Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Avoid generating NaNs in UnwoundPathSum (#3943)

Browse Source 
* Avoid generating NaNs in UnwoundPathSum.

Kudos to Jakub Zakrzewski for tracking down the bug.

* Add a test
This commit is contained in:
Kodi Arfer 2019-01-03 18:04:46 -05:00 committed by Philip Hyunsu Cho
parent 55bc149efb
commit 6a569b8cd9
2 changed files with 37 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
R-package/tests/testthat/test_interactions.R
View File

@ -81,6 +81,39 @@ test_that("predict feature interactions works", {
expect_lt(max(abs(intr - gt_intr)), 0.1)
})
test_that("SHAP contribution values are not NAN", {
d <- data.frame(
x1 = c(-2.3, 1.4, 5.9, 2, 2.5, 0.3, -3.6, -0.2, 0.5, -2.8, -4.6, 3.3, -1.2,
-1.1, -2.3, 0.4, -1.5, -0.2, -1, 3.7),
x2 = c(291.179171, 269.198331, 289.942097, 283.191669, 269.673332,
294.158346, 287.255835, 291.530838, 285.899586, 269.290833,
268.649586, 291.530841, 280.074593, 269.484168, 293.94042,
294.327506, 296.20709, 295.441669, 283.16792, 270.227085),
y = c(9, 15, 5.7, 9.2, 22.4, 5, 9, 3.2, 7.2, 13.1, 7.8, 16.9, 6.5, 22.1,
5.3, 10.4, 11.1, 13.9, 11, 20.5),
fold = c(2, 2, 2, 1, 2, 2, 1, 2, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2))
ivs <- c("x1", "x2")
fit <- xgboost(
verbose = 0,
params = list(
objective = "reg:linear",
eval_metric = "rmse"),
data = as.matrix(subset(d, fold == 2)[, ivs]),
label = subset(d, fold == 2)$y,
nthread = 1,
nrounds = 3)
shaps <- as.data.frame(predict(fit,
newdata = as.matrix(subset(d, fold == 1)[, ivs]),
predcontrib = T))
result <- cbind(shaps, sum = rowSums(shaps), pred = predict(fit,
newdata = as.matrix(subset(d, fold == 1)[, ivs])))
expect_true(identical(TRUE, all.equal(result$sum, result$pred, tol = 1e-6)))
})
test_that("multiclass feature interactions work", {
dm <- xgb.DMatrix(as.matrix(iris[,-5]), label=as.numeric(iris$Species)-1)

5
src/tree/tree_model.cc
View File

@ -293,9 +293,12 @@ bst_float UnwoundPathSum(const PathElement *unique_path, unsigned unique_depth,
total += tmp;
next_one_portion = unique_path[i].pweight - tmp * zero_fraction * ((unique_depth - i)
/ static_cast<bst_float>(unique_depth + 1));
} else {
} else if (zero_fraction != 0) {
total += (unique_path[i].pweight / zero_fraction) / ((unique_depth - i)
/ static_cast<bst_float>(unique_depth + 1));
} else {
CHECK_EQ(unique_path[i].pweight, 0)
<< "Unique path " << i << " must have zero weight";
}
}
return total;