我正在进行多类分类,类别不平衡。
我注意到 f1 总是小于准确率和召回率的直接调和平均值,在某些情况下,f1 甚至小于准确率和召回率。
仅供引用,我调用了 metrics.precision_score(y,pred)
以获得精度等。
我知道微观/宏观平均值的差异,并使用 precision_recall_fscore_support()
的类别结果测试它们不是微观。
不确定这是由于使用了宏平均值还是其他原因?
更新后的详细结果如下:
n_samples:75,n_features:250
MultinomialNB(alpha=0.01, fit_prior=True)
2 倍简历:
第一次运行:
F1: 0.706029106029
Precision: 0.731531531532
Recall: 0.702702702703
precision recall f1-score support
0 0.44 0.67 0.53 6
1 0.80 0.50 0.62 8
2 0.78 0.78 0.78 23
avg / total 0.73 0.70 0.71 37
第二次运行:
F1: 0.787944219523
Precision: 0.841165413534
Recall: 0.815789473684
precision recall f1-score support
0 1.00 0.29 0.44 7
1 0.75 0.86 0.80 7
2 0.82 0.96 0.88 24
avg / total 0.84 0.82 0.79 38
总体:
Overall f1-score: 0.74699 (+/- 0.02)
Overall precision: 0.78635 (+/- 0.03)
Overall recall: 0.75925 (+/- 0.03)
关于微观/宏观平均的定义来自 Scholarpedia :
In multi-label classification, the simplest method for computing an aggregate score across categories is to average the scores of all binary task. The resulted scores are called macro-averaged recall, precision, F1, etc. Another way of averaging is to sum over TP, FP, TN, FN and N over all the categories first, and then compute each of the above metrics. The resulted scores are called micro-averaged. Macro-averaging gives an equal weight to each category, and is often dominated by the system’s performance on rare categories (the majority) in a power-law like distribution. Micro-averaging gives an equal weight to each document, and is often dominated by the system’s performance on most common categories.
这是电流 open issue在 Github 上,#83。
以下示例演示了微观、宏观和加权(Scikit-learn 中的当前)平均可能有何不同:
y = [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2]
pred = [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 2, 0, 1, 2, 2, 2, 2]
混淆矩阵:
[[9 3 0]
[3 5 1]
[1 1 4]]
Wei Pre: 0.670655270655
Wei Rec: 0.666666666667
Wei F1 : 0.666801346801
Wei F5 : 0.668625356125
Mic Pre: 0.666666666667
Mic Rec: 0.666666666667
Mic F1 : 0.666666666667
Mic F5 : 0.666666666667
Mac Pre: 0.682621082621
Mac Rec: 0.657407407407
Mac F1 : 0.669777037588
Mac F5 : 0.677424801371
上面的F5是F0.5的简写...
最佳答案
您能否使用以下输出更新您的问题:
>>> from sklearn.metrics import classification_report
>>> print classification_report(y_true, y_predicted)
这将显示每个单独类别的准确率和召回率以及支持度,从而帮助我们理解平均的工作原理并决定这是否是一种适当的行为。
关于python - F1 小于 Scikit-learn 中的精度和召回率,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/8284456/