I want to use NMF to separate true sources from data. My data is in group structure with overlap elements. For example (in the smaller version)
- group1: contains A,B,C,D,E,F,G patterns
- group2: contains B,C,G,H,I,J patterns
- group3: contains B,K,L patterns
In the data, A,B,C,D are detected, and group A should be the result as true source. I try using l1-norm as a penalty term for group-sparsity but the result does not look nice yet. I think the problem is l1 may not be enough to restrain since B appears in every groups. Plus, E,F,G in group A are allowed to be missing.
What can I do to make my NMF better in this particular situation? In fact, each group has about 20-400 patterns and overlap can be about 10%