The technological advancements of the past decades have allowed transforming an increasing part of our daily actions and decisions into storable data, leading to a radical change in the scale and scope of available data in relation to virtually any object of study. In the field of discrete choice analysis, such abundance of data has the potential to expand our understanding of human behavior, but this prospect is limited by the poor scalability of discrete choice models (DCMs). This thesis presents a series of innovative methodological developments for the specification and estimation of DCMs and other statistical models using large-scale datasets. Our main contributions consist in practical methods inspired from the success of machine learning in harnessing and exploiting ever-larger amounts of data. By making these methods publicly available, we offer valuable tools to researchers and practitioners across various domains.