Minimal datasets to reproduce the line/point graphs in the manuscript
"Machine learning based approach to ... NMF analysis of spatially-resolved ARPES of graphene on SiC"
(STAM Methods, accepted)

Files
-----
fig2b.csv
  Reproduces Fig. 2(b): Summation of squared error between experimental ARPES
  spectra and NMF-reconstructed spectra, as a function of the number of NMF
  basis vectors (k = 2-24).
  Columns:
    Number of Base Vectors        : number of NMF components k
    Summation of Squared Error     : sum over all measurement positions of the
                                      mean-squared reconstruction error
                                      (normalized intensity scale, see Methods)

fig3d.csv
  Reproduces Fig. 3(d): contribution ratio of the activation vector H for the
  basis vectors W0 and W3 (NMF, k = 7) at the five representative positions
  A-E indicated in Fig. 3(a),(b) (x = 14.15-16.15 mm, z = 1.0 mm).
  Columns:
    Position               : map position label (A-E), corresponds to Fig. 3(a),(b)
    W0_activation_ratio     : H(W0) normalized to percent of total activation at that position
    W3_activation_ratio     : H(W3) normalized to percent of total activation at that position

fig4d.csv
  Reproduces Fig. 4(d): contribution ratio of the activation vector H for the
  basis vectors W1 and W4 (NMF, k = 7) at the five representative positions
  F-J indicated in Fig. 4(a),(b) (x = 15.15 mm, z = -6.6 to -2.6 mm).
  Columns:
    Position               : map position label (F-J), corresponds to Fig. 4(a),(b)
    W1_activation_ratio     : H(W1) normalized to percent of total activation at that position
    W4_activation_ratio     : H(W4) normalized to percent of total activation at that position

Methods (summary)
------------------
NMF (k = 7, init='random', max_iter=2000, random_state=1) was applied to the
PCA-reconstructed (n=390, i.e. full-rank), normalized spatially-resolved ARPES
dataset of graphene on 6H-SiC(0001) (390 spatial positions, each a 52 x 123
binding-energy x angle map after 3x3 rebinning). For Fig. 2(b), NMF was run
for k = 1-24 and the reconstruction error was evaluated on a 0-10000 normalized
intensity scale. For Figs. 3(d)/4(d), the activation matrix H (k=7) was
normalized at each spatial position to percentages summing to 100.
