In some paper, such as, Nonnegative Matrix Factorization: A Comprehensive Review, I see the interpretability of Nonnegative matrix factorization (NMF). However, I don't know the means of this. How to show the interpretability of NMF by a small qualitative example on a toy data?
In addition, what is interpretability? Especially in Non-negative Matrix Factorization, how to understand interpretability?
Consider an example of prediction of user movie ranking, having initially one sparse matrix "users-movies". Basic assumption is that each user is associated with a vector of preferable genre and each movie is associated with the vector of the same length, that shows its genre mixture.
In such a case nonnegative matrix decomposition has clear meaning, like the one written above, that is it's interpretable in some way.