在 xgb 中使用 f-score

Question

我正在尝试使用 scikit-learn 的 f-score 作为 xgb 分类器中的评估指标。这是我的代码：

clf = xgb.XGBClassifier(max_depth=8, learning_rate=0.004,
                            n_estimators=100,
                            silent=False,   objective='binary:logistic',
                            nthread=-1, gamma=0,
                            min_child_weight=1, max_delta_step=0, subsample=0.8,
                            colsample_bytree=0.6,
                            base_score=0.5,
                            seed=0, missing=None)
scores = []
predictions = []
for train, test, ans_train, y_test in zip(trains, tests, ans_trains, ans_tests):
        clf.fit(train, ans_train, eval_metric=xgb_f1,
                    eval_set=[(train, ans_train), (test, y_test)],
                    early_stopping_rounds=900)
        y_pred = clf.predict(test)
        predictions.append(y_pred)
        scores.append(f1_score(y_test, y_pred))

def xgb_f1(y, t):
    t = t.get_label()
    return "f1", f1_score(t, y)

但是出现错误：Can't handle mix of binary and continuous

Answer 1

问题是 f1_score 正在尝试比较 非二进制与二进制 目标，默认情况下此方法进行二进制平均。来自 documentation "average : string, [None, 'binary' (default), 'micro', 'macro', 'samples', 'weighted' ]”。

无论如何，错误是说你的预测是这样连续的 [0.001, 0.7889,0.33...] 但你的目标是二元的 [0,1,0...]。因此，如果您知道阈值，我建议您在将结果发送到 f1_score 函数之前对其进行预处理。阈值的通常值为 0.5.

评估函数的测试示例。不再输出错误：

def xgb_f1(y, t, threshold=0.5):
    t = t.get_label()
    y_bin = [1. if y_cont > threshold else 0. for y_cont in y] # binarizing your output
    return 'f1',f1_score(t,y_bin)

根据@smci 的建议，less_verbose/more_efficient 解决方案可能是：

def xgb_f1(y, t, threshold=0.5):
    t = t.get_label()
    y_bin = (y > threshold).astype(int) # works for both type(y) == <class 'numpy.ndarray'> and type(y) == <class 'pandas.core.series.Series'>
    return 'f1',f1_score(t,y_bin)

在 xgb 中使用 f-score

Using f-score in xgb

python

evaluation-function

xgboost