xgboost converter updates#106
Open
thesps wants to merge 4 commits into
Open
Conversation
… percentage of correctly reached leaves. Then require a numerical agreement only on correctly reached leaves.
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The XGBoost converter hasn't worked that well since XGBoost version 2 (now at 3.3.0). This PR implements some missed feature - the model initial prediction / base score, and updates the tests.
Before this update, using the new test script, all of the tests failed based on numerical mismatches. Now, 3 of the tests pass and 3 of the tests fail due to leaf mismatches. I observed that even a single tree model would show numerical mismatches - hinting that reaching a different leaf was an issue as well as the different numerical values from the missing base score. The issue of reaching a wrong leaf is not fixed by this PR, but the testing is now able to help isolate the two types of issues.
Reaching a wrong leaf will be addressed in a future update, and seems to occur specifically for values which are very close to a decision threshold.