Comparing sample-level prediction accuracy on Gevers et al. Data given increasing training sizes.

Unlike OTU/k-mer based classifiers, which are trained at sample-level, our proposed model is trained at the read level before read-level results are then fused by sample-level predictor. This comparison, over 40, 160, and 400 samples in the training data shows that the read-level classifier learns predictive taxa/information and the sample-level prediction for the proposed methods are competitive with prediction from OTU tables and will allow interpretable representations shown in the subsequent sections. The obtained accuracy values are averaged, and the standard deviation is computed, over 5 experiments in which we randomly selected training-testing data splits with replacement.