(self)
| 10 | |
| 11 | class TestSHAP: |
| 12 | def test_feature_importances(self) -> None: |
| 13 | rng = np.random.RandomState(1994) |
| 14 | data = rng.randn(100, 5) |
| 15 | target = np.array([0, 1] * 50) |
| 16 | |
| 17 | features = ["Feature1", "Feature2", "Feature3", "Feature4", "Feature5"] |
| 18 | |
| 19 | dm = xgb.DMatrix(data, label=target, feature_names=features) |
| 20 | params = { |
| 21 | "objective": "multi:softprob", |
| 22 | "eval_metric": "mlogloss", |
| 23 | "eta": 0.3, |
| 24 | "num_class": 3, |
| 25 | } |
| 26 | |
| 27 | bst = xgb.train(params, dm, num_boost_round=10) |
| 28 | |
| 29 | # number of feature importances should == number of features |
| 30 | scores1 = bst.get_score() |
| 31 | scores2 = bst.get_score(importance_type="weight") |
| 32 | scores3 = bst.get_score(importance_type="cover") |
| 33 | scores4 = bst.get_score(importance_type="gain") |
| 34 | scores5 = bst.get_score(importance_type="total_cover") |
| 35 | scores6 = bst.get_score(importance_type="total_gain") |
| 36 | assert len(scores1) == len(features) |
| 37 | assert len(scores2) == len(features) |
| 38 | assert len(scores3) == len(features) |
| 39 | assert len(scores4) == len(features) |
| 40 | assert len(scores5) == len(features) |
| 41 | assert len(scores6) == len(features) |
| 42 | |
| 43 | # check backwards compatibility of get_fscore |
| 44 | fscores = bst.get_fscore() |
| 45 | assert scores1 == fscores |
| 46 | |
| 47 | dtrain, dtest = tm.load_agaricus(__file__) |
| 48 | |
| 49 | def fn(max_depth: int, num_rounds: int) -> None: |
| 50 | # train |
| 51 | params = {"max_depth": max_depth, "eta": 1} |
| 52 | bst = xgb.train(params, dtrain, num_boost_round=num_rounds) |
| 53 | |
| 54 | # predict |
| 55 | preds = bst.predict(dtest) |
| 56 | contribs = bst.predict(dtest, pred_contribs=True) |
| 57 | |
| 58 | # result should be (number of features + BIAS) * number of rows |
| 59 | assert contribs.shape == (dtest.num_row(), dtest.num_col() + 1) |
| 60 | |
| 61 | # sum of contributions should be same as predictions |
| 62 | np.testing.assert_array_almost_equal(np.sum(contribs, axis=1), preds) |
| 63 | |
| 64 | # for max_depth, num_rounds in itertools.product(range(0, 3), range(1, 5)): |
| 65 | # yield fn, max_depth, num_rounds |
| 66 | |
| 67 | # check that we get the right SHAP values for a basic AND example |
| 68 | # (https://arxiv.org/abs/1706.06060) |
| 69 | X = np.zeros((4, 2)) |
nothing calls this directly
no test coverage detected