A Cubist::cubist() model fitted with a factor predictor produces a formula that R cannot evaluate. The split condition keeps the quoted column name and gets no threshold:
set.seed(1)
n <- 120
df <- data.frame(
y = rnorm(n),
x = rnorm(n),
f = factor(sample(c("a", "b", "c", "d"), n, TRUE))
)
df$y <- df$y + as.numeric(df$f) + 0.5 * df$x
mf <- model.frame(y ~ x + f, df)
model <- Cubist::cubist(x = mf[, -1, drop = FALSE], y = mf[[1]])
tidypredict_fit(model)
#> (ifelse(`"f"` <= NA_real_, 1.384984 + x * 0.58, 0) + ifelse(`"f"` <=
#> NA_real_, 3.7233071 + x * 0.55, 0))/((`"f"` <= NA_real_) +
#> (`"f"` <= NA_real_))
rlang::eval_tidy(tidypredict_fit(model), df)
#> Error: object '"f"' not foundTwo things go wrong in the same place: the column name is carried through with its surrounding quotes, so it is backtick-quoted into a name that does not exist, and the categorical split value never becomes a threshold, so it is NA_real_. Cubist writes categorical rules as f in {a, b} style conditions rather than as a numeric cut, and the parser appears to read them with the numeric path.
Found by sweeping a 4-level unordered factor across every supported regression class rather than one model at a time. The other 15 classes in that sweep either matched their own predict() to ~1e-15 or failed in a way already tracked (#282 randomForest, #283 ranger); this one has no existing issue. #285 and #286 are about the extrapolation clamp and committee assignment and do not cover it.
Every categorical test in test-model-cubist.R goes through the x/y interface with numeric predictors only, which is why the suite is green.
A
Cubist::cubist()model fitted with a factor predictor produces a formula that R cannot evaluate. The split condition keeps the quoted column name and gets no threshold:Two things go wrong in the same place: the column name is carried through with its surrounding quotes, so it is backtick-quoted into a name that does not exist, and the categorical split value never becomes a threshold, so it is
NA_real_. Cubist writes categorical rules asf in {a, b}style conditions rather than as a numeric cut, and the parser appears to read them with the numeric path.Found by sweeping a 4-level unordered factor across every supported regression class rather than one model at a time. The other 15 classes in that sweep either matched their own
predict()to ~1e-15 or failed in a way already tracked (#282 randomForest, #283 ranger); this one has no existing issue. #285 and #286 are about the extrapolation clamp and committee assignment and do not cover it.Every categorical test in
test-model-cubist.Rgoes through the x/y interface with numeric predictors only, which is why the suite is green.