Notes. In this example, we have trained models to select quiescent galaxies from the Int Wide mock catalogue in the redshift bin 2 < z < 2.5, using ugriz, Euclid, W1, W2, and 20 cm photometry and colours, and with pre-binning by redshift, as described in Sect. 5.1.2. A single random seed is used for the train/test split and base learners to allow a relatively controlled comparison between methods. The columns are as follows: (1) The learning algorithm; also shown are the final F1-scores after the models produced by the 5 default base-learners have been ensembled using meta-learners; (2) F1 -score when missing values are imputed with the constant value -99.9; (3) F1 -score when imputing with the average value of a feature; (4) F1 -score when missing values are dynamically imputed with either the constant value −99.9 (tree-based learners) or the mean of a feature (MLPClassifier and KNeighborsClassifier); (5) F1-score when imputing with the median value of a feature; (6) F1-score when imputing with the minimum value of a feature. Some F1-scores differ significantly to those presented in Sect. 5, since here we use only a single random seed instead of averaging results over multiple pipeline runs that use different random seeds. Note that in this test, MLPClassifier was unable to correctly identify any quiescent galaxies when missing values were imputed with −99.9; nonetheless, this failure did not appear to be detrimental to the final meta-learner ensemble.