problem with maxInteractDeg

[debis93]

Hi,
I found the problem that, when the second tuning parameter changes (maxInteractDeg), the model estimated with the same "deg" remains unchanged.
In fact, from this example the design matrix remains unchanged.

x <- iris[,1:4]
dim(getPoly(x, deg = 1, maxInteractDeg = 1)$xdata)
dim(getPoly(x, deg = 1, maxInteractDeg = 2)$xdata)
dim(getPoly(x, deg = 1, maxInteractDeg = 3)$xdata)
dim(getPoly(x, deg = 2, maxInteractDeg = 1)$xdata)
dim(getPoly(x, deg = 2, maxInteractDeg = 2)$xdata)
dim(getPoly(x, deg = 2, maxInteractDeg = 3)$xdata)

What could this problem depend on?
Currently I do not find it useful to vary the model according to the degree of interaction.

[matloff]

This may be a bug, introduced during a recent update. I'll check.

[rdrr1990]

I just pushed get_poly(), which addresses this issue. I have not replaced getPoly() just yet because I want to be sure that it doesn't introduce any little issues downstream. In the meantime, the following examples should make it clear how to use it with most estimators. Note that if you use it for CV, please take care to generate the test model matrix as shown below.

X <- get_poly(mtcars, 2)
W <- get_poly(Xy=mtcars, 2) # treats final column as response, as other parts of polyreg do
ncol(W) < ncol(X)                  # TRUE
X_train <- get_poly(mtcars[1:20,], 4, 2)
X_test <- get_poly(mtcars[21:32,],
                    model_formula = attributes(X_train)$formula)

[luca-scr]

I had the same issue so I tried to use get_poly() but it seems there are some problems:

1. there are two files get_poly.R, one in the root folder and one in the R folder, which is the one to use?
2. the function get_poly() is not exported so it can't be used. Strangely enough, it is not available even if I call it using polyreg:::get_poly()
3. I tried to source the get_poly.Rbut it doesn't work because some called functions are not exported, for instance, N_distinct().

Update: using a hack I have been able to run get_poly() by adding the following lines at the beginning of the file:

pow <- polyreg:::pow
model_matrix <- polyreg:::model_matrix
is_continuous <- polyreg:::is_continuous
N_distinct <- polyreg:::N_distinct
get_interactions <- polyreg:::get_interactions

First, having names describing the terms is a great and useful addition! Thanks.

However, I noticed two issues in usage of get_poly():

1. the object returned is quite different from that returned by getPoly(), so I guess it wouldn't be possible to use get_poly() as is in polyFit().

2. consider the following example:

x <- iris[,1:4]
xp <- get_poly(x, deg = 1, maxInteractDeg = 1)
dim(xp)
# [1] 150  12
colnames(xp)
#  [1] "Sepal.Length"         "Sepal.Width"          "Petal.Length"        
#  [4] "Petal.Width"          "pow(Sepal.Length, 2)" "pow(Sepal.Width, 2)" 
#  [7] "pow(Petal.Length, 2)" "pow(Petal.Width, 2)"  "pow(Sepal.Length, 1)"
# [10] "pow(Sepal.Width, 1)"  "pow(Petal.Length, 1)" "pow(Petal.Width, 1)" 

Problems:

• the terms pow(*,1) shouldn't be included (they are already at the beginning)
• why terms pow(*,2) if deg = 1?
  I think these problems could be easily solved by changing the code for(i in 2:maxDeg) to for(i in seq(maxDeg)). This a common problem in R usage. Consider

maxDeg = 1
2:maxDeg
# [1] 2 1         # <- this is wrong
seq(maxDeg)
# [1] 1           # <- now it is ok

And perhaps the extra linear terms could be avoided using for(i in seq(maxDeg)[-1]) because

seq(maxDeg)[-1]
# integer(0)

Sorry for this long message but I hope it helps to clear and fix the issue.

[matloff]

The get_poly() function is under construction, hence not listed in NAMESPACE. It will eventually replace getPoly(). Thanks for your support and your patience.

[rdrr1990]

@luca-scr thanks for the notes--they are very helpful as we build the replacement!