X ~ W H with W, H >= 0 by Lee-Seung multiplicative updates: Frobenius loss,
or loss = "kl" for the Poisson log-likelihood of ESL eq 14.73 with the
updates of eq 14.74 (W then H). The default start is the same deterministic
LCG as the Python arm.
Usage
morie_esl_nmf(
X,
k,
max_iter = 500,
tol = 1e-10,
seed = 13,
loss = c("frobenius", "kl"),
W0 = NULL,
H0 = NULL
)
Arguments
- X
Non-negative matrix.
- k
Rank.
- max_iter, tol
Update controls.
- seed
LCG seed for the default start.
- loss
"frobenius" or "kl".
- W0, H0
Optional starting factors.
Value
Named list: estimate, W, H, frobenius_error, relative_error,
iterations, converged, n, p, k, loss, and for "kl" kl_divergence, loglik,
divergence_path.
References
Lee, D. D. & Seung, H. S. (2001). NIPS 13, 556-562.
Examples
X <- outer(1:6, 1:5, function(i, j) abs(sin(i * j / 3)) * 3 + (i + j) %% 3)
morie_esl_nmf(X, 2, loss = "kl")$kl_divergence
#> [1] 4.643192
