Add SHAP calculation to GBT regression #2460
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
cpp/daal/include/algorithms/decision_forest/decision_forest_classification_model_builder.h
Outdated
Show resolved
Hide resolved
cpp/daal/src/algorithms/dtrees/gbt/regression/gbt_regression_predict_dense_default_batch_impl.i
Outdated
Show resolved
Hide resolved
cpp/daal/src/algorithms/dtrees/gbt/regression/gbt_regression_predict_dense_default_batch_impl.i
Outdated
Show resolved
Hide resolved
cpp/daal/src/algorithms/dtrees/gbt/regression/gbt_regression_predict_dense_default_batch_impl.i
Outdated
Show resolved
Hide resolved
cpp/daal/src/algorithms/dtrees/gbt/regression/gbt_regression_predict_dense_default_batch_impl.i
Outdated
Show resolved
Hide resolved
cpp/daal/src/algorithms/dtrees/gbt/regression/gbt_regression_predict_dense_default_batch_impl.i
Outdated
Show resolved
Hide resolved
cpp/daal/src/algorithms/dtrees/gbt/regression/gbt_regression_predict_dense_default_batch_impl.i
Outdated
Show resolved
Hide resolved
cpp/daal/src/algorithms/dtrees/gbt/regression/gbt_regression_predict_dense_default_batch_impl.i
Outdated
Show resolved
Hide resolved
cpp/daal/src/algorithms/dtrees/gbt/regression/gbt_regression_predict_dense_default_batch_impl.i
Outdated
Show resolved
Hide resolved
ahuber21
force-pushed
the
dev/ahuber/shap-values-DecisionTreeTable
branch
from
e2dc13e
to
cc21b23
Compare
ahuber21
force-pushed
the
dev/ahuber/shap-values-DecisionTreeTable
branch
from
88b770b
to
00d8087
Compare
ahuber21
requested review from
Alexsandruss and
samir-nasibli
as code owners
|
/intelci: run |
|
@Alexsandruss please use this CI http://intel-ci.intel.com/ee5dea85-3976-f168-8976-a4bf010d0e2e |
|
@ahuber21 - please rebase/merge from master - currently pr shows changes that already in master as a diff , hard to review |
add weights to GbtDecisionTree Include TreeShap recursion steps fix buffer overflow in memcpy Add cover to GbtDecisionTree from model builder fix some index offsets, correct results for trees up to depth=5 fix: nodeIsDummyLeaf is supposed to check left child remove some debug statements chore: apply oneDAL code style predictContribution wrapper with template dispatching increase speed by reducing number of cache misses use thread-local result accessor backup commit with 13% speedup wrt xgboost add preShapContributions/predShapInteractions as function parameter Revert "introduce pred_contribs and pred_interactions SHAP options" This reverts commit 483aa5b. remove some debug content reset env_detect.cpp to origin/master remove std::vector<float> test by introducing thread-local NumericTable Move treeshap into separate translation unit - caution: treeShap undefined in libonedal builds but segfaults Fix function arguments respect predShapContributions and predShapInteractions options and check for legal combinations tmp: work on pred_interactions
… function arguments
|
Pushed updates after applying review comments. Only the usage of |
Vika-F
reviewed
Oct 20, 2023
cpp/daal/include/algorithms/gradient_boosted_trees/gbt_regression_predict_types.h
Outdated
Show resolved
Hide resolved
Vika-F
reviewed
Oct 20, 2023
Co-authored-by: Victoriya Fedotova <viktoria.nn@gmail.com>
|
Evaluation of CI failures: LinuxDaal4py is expected to fail, we depend on changes from intel/scikit-learn-intelex#1399 LinuxMakeDPCPP: Happens in other pipelines as well. Needs to be investigated separately. |
Vika-F
approved these changes
Oct 27, 2023
Vika-F
added a commit
to Vika-F/daal
that referenced
this pull request
Oct 27, 2023
Add SHAP calculation to GBT regression (oneapi-src#2460)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR explores SHAP values (originally added in XGBoost here: dmlc/xgboost#2438). I have taken the same algorithm and modified it to work with oneDAL datastructures. Most performance bottlenecks were removed. However, there is still room for improvement. (For instance a more efficient use of vector instructions in the
unwoundPathSum()andunwindPath()helper functions.I am creating this as a draft because I am aware of a few things that must be changed. Feel free to review regardless. All comments are helpful at this point.
Things to be addressed (incomplete list, likely to be extended)
coverwas added for regression, since it is needed to calculate SHAP values. I have added a dummycover = 0in some places to fix compilation errors, but that obviously needs improvement. MovingtreeShap()to a common base class and enabling it for both classification and regression is probably the best solution.predShapContributionsandpredShapInteractionsmodel parameters are available. The code inPredictRegressionTask::run()must be updated accordingly. Additionally, the fitting size of result memory must be allocated in gbt_regression_predict_result_fpt.cpp (also hardcoded to accommodate SHAP contributions).