Strange behaviour of Symmetric NMF with L1-norm distance

14 Views Asked by At

I am trying to validate a symmetric NMF model based on the L1-norm distance, that is

$$min_{U\geq0} \|X - UU^T\|_1$$

However, after implementing it, I got very strange results. The model works very well on synthetic data and fails with real world data ( mainly, face images data). I tried both dense similarity matrices and sparse ones, but nothing help! Does anyone have experienced similar issues?