the faster and more accurate analytical ability of cost of transport, including the waiting times). Using GPU (TensorFlow) would optimize the model execution time further when we use real-time data, depending on the matrix size. The SHAP analysis for the XGBoost regression model is also included here.
To use the code, you must install Python 3 and the required libraries given in the requirements file. Contributions to the code are welcome. If you have any questions or suggestions, please open an issue in the repository. WSP Sweden acquired the trip data from Telia company and has been used for this research in confidentiality. It can be made available upon request and a proper confidentiality agreement with Telia and WSP Sweden organization.