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

Re-implement ROC-AUC. (#6747)

Browse Source 
* Re-implement ROC-AUC.

* Binary
* MultiClass
* LTR
* Add documents.

This PR resolves a few issues:
  - Define a value when the dataset is invalid, which can happen if there's an
  empty dataset, or when the dataset contains only positive or negative values.
  - Define ROC-AUC for multi-class classification.
  - Define weighted average value for distributed setting.
  - A correct implementation for learning to rank task.  Previous
  implementation is just binary classification with averaging across groups,
  which doesn't measure ordered learning to rank.
This commit is contained in:
Jiaming Yuan
2021-03-20 16:52:40 +08:00
committed by GitHub
parent 4ee8340e79
commit bcc0277338
27 changed files with 1622 additions and 461 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
src/tree/param.h
View File

@@ -15,6 +15,7 @@
#include "xgboost/parameter.h"
#include "xgboost/data.h"
#include "../common/math.h"
namespace xgboost {
namespace tree {
@@ -264,14 +265,11 @@ XGBOOST_DEVICE inline static T1 ThresholdL1(T1 w, T2 alpha) {
return 0.0;
}
template <typename T>
XGBOOST_DEVICE inline static T Sqr(T a) { return a * a; }
// calculate the cost of loss function
template <typename TrainingParams, typename T>
XGBOOST_DEVICE inline T CalcGainGivenWeight(const TrainingParams &p,
T sum_grad, T sum_hess, T w) {
return -(T(2.0) * sum_grad * w + (sum_hess + p.reg_lambda) * Sqr(w));
return -(T(2.0) * sum_grad * w + (sum_hess + p.reg_lambda) * common::Sqr(w));
}
// calculate weight given the statistics
@@ -296,9 +294,9 @@ XGBOOST_DEVICE inline T CalcGain(const TrainingParams &p, T sum_grad, T sum_hess
}
if (p.max_delta_step == 0.0f) {
if (p.reg_alpha == 0.0f) {
return Sqr(sum_grad) / (sum_hess + p.reg_lambda);
return common::Sqr(sum_grad) / (sum_hess + p.reg_lambda);
} else {
return Sqr(ThresholdL1(sum_grad, p.reg_alpha)) /
return common::Sqr(ThresholdL1(sum_grad, p.reg_alpha)) /
(sum_hess + p.reg_lambda);
}
} else {

2
src/tree/split_evaluator.h
View File

@@ -114,7 +114,7 @@ class TreeEvaluator {
}
// Avoiding tree::CalcGainGivenWeight can significantly reduce avg floating point error.
if (p.max_delta_step == 0.0f && has_constraint == false) {
return Sqr(ThresholdL1(stats.sum_grad, p.reg_alpha)) /
return common::Sqr(ThresholdL1(stats.sum_grad, p.reg_alpha)) /
(stats.sum_hess + p.reg_lambda);
}
return tree::CalcGainGivenWeight<ParamT, float>(p, stats.sum_grad,