Model performance estimated through cross validation for the combination of Mg II h&k, C II, and Si IV. As baseline for comparison we have used Mg II h&k (left panel), reaching a lower TSS than in the single line experiments with VAE preprocessing. In the experiments with the completely newly trained convolutional neural network (centre panel) and the pre-learned features (right panel) for the line combination we can observe an increase compared to the baseline experiment with Mg II h&k.