如何针对Optuna中的多个指标进行优化

Question

如何在Optuna的objective函数内同时优化多个指标?例如，我正在训练一个LGBM分类器，并希望为所有常见的分类指标(如F1、精度、召回率、准确性、AUC等)找到最佳超参数集。

def objective(trial):
# Train
gbm = lgb.train(param, dtrain)
preds = gbm.predict(X_test)
pred_labels = np.rint(preds)
# Calculate metrics
accuracy = sklearn.metrics.accuracy_score(y_test, pred_labels)
recall = metrics.recall_score(pred_labels, y_test)
precision = metrics.precision_score(pred_labels, y_test)
f1 = metrics.f1_score(pred_labels, y_test, pos_label=1)
...

我该怎么做?

Answer 1

在定义网格并使用这些参数拟合模型并生成预测之后，计算您想要优化的所有指标:

def objective(trial):
param_grid = {"n_estimators": trial.suggest_int("n_estimators", 2000, 10000, step=200)}
clf = lgbm.LGBMClassifier(objective='binary', **param_grid)
clf.fit(X_train, y_train)
preds = clf.predict(X_valid)
probs = clf.predict_proba(X_valid)

# Metrics
f1 = sklearn.metrics.f1_score(y_valid, press)
accuracy = ...
precision = ...
recall = ...
logloss = ...

并按您想要的顺序返回它们:

def objective(trial):
...
return f1, logloss, accuracy, precision, recall

然后，在研究对象中，指定您是希望最小化还是最大化directions的每个度量，如下所示:

study = optuna.create_study(directions=['maximize', 'minimize', 'maximize', 'maximize', 'maximize'])
study.optimize(objective, n_trials=100)

有关详细信息，请参阅文档中的使用Optuna进行多目标优化。