Avoid generating NaNs in UnwoundPathSum

[Kodiologist]

This is my attempt to fix the bug creating NaN SHAP values (#3333). I have an example R script (below) which produces NaN SHAPs without this patch and doesn't produce any with it, but I don't know whether my change is statistically correct.

The script is:

library(xgboost)

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")

set.seed(50)

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])))
print(result)

message(if (identical(TRUE, all.equal(result$sum, result$pred, tol = 1e-6)))
       "Looks good"
   else
       "Oh no")

Its output without the change is:

          x1  x2     BIAS sum      pred
1  0.2951230 NaN 6.967499 NaN  5.478699
2 -0.6753536 NaN 6.967499 NaN  5.478699
3  0.2951230 NaN 6.967499 NaN  5.478699
4  0.8563597 NaN 6.967499 NaN 11.253153
Oh no

Its output with the change is:

          x1         x2     BIAS       sum      pred
1  0.2951230 -1.7839233 6.967499  5.478699  5.478699
2 -0.6753536 -0.8134466 6.967499  5.478699  5.478699
3  0.2951230 -1.7839233 6.967499  5.478699  5.478699
4  0.8563597  3.4292932 6.967499 11.253152 11.253153
Looks good

CC @slundberg, @allanjust, @liuyanguu

Fixes #3333.

[slundberg]

Thanks @Kodiologist! I think your change should retain the correctness of the method. If this resolves the NaNs then the issue must have been when the proportion of training samples and/or depth weighting caused an underflow. In that case they would fail to cancel out. Unless anyone else has comments I think this should be merged.

[Kodiologist]

I remembered that the SHAPs for an observation always sum to equal the prediction for that observation, so we can use that as a check. Turns out, it works! Hooray! I've incorporated it into the test script above, which should probably be put in XGBoost's test suite in some form.

By the way, I think the Makefile for this project may not be tracking all dependencies correctly, because I found that I had to do make clean before each make in order for the R package to see changes to the C++ code outside its own interface code.

[hcho3]

@Kodiologist Thanks a lot for your contribution. Before merging this in, let me add your sample snippet as a unit test so that the bug doesn't creep in in the future. As for dependency management, yes, I think there is some issue with Makefile. I usually use CMake instead:

mkdir build
cd build
cmake .. -DR_LIB=ON
make -j10
make install

Then every change in C++ code will cause make install to install fresh binary into R package directory.

[Kodiologist]

Oh, okay. You might want to add the CMake call to https://xgboost.readthedocs.io/en/latest/build.html#building-on-ubuntu-debian .

[slundberg]

@hcho3 did you ever end up adding that sample snippet?

[slundberg]

Also looks like the relevant code got moved to tree_model.cc since the initial PR

[hcho3]

@slundberg Sorry, I've been away for vacation. I'll rebase this PR against the latest master now.

[hcho3]

@Kodiologist @slundberg Okay, looks like the re-basing was successful. I'll add the test snippet soon.

[codecov-io]

Codecov Report

Merging #3943 into master will decrease coverage by 0.01%.
The diff coverage is 0%.

@@            Coverage Diff             @@
##           master    #3943      +/-   ##
==========================================
- Coverage   60.79%   60.77%   -0.02%     
==========================================
  Files         130      130              
  Lines       11699    11702       +3     
==========================================
  Hits         7112     7112              
- Misses       4587     4590       +3

Impacted Files	Coverage Δ
src/tree/tree_model.cc	17.91% <0%> (-0.23%)	⬇️
tests/cpp/linear/test_linear.cc	100% <0%> (ø)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f368d0d...1928ffc. Read the comment docs.

[hcho3]

@Kodiologist @slundberg I added the test script. It should be good to merge once all checks pass.

[hcho3]

@Kodiologist Merged. Thanks for your contribution.

[astrixg]

I think this issue still persists. I upgraded to the latest pip library v0.27.0, but still the shap.TreeExplainer.shap_values() function is generating NaN values for some features.

[hcho3]

@astrixg The latest version is 0.81. Can you upgrade? If the issue persists, you should create a new issue.

[astrixg]

https://pypi.org/project/shap/
https://anaconda.org/conda-forge/shap

@hcho3 Its 0.28 on pip and 0.27 on conda. Am I installing from the wrong source?

[Kodiologist]

@astrixg hcho3 is talking about the xgboost package, not shap. However, upgrading to xgboost 0.81 won't help you because that release predates this PR. To get this fix now, you need to compile and install xgboost master.