DESCRIPTION

Students:  Please note instructions in paper copy of quiz.  Save your
files often, make sure OMSI fills your entire screen at all times, etc.

A question may not fully fit into your OMSI question box, which is not
scrollable.  You can try adjusting the relative size of the question and
answer boxes, but remember that clicking CopyQtoA will copy the entire
question box to the answer box.

QUESTION

(30 pts) Which can accommodate prediction of new users' ratings of old
movies?  (choose as many as apply.)  "Old" and "new" refer to our
original matrix A.  No covariates used.

(i) lm(), and the new user only has rated old movies
(ii) lm(), and the new user only has rated new movies
(iii) lm(), and the new user has rated both old and new movies
(iv) NMF, and the new user only has rated old movies
(v) NMF, and the new user only has rated new movies
(vi) NMF, and the new user has rated both old and new movies

QUESTION -ext .R -run 'Rscript omsi_answer2.R'

(35 pts) The code below will run NMF on the InstEval data using
recosystem, and then will complete the matrix and evaluate the accuracy
(on the same data, no cross-validation).  The accuracy measure is
proportion of correctly predicted ratings, not MAPE.  Use rank 20.

gy <- function() {

   library(lme4)
   data(InstEval)
   ie <- InstEval
   
   # easier if IDs are 1,2,3,..., and ratings are numeric
   factorToNumToFactor <- function(f) as.factor(as.numeric(f))
   ie$s <- factorToNumToFactor(ie$s)
   ie$d <- factorToNumToFactor(ie$d)
   ie$y <- as.numeric(ie$y)
   
   library(recosystem)
   r <- Reco()
   # training set will be all of ie, but need in recosystem form

   
   r$train(ie.trn,       )
   result <- r$output(out_memory(),out_memory())
  
   # for a given row in ie, find the predicted rating
   findPred <- function(row) {
      row <- as.numeric(row)  # factors treated as character mode, need nums
      userID <- row[1]
      instID <- row[2]
      round(       )
   }

   # get vector of predicted ratings for ie
   preds <- apply(ie,1,findPred)

   mean(       )

}

print(gy())

QUESTION -ext .R -run 'Rscript omsi_answer3.R'

(35 pts) The nmf() function below will return W and H for the
approximate factorization WH, using the ALS algorithm.  It is assumed
that the A matrix in the first argument is fully known.  The second and
third arguments are the initial guesses, and nIters is the number of
iterations; A find-H, find-W cycle counts as one iteration.  Note:  Do
not constrain W and H to be nonnegative; this question is really about
ALS, not NMF.

nmf <- function(a,w0,h0,nIters) {
   w <- w0; h <- h0
   nch <- ncol(h)
   nrw <- nrow(w)
   for (iter in 1:nIters) {

 

   
   }
   list(w=w, h=h)
}

a <- rbind(2:5,c(4,3,5,1),c(2,2,1,4),1:4)
w0 <- matrix(1:8,ncol=2) 
h0 <- matrix(1:8,nrow=2) 
print(nmf(a,w0,h0,2))