I'm using dictionary learning for structured data, something like the credit card fraud and iris dataset.

As we all know dictionary learning is similar to principal component analysis and it can be used to perform dimensionality reduction in the form of X = DA, given X is the training matrix, D is the trained dictionary and A is the sparse coding. After we have obtained a well-trained dictionary, we obtained a sparse coding approximated from this dictionary, with a few non-zero coefficients.

But how do we know which features in the matrix X contributes more or what are the reduced features? In PCA, we can examine with the variance but it seems like not the case in dictionary learning.

How do we achieve interpretability for non-image task?

(I'm using scikit-learn)