Evolved massive stars dominate stellar feedback and drive the chemodynamical evolution of their host galaxies. Therefore, the identification of new evolved massive stars is an essential task that will soon become impractical, as next-generation observing facilities deliver large-area photometric surveys at an unprecedented scale. The need for automated photometric classifiers that can efficiently exploit unlabelled data (which will be the most abundant in these new surveys) is thus evident. In this work, we explore the potential of semi-supervised learning techniques to improve the classification of evolved massive stars by leveraging unlabelled data. Our findings show that cluster-based semi-supervised methods outperform supervised classifiers when only a few labelled examples are available.