Non-Negative Matrix Factorization with Kernel Covariates (original) (raw)

Lifecycle: experimental GitHub version

nmfkc is an R package that extends Non-negative Matrix Factorization (NMF) by incorporating covariates using kernel methods. It supports advanced features like rank selection via cross-validation, time-series modeling (NMF-VAR), supervised classification (NMF-LAB), feed-forward + feedback structural modeling with equilibrium interpretation (NMF-FFB; formerly NMF-SEM), and mixed-effects modeling with random effects (NMF-RE).

Installation

Help and Usage

Citation

Quick Example

library(nmfkc)

# Decompose a matrix Y into basis X and coefficient B with rank = 2
X_true <- cbind(c(1, 0, 1), c(0, 1, 0))
B_true <- cbind(c(1, 0), c(0, 1), c(1, 1))
Y <- X_true %*% B_true

res <- nmfkc(Y, rank = 2, epsilon = 1e-6)
plot(res)     # Convergence plot
summary(res)  # Summary statistics

See browseVignettes("nmfkc") for detailed examples covering rank selection, kernel NMF, time-series, classification, NMF-FFB, and NMF-RE.

Comparison with Standard NMF

Feature Standard NMF nmfkc
Handles covariates No Yes (Linear / Kernel)
Feed-forward + feedback modeling No Yes (NMF-FFB)
Mixed-effects / Random effects No Yes (NMF-RE)
Classification No Yes (NMF-LAB)
Time series modeling No Yes (NMF-VAR)
Nonlinearity No Yes (Kernel)
Clustering support Limited Yes (Hard/Soft)
Rank selection / CV Limited (ad hoc) Yes (Element-wise CV, Column-wise CV)

Statistical Model

The nmfkc package builds upon the standard NMF framework by incorporating external information (covariates):

\[Y(P,N) \approx X(P,Q) \times C(Q,R) \times A(R,N)\]

Extensions

References