In the era of new telescopes, new challenges are being presented, some of which can be overcome through machine learning. Spectroscopic follow-up of every supernova discovered in LSST’s vast data stream is unrealistic. Reliable classification of Type Ia Supernovae is important if we want to use these discoveries for constraining cosmological parameters. My work involves developing methods to optimise a training sample of supernovae for photometric classification with machine learning. This training data will come from spectra obtained with 4MOST by a rapid follow-up of LSST’s early discoveries. In TiDES (Time-Domain Extragalactic Survey), we are therefore working towards maximising survey overlap with LSST to acquire a large, good-quality training set. Performance of classification is dependent on multiple factors, including size of the training sample, machine learning algorithms used and class representativeness in terms of supernova magnitudes, redshifts, as well as the common machine learning challenge of balancing classes (i.e. the different types).