with non negative

R/NMF_optim.R

@@ -7,6 +7,7 @@
 #' @param k Number of components.
 #' @param W Initial matrix W (optional, defaults to random non-negative values).
 #' @param H Initial matrix H (optional, defaults to random non-negative values).
+#' @param notNegative Logical. If TRUE, the factorization is constrained to non-negative elements.
 #' @param max_iter Maximum number of iterations (default is 100).
 #' @return A list containing factorized matrices W and H.
 #'
@@ -22,7 +23,7 @@
 #'
 #' @export
 
-nmf_conjugate_gradient <- function(V, k, W = NULL, H = NULL, max_iter = 100) {
+nmf_conjugate_gradient <- function(V, k, W = NULL, H = NULL, notNegative = FALSE, max_iter = 100) {
 
   # Initialize W and H with random non-negative values
   if(is.null(W) || is.null(H)){
@@ -63,13 +64,14 @@ nmf_conjugate_gradient <- function(V, k, W = NULL, H = NULL, max_iter = 100) {
   # Perform projected conjugate gradient optimization
   result <- optim(par = theta, fn = obj_fun, gr = grad_obj_fun, method = "L-BFGS-B", control = list(maxit = max_iter))
 
-  # Extract factorized matrices W and H and enforce non-negativity
-  # W <- project_to_non_negative(matrix(result$par[1:(nrow(V) * k)], nrow = nrow(V), ncol = k))
-  # H <- project_to_non_negative(matrix(result$par[(nrow(V) * k + 1):length(result$par)], nrow = k, ncol = ncol(V)))
-
   W <- matrix(result$par[1:(nrow(V) * k)], nrow = nrow(V), ncol = k)
   H <- matrix(result$par[(nrow(V) * k + 1):length(result$par)], nrow = k, ncol = ncol(V))
 
+  # Extract factorized matrices W and H and enforce non-negativity
+  if(notNegative){
+    W <- project_to_non_negative(W)
+    H <- project_to_non_negative(H)
+  }
   n <- length(V)
   squared_diff <- (V - W%*%H)^2
   mse <- sum(squared_diff) / n