Plotting with linear = TRUE sometimes fails about rforestry

[sjsw]

It looks like this was actually caused by an outdated version of glmnet and not by forestry itself.

[sjsw]

Actually it seems it is not related to glment version, rather the error is intermittent because it depends on the seed used to construct the forest. I think it probably arises when a tree contains leaves that are grown to purity, which causes glmnet to fail when called in the plotting method. This could be fixed by patching the plotting function to not call glmnet when the variance of the outcome in a leaf is zero.

[linanqiu]

For the case where you have this error:

Error in if (any(lambda < 0)) stop("lambdas should be non-negative") : 
  missing value where TRUE/FALSE needed

The cause is:

Rforestry/R/plottree.R

Lines 238 to 252 in cecf729

	if (forestry_tree@linear) {
	# ridge forest
	for (leaf_id in node_info$node_id[node_info$is_leaf]) {
	# leaf_id = 5
	###
	this_ds <- dta_x[leaf_idx[[leaf_id]],
	forestry_tree@linFeats + 1]
	encoder <- onehot::onehot(this_ds)
	remat <- predict(encoder, this_ds)
	###
	y_leaf <- dta_y[leaf_idx[[leaf_id]]]
	plm <- glmnet::glmnet(x = remat,
	y = y_leaf,
	lambda = forestry_tree@overfitPenalty * sd(y_leaf)/nrow(remat),
	alpha = 0)

As you said, y_leaf is a single element vector. sd(y_leaf) = NA, causing lambda = NA.

Browse[1]> n
debug at C:/Users/Linan Qiu/Desktop/temp.R#227: plm <- glmnet::glmnet(x = remat, y = y_leaf, lambda = forestry_tree@overfitPenalty * 
    sd(y_leaf)/nrow(remat), alpha = 0)
Browse[2]> forestry_tree@overfitPenalty * sd(y_leaf)/nrow(remat)
[1] NA
Browse[2]> sd(y_leaf)
[1] NA
Browse[2]> y_leaf
[1] -19.65548

[linanqiu]

For the case where you have the other error:

 Error in elnet(xd, is.sparse, ix, jx, y, weights, offset, type.gaussian,  : 
  y is constant; gaussian glmnet fails at standardization step 

The cause is still that couple of lines, but that's when y_leaf has length > 1 but consist of all the same elements:

# added these print statements to the if block above:

#      message(glue('overfitPenalty: {forestry_tree@overfitPenalty}'))
#      message(glue('y_leaf: {paste0(y_leaf, collapse = ", ")}, sd(y_leaf): {sd(y_leaf)}'))
#      message(glue('nrow(remat): {nrow(remat)}'))

overfitPenalty: 1
y_leaf: -9.56037240704544, -9.56037240704544, -9.56037240704544, sd(y_leaf): 0
nrow(remat): 3

Y is constant, which is what glmnet is complaining about. Could be because glmnet does some standardization (https://stackoverflow.com/questions/54755857/glmnet-error-nulldev-0-stopy-is-constant-gaussian-glmnet-fails-at-standa) and fails at sd(y_leaf) = 0

These two cases seem to exhaust the errors I've seen by clicking source a dozen times 😛

Seems like the right thing to do for a linear aggregator at leaf in these cases is just to return Y?

[sjsw]

This LGTM, Linan! :)