Precision (top), recall (middle), and F₁ score (bottom) per class as a function of the fraction of the training dataset (1.55 million sources) used to train the random forest. Balancing the classes was done by taking 20% of the galaxies in the training set. All models were evaluated on the test dataset of 1.55 million spectroscopically confirmed sources, without balancing the classes. Class weights inversely proportional to the class frequency were used in the training in all cases.