Who do we choose to spend our leisure time with? Insights from discrete choice and machine learning models

This thesis investigates the determinants of social contact selection in leisure
activities, integrating traditional discrete choice models (DCM) with machine
learning (ML) techniques to enhance model specification and explanatory power.
The research utilises a dataset collected through snowball sampling in Switzerland, capturing a wide range of respondents’ characteristics and the characteris-
tics of their social network members. Employing multinomial logit models, the
study reveals how dyadic variables such as age homophily, gender homophily, relationship duration and so on influence leisure activity choices with various social
contacts. Furthermore, the incorporation of machine learning techniques, particularly Shapley Additive explanations (SHAP), enriches the model. SHAP high-
lights predictors that might otherwise be overlooked. It also provides insight into
the direction and impact of these predictors, facilitating their interpretation before running a choice model. The findings extend the current understanding of
social interaction patterns, advocating for consideration of social networks in data
collection and modelling of who people interact with. This thesis also uses machine learning to assist choice modelling, offering an additional tool for analysing
social contact preferences in leisure contexts, which implies an additional set (potentially large) of explanatory variables where machine learning models could be
useful.