LightGBMカスタムメトリクスのサンプル

毎回リファレンス確認しながら書いてるカスタムメトリクス。
ローカルにソース転がってたので、備忘録として挙げておく。

＃ サンプルコード

from sklearn.metrics import roc_auc_score
from sklearn.metrics import average_precision_score

def calc_accuracy(y_pred, y_true):
    metric = np.mean(y_true == y_pred)
    return 'accuracy', metric, True

def calc_precision(y_pred, y_true):
    metric = precision_score(y_true, y_pred)
    return 'precison', metric, True

def calc_recall(y_pred, y_true):
    metric = recall_score(y_true, y_pred)
    return 'recall', metric, True

def calc_fscore(fscores):
    metric = np.max(fscores)
    return 'fscore', metric, True

def calc_auc(y_pred, y_true):
    metric = roc_auc_score(y_true, y_pred)
    return 'auc', metric, True

def calc_prauc(y_pred, y_true):
    metric = average_precision_score(y_true, y_pred)
    return 'prauc', metric, True

# CV用評価関数算出メソッド
def acquire_custom_metircs(preds ,data):
    precisions, recalls, thresholds = precision_recall_curve(data.get_label(), preds, pos_label=1)
    thresholds = np.append(thresholds, 1)
    fscores = 2 * precisions * recalls / (precisions + recalls)
    threshold_max_fscore = thresholds[np.argmax(fscores)]

    y_true = data.get_label()
    y_pred = np.where(preds > 0.5, 1, 0)

    return [
        calc_accuracy(y_pred, y_true),  ＃ Accuracy
        calc_precision(y_pred, y_true),  ＃ Precision
        calc_recall(y_pred, y_true),  # Recall
        calc_fscore(fscores),  # F1 score
        calc_auc(y_pred, y_true),  # ROC AUC
        calc_prauc(y_pred, y_true)  # PR AUC
    ]

# CV
histroy = lgb.cv(
    best,
    dtrain, 
    folds=StratifiedShuffleSplit(n_splits=5, test_size=0.2, random_state=0), 
    feval=acquire_custom_metircs,  ＃ CVで使う評価関数に自作関数を指定
    num_boost_round=cons_val_num_boost_round, 
    early_stopping_rounds=cons_val_early_stopping_rounds, 
    verbose_eval=cons_val_verbose_eval, 
    show_stdv=False, 
    seed=0
)