behavior_modelling_workshop

Comparison of various Classification models for mode choice This repository contains the code used in Mode choice prediction using Traditional and Machine Learning methods.

Requirements

• Python == 3.6.7
• Tensorflow == 1.14.0
• Keras == 2.2.4
• Scipy == 1.2.1
• Statsmodels == 0.9.0
• Sci-kit learn == 0.21.1

Dataset #1

Person Trip (PT) data PT_NEW_without_unknowns.zip for city of Yokohama is used. After data cleaning and removal of entries with empty values, the dataset consists of 1443009 entries. The complete dataset is divided into 2 parts as Train and Test set in the ratio 85:15.

Dataset #2

Person Probe (PP) data PP_Data_Clean_English_new.csv for city of Yokohama is also used in the study. After data cleaning, the dataset consists of 1522 entries. Further, the dataset is divided into 2 parts as Train and Test set in the ratio 70:30.

Prediction models

• Multinomial Logit (MNL) is utilised from statsmodel library.
• Random Forests (RF) from Sci-kit learn.
• Extreme Gradient Boosting (XGB) from xgboost.
• Artificial Neural Networks (ANN) from Keras.

Tips to reiterate the Notebook

• Extract PT-data csv file from PT_NEW_without_unknowns.zip.
• All the four models tested for each of the datasets and are given as single notebook.

• PT_data_analysis-overall.ipynb corresponds to PT data analysis.
• PP_data_analysis-overall.ipynb corresponds to PP data analysis.

Notes

• Train and Test data used in each model is same, therefore the metrics obtained from such models can be compared with each other.
• MNL model is used in this exercise only to predict the mode choice behavior in this stuy and not to study interaction between predictor variables.