Synergy Between Sufficient Changes and Sparse Mixing Procedure for Disentangled Representation Learning