It is easy and yes, its detail may be followed as here: https://towardsdatascience.com/the-art-of-finding-the-best-features-for-machine-learning-a9074e2ca60d
If the relative target value is calculated in an unsupervised manner, that is not using the information from the actual ground truth, then I don't see any reason as why it can't be used as a feature.