I have suggested a semi-supervised approach for the hierarchical multi-label classification task. I have included the MLSMOTE oversampling technique as a pre-processing step, and then evaluate the impact of the oversampling on the semi-supervised learning Below, the obtained results in terms of Micro-f1 and Macro-f1 scores.
These two figures are under the labeling ratio 50% labeled, 50% unlabeled.
My question is, is it possible to say that oversampling is beneficial at the end of co-training, because at the last iteration, we get the same Macro-F1 scores, we obtain higher Micro-F1 scores for the method without oversampling ? the reason could be that the latter is biased towards the majority labels which increases the Micro-f1 scores in detriment of Macro-F1 scores. Thanks for help.