Figure 2

Properties of NMF on epigenomic data. (A) Average Sparsity of factor matrices W and H as a function of c calculated using Hoyer’s formula (see Results). (B) Approximate nesting of "basis patterns" for factorizations using different factorization ranks c. Nesting occurs if a lower rank basis vector is approximately derived by adding two higher rank basis vectors. (C) NMF reconstruction error and sparsity on 1000 random initializations of matrices H and W. NNDSVD – deterministic non-negative double singular value decomposition, random – uniformly (0 to 1 range) random non-negative matrices. The 1000 NNDSVDar randomizations find approximately the same factorization as NNDSVD and are not shown.