기계 학습 모델에 대한 최상의 임계값 설정: 0.5가 이진 분류에 대한 최상의 임계값입니까?

import matplotlib.pyplot as plt
 import matplotlib as mpl
 from sklearn import datasets
 from sklearn.linear_model import LogisticRegression
 from sklearn.model_selection import train_test_split
 from sklearn_evaluation.plot import ConfusionMatrix
 
 # matplotlib settings
 mpl.rcParams['figure.figsize'] = (4, 4)
 mpl.rcParams['figure.dpi'] = 150
 
 # create sample dataset
 X, y = datasets.make_classification(1000, 10, n_informative=5, class_sep=0.4)
 X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)
 
 # fit model
 clf = LogisticRegression()
 _ = clf.fit(X_train, y_train)

# predict on the test set
 y_pred = clf.predict(X_test)
 
 # plot confusion matrix
 cm_dot_five = ConfusionMatrix(y_test, y_pred)
 cm_dot_five

y_score = clf.predict_proba(X_test)

cm_dot_four = ConfusionMatrix(y_score[:, 1] >= 0.4, y_pred)

cm_dot_five + cm_dot_four

curl -O https://raw.githubusercontent.com/ploomber/posts/master/threshold/fit.ipynb?utm_source=medium&utm_medium=blog&utm_campaign=threshold

ploomber cloud nb fit.ipynb

ploomber cloud status @latest --summary
 
 status count
 -------- -------
 finished 20
 
 Pipeline finished. Check outputs:
 $ ploomber cloud products

ploomber cloud download 'threshold-selection/*.csv' --summary

 from glob import glob
 
 import pandas as pd
 import numpy as np
 paths = glob('threshold-selection/**/*.csv')
 metrics = [pd.read_csv(path) for path in paths]
 
 for idx, df in enumerate(metrics):
plt.plot(df.threshold, df.precision, color='blue', alpha=0.2,
label='precision' if idx == 0 else None)
plt.plot(df.threshold, df.recall, color='green', alpha=0.2,
label='recall' if idx == 0 else None)
plt.plot(df.threshold, df.f1, color='orange', alpha=0.2,
label='f1' if idx == 0 else None)
 
 
 plt.grid()
 plt.legend()
 plt.xlabel('Threshold')
 plt.ylabel('Metric value')
 
 for handle in plt.legend().legendHandles:
handle.set_alpha(1)
 
 ax = plt.twinx()
 
 for idx, df in enumerate(metrics):
ax.plot(df.threshold, df.n_flagged,
label='flagged' if idx == 0 else None,
color='red', alpha=0.2)
 
 plt.ylabel('Flagged')
 ax.legend(loc=0)
 ax.legend().legendHandles[0].set_alpha(1)

shape = (df.shape[0], len(metrics))
 precision = np.zeros(shape)
 recall = np.zeros(shape)
 f1 = np.zeros(shape)
 n_flagged = np.zeros(shape)
 for i, df in enumerate(metrics):
precision[:, i] = df.precision.values
recall[:, i] = df.recall.values
f1[:, i] = df.f1.values
n_flagged[:, i] = df.n_flagged.values
 precision_ = np.quantile(precision, q=0.5, axis=1)
 recall_ = np.quantile(recall, q=0.5, axis=1)
 f1_ = np.quantile(f1, q=0.5, axis=1)
 n_flagged_ = np.quantile(n_flagged, q=0.5, axis=1)
 plt.plot(df.threshold, precision_, color='blue', label='precision')
 plt.plot(df.threshold, recall_, color='green', label='recall')
 plt.plot(df.threshold, f1_, color='orange', label='f1')
 
 plt.fill_between(df.threshold, precision_interval[0],
precision_interval[1], color='blue',
alpha=0.2)
 
 plt.fill_between(df.threshold, recall_interval[0],
recall_interval[1], color='green',
alpha=0.2)
 
 
 plt.fill_between(df.threshold, f1_interval[0],
f1_interval[1], color='orange',
alpha=0.2)
 plt.xlabel('Threshold')
 plt.ylabel('Metric value')
 plt.legend()
 
 ax = plt.twinx()
 ax.plot(df.threshold, n_flagged_, color='red', label='flagged')
 ax.fill_between(df.threshold, n_flagged_interval[0],
n_flagged_interval[1], color='red',
alpha=0.2)
 
 ax.legend(loc=3)
 
 plt.ylabel('Flagged')
 plt.grid()

idx = np.argmax(f1_)
 prec_lower, prec_upper = precision_interval[0][idx], precision_interval[1][idx]
 rec_lower, rec_upper = recall_interval[0][idx], recall_interval[1][idx]
 threshold = df.threshold[idx]
 
 print(f'Max F1 score: {f1_[idx]:.2f}')
 print('Metrics when maximizing F1 score:')
 print(f' - Threshold: {threshold:.2f}')
 print(f' - Precision range: ({prec_lower:.2f}, {prec_upper:.2f})')
 print(f' - Recall range: ({rec_lower:.2f}, {rec_upper:.2f})')
 
 #结果
 Max F1 score: 0.71
 Metrics when maximizing F1 score:
- Threshold: 0.26
- Precision range: (0.58, 0.61)
- Recall range: (0.86, 0.90)

idx = np.argmax(n_flagged_ <= 5000)
 
 prec_lower, prec_upper = precision_interval[0][idx], precision_interval[1][idx]
 rec_lower, rec_upper = recall_interval[0][idx], recall_interval[1][idx]
 threshold = df.threshold[idx]
 
 print('Metrics when limiting to a maximum of 5,000 flagged events:')
 print(f' - Threshold: {threshold:.2f}')
 print(f' - Precision range: ({prec_lower:.2f}, {prec_upper:.2f})')
 print(f' - Recall range: ({rec_lower:.2f}, {rec_upper:.2f})')
 
 # 结果
 Metrics when limiting to a maximum of 5,000 flagged events:
- Threshold: 0.82
- Precision range: (0.77, 0.81)
- Recall range: (0.25, 0.36)