Balance features relevance in unsupervised learning

Hello everyone,

I would like to perform unsupervised learning with structured data, where some features are more relevant than others.
Is there any way to specify that a feature is more important? It would correspond to giving to the network some prior knowledge; I understand that the network may fit on features that may seem not relevant for humans (which is fine, as long as it still uses the important features).
I could also use a mix of supervised and non-supervised training, but I am curious to see if this feature exist somehow.

Thank you for your replies :slight_smile: