python - 如何针对 Optuna 中的多个指标进行优化

Question

如何在 Optuna 的 objective 函数中同时优化多个指标。例如，我正在训练一个 LGBM 分类器，并希望为所有常见分类指标(如 F1、精度、召回率、准确率、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，指定是否要将每个指标最小化或最大化到 directions，如下所示:

study = optuna.create_study(directions=['maximize', 'minimize', 'maximize', 'maximize', 'maximize'])

study.optimize(objective, n_trials=100)

更多详情，见Multi-objective Optimization with Optuna在文档中。