Imbalanced dataset - feature extraction and unsupervised learning

[FredS1000]

Hey everybody,
this discussion could be a bit more general than the issue i found regarding this issue:

#522

I have an imbalanced dataset on audio files with two classes. I am struggeling to improve my accuracy on the class with minor representation. Currently i am preprocessing the audio files to fbank via the preprocessing library and analysing the resulting data by a stacked cnn.
On the one hand i want more control about the preprocessing to do feature extraction (Mel- Spectrum Coefficients e.g.) and use tree based methods to analyze the results. On the other hand i would like to do some unsupervised training on the larger part of the data and do anomaly detection (autoencoders with manually built thresholds).

Are there any suggestions to do this in ludwig?
Thanks a lot. I hope this is the right place to ask :)