The ability to predict travel mode choices given trip features is vital for addressing mobility needs. Hence, travel mode choice classification models are developed. However, the number of real trip instances used to train such models is typically limited. Moreover, changes in the probability of observing individual trip features occur. This is partly due to changes in public transport services. As evolving interest in transport modes is also observed, this means that both virtual and real concept drift typically occur in trip data. In this work, we focus on the challenge of predicting mode choices with a limited number of real trip records present in trip data streams reflecting the evolution of transport systems and travellers’ preferences. We address this challenge by proposing a method generating synthetic unlabelled trip examples including estimated features of trip mode alternatives. Such instances can be inserted in arbitrary volumes and at different period(s). Hence, this work also addresses how real data streams should be mixed with synthetic instances and how the impact of such scenarios can be analysed. We evaluate the data generation method with five real urban mode choice prediction tasks. We identify which synthetic data scenarios are the most beneficial to the semi-supervised learning process. Our results show that mode choice models of typically higher accuracy are obtained once synthetic instances are used, with accuracy gains even exceeding 9%. The results of our experiments also contribute to the possible future development of semi-supervised stream mining methods.