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

Re-implement PR-AUC. (#7297)

Browse Source 
* Support binary/multi-class classification, ranking.
* Add documents.
* Handle missing data.
This commit is contained in:
Jiaming Yuan
2021-10-26 13:07:50 +08:00
committed by GitHub
parent a6bcd54b47
commit d4349426d8
12 changed files with 1035 additions and 655 deletions
Download Patch File Download Diff File Expand all files Collapse all files

60
tests/cpp/metric/test_rank_metric.cc
View File

@@ -24,66 +24,6 @@ TEST(Metric, AMS) {
}
#endif
TEST(Metric, DeclareUnifiedTest(AUCPR)) {
auto tparam = xgboost::CreateEmptyGenericParam(GPUIDX);
xgboost::Metric *metric = xgboost::Metric::Create("aucpr", &tparam);
ASSERT_STREQ(metric->Name(), "aucpr");
EXPECT_NEAR(GetMetricEval(metric, {0, 0, 1, 1}, {0, 0, 1, 1}), 1, 1e-10);
EXPECT_NEAR(GetMetricEval(metric, {0.1f, 0.9f, 0.1f, 0.9f}, {0, 0, 1, 1}),
0.5f, 0.001f);
EXPECT_NEAR(
GetMetricEval(metric,
{0.4f, 0.2f, 0.9f, 0.1f, 0.2f, 0.4f, 0.1f, 0.1f, 0.2f, 0.1f},
{0, 0, 0, 0, 0, 1, 0, 0, 1, 1}),
0.2908445f, 0.001f);
EXPECT_NEAR(GetMetricEval(
metric, {0.87f, 0.31f, 0.40f, 0.42f, 0.25f, 0.66f, 0.95f,
0.09f, 0.10f, 0.97f, 0.76f, 0.69f, 0.15f, 0.20f,
0.30f, 0.14f, 0.07f, 0.58f, 0.61f, 0.08f},
{0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 1, 1}),
0.2769199f, 0.001f);
EXPECT_ANY_THROW(GetMetricEval(metric, {0, 1}, {}));
EXPECT_ANY_THROW(GetMetricEval(metric, {0, 0}, {0, 0}));
EXPECT_ANY_THROW(GetMetricEval(metric, {0, 0}, {1, 1}));
// AUCPR with instance weights
EXPECT_NEAR(GetMetricEval(
metric, {0.29f, 0.52f, 0.11f, 0.21f, 0.219f, 0.93f, 0.493f,
0.17f, 0.47f, 0.13f, 0.43f, 0.59f, 0.87f, 0.007f},
{0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 1, 0},
{1, 2, 7, 4, 5, 2.2f, 3.2f, 5, 6, 1, 2, 1.1f, 3.2f, 4.5f}), // weights
0.694435f, 0.001f);
// AUCPR with groups and no weights
EXPECT_NEAR(GetMetricEval(
metric, {0.87f, 0.31f, 0.40f, 0.42f, 0.25f, 0.66f, 0.95f,
0.09f, 0.10f, 0.97f, 0.76f, 0.69f, 0.15f, 0.20f,
0.30f, 0.14f, 0.07f, 0.58f, 0.61f, 0.08f},
{0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 0, 1, 1},
{}, // weights
{0, 2, 5, 9, 14, 20}), // group info
0.556021f, 0.001f);
// AUCPR with groups and weights
EXPECT_NEAR(GetMetricEval(
metric, {0.29f, 0.52f, 0.11f, 0.21f, 0.219f, 0.93f, 0.493f,
0.17f, 0.47f, 0.13f, 0.43f, 0.59f, 0.87f, 0.007f}, // predictions
{0, 1, 1, 0, 0, 1, 1, 0, 0, 0, 0, 0, 1, 0},
{1, 2, 7, 4, 5, 2.2f, 3.2f, 5, 6, 1, 2, 1.1f, 3.2f, 4.5f}, // weights
{0, 2, 5, 9, 14}), // group info
0.8150615f, 0.001f);
// Exception scenarios for grouped datasets
EXPECT_ANY_THROW(GetMetricEval(metric,
{0, 0.1f, 0.3f, 0.5f, 0.7f},
{1, 1, 0, 0, 0},
{},
{0, 2, 5}));
delete metric;
}
TEST(Metric, DeclareUnifiedTest(Precision)) {
// When the limit for precision is not given, it takes the limit at
// std::numeric_limits<unsigned>::max(); hence all values are very small