The meaning of parameter optimization or cross-validation evaluation indicators in machine learning

In Scikit-Learn, the cross_val_score function supports a variety of different scoring criteria (scoring parameters). Here are some common scoring criteria and their application scenarios:

Reference:
https://blog.csdn.net/worther/article/details/126909270

https://zhuanlan.zhihu.com/p/509437755

• Classification question:

•   accuracy: 准确率
    roc_auc, roc_auc_ovo, roc_auc_ovr: ROC曲线下面积
    average_precision: 平均精度
    f1, f1_macro, f1_micro, f1_weighted: F1分数
    precision, precision_macro, precision_micro, precision_weighted: 精度
    recall, recall_macro, recall_micro, recall_weighted: 召回率
    balanced_accuracy: 平衡准确率
  

• Regression problem:

•   neg_mean_squared_error: 均方误差（负值）
    neg_mean_absolute_error: 平均绝对误差（负值）
    neg_root_mean_squared_error: 均方根误差（负值）
    neg_median_absolute_error: 中位绝对误差（负值）
    r2: 决定系数（R²）
    explained_variance: 解释方差
    max_error: 最大误差
  

• Clustering problem:

•   adjusted_rand_score: 调整后的Rand指数
    homogeneity_score: 同质性得分
    completeness_score: 完整性得分
    v_measure_score: V-measure得分
    adjusted_mutual_info_score: 调整后的互信息得分
    normalized_mutual_info_score: 标准化互信息得分
  

Each of these scoring standards has its own applicable scenarios and characteristics. Which scoring criteria you choose depends on your specific task and model evaluation needs. For example, in a regression task, if you care about the size of the prediction error, you can choose neg_mean_squared_error or neg_mean_absolute_error; and if you care about the proportion of variance explained by the model, you can choose r2.

One word per text

come on