Abstract

This master thesis explored the use of machine learning to understand and identify traveler

behaviors based on survey data. With the growing importance of sustainable travel in societal

development, it is crucial to analyze transportation habits effectively. The study used data from

Uppsala’s public transportation system (UL) collected between 2017 and 2023, including

attributes such as age, gender, residential area, car availability and satisfaction with public

transportation. Four traveler types were defined: drivers, switchers, public transport users, and

infrequent travelers. The primary objective was to develop machine learning models capable of

predicting and categorizing individuals into these traveler types while identifying the key factors

influencing these classifications. Three types of machine learning models were evaluated:

Artificial Neural Networks with Multilayer Perceptrons and Backpropagation (ANN-MLP-BP),

Gradient Boosting Trees (GBT), and Logistic Regression (LR). The study found that ANN-MLP-

BP models, particularly those trained on 15 attributes, performed the best, achieving higher F1-

scores through both macro and weighted averages. The two most important key factors was

identified as postal code and car availability. Additionally, the study addressed the challenge of

imbalanced data by training and comparing models on both the complete and an undersampled

dataset. Results indicated that models trained on the complete dataset excelled in classifying

the majority classes, drivers and switchers, while those trained on the undersampled dataset

better identified the minority classes, public transport users and infrequent travelers. This

highlights the importance of choosing datasets and the potential of controlled undersampling to

enhance model performance for underrepresented classes.