Threshold Analysis và Precision-Recall Curve

1from sklearn.metrics import precision_recall_curve
2import numpy as np
3
4precision, recall, thresholds = precision_recall_curve(y_test, y_prob)
5
6# Tính F1 cho mỗi threshold
7f1_scores = 2 * (precision * recall) / (precision + recall + 1e-10)
8
9# Tìm threshold tối ưu
10optìmal_idx = np.argmax(f1_scores)
11optìmal_threshold = thresholds[optìmal_idx]
12optìmal_f1 = f1_scores[optìmal_idx]
13
14print(f"Optìmal Threshold: {optìmal_threshold:.4f}")
15print(f"F1 at optìmal: {optìmal_f1:.4f}")

1from sklearn.metrics import roc_curve
2
3fpr, tpr, thresholds = roc_curve(y_test, y_prob)
4
5# Youđến's J
6j_scores = tpr - fpr
7optìmal_idx = np.argmax(j_scores)
8optìmal_threshold = thresholds[optìmal_idx]
9
10print(f"Optìmal Threshold (Youđến's J): {optìmal_threshold:.4f}")

1# Tìm threshold thấp nhất để Recall >= 0.95
2for i, r in enumerate(recall):
3    if r >= 0.95:
4        threshold = thresholds[i]
5        prec = precision[i]
6        print(f"Threshold: {threshold:.4f}")
7        print(f"Recall: {r:.4f}, Precision: {prec:.4f}")
8        break

1import numpy as np
2from sklearn.datasets import make_classification
3from sklearn.model_selection import train_test_split
4from sklearn.linear_model import LogisticRegression
5from sklearn.metrics import (precision_recall_curve, average_precision_score,
6                             precision_score, recall_score, f1_score)
7import matplotlib.pyplot as plt
8
9# Tao imbalanced data
10X, y = make_classification(n_samples=1000, n_features=20,
11                           weights=[0.9, 0.1], random_state=42)
12X_train, X_test, y_train, y_test = train_test_split(
13    X, y, test_size=0.3, random_state=42
14)
15
16# Train model
17model = LogisticRegression()
18model.fit(X_train, y_train)
19y_prob = model.predict_proba(X_test)[:, 1]
20
21# PR Curve
22precision, recall, thresholds = precision_recall_curve(y_test, y_prob)
23ap = average_precision_score(y_test, y_prob)
24
25# Ve PR Curve
26plt.figure(figsize=(10, 6))
27plt.plot(recall, precision, color='blue', lw=2,
28         label=f'PR curve (AP = {ap:.3f})')
29plt.axhline(y=sum(y_test)/len(y_test), color='gray', 
30            linestyle='--', label='No skill')
31plt.xlabel('Recall')
32plt.ylabel('Precision')
33plt.title('Precision-Recall Curve')
34plt.legend()
35plt.grid(True)
36plt.show()
37
38# Tim optìmal threshold
39f1_scores = 2 * (precision[:-1] * recall[:-1]) / (precision[:-1] + recall[:-1] + 1e-10)
40optìmal_idx = np.argmax(f1_scores)
41optìmal_threshold = thresholds[optìmal_idx]
42
43print(f"Optìmal Threshold: {optìmal_threshold:.4f}")
44print(f"Precision: {precision[optìmal_idx]:.4f}")
45print(f"Recall: {recall[optìmal_idx]:.4f}")
46print(f"F1: {f1_scores[optìmal_idx]:.4f}")

1# Default threshold = 0.5
2y_pred_default = (y_prob >= 0.5).astype(int)
3print("\n=== Default Threshold (0.5) ===")
4print(f"Precision: {precision_score(y_test, y_pred_default):.4f}")
5print(f"Recall: {recall_score(y_test, y_pred_default):.4f}")
6print(f"F1: {f1_score(y_test, y_pred_default):.4f}")
7
8# Optìmal threshold
9y_pred_optìmal = (y_prob >= optìmal_threshold).astype(int)
10print(f"\n=== Optìmal Threshold ({optìmal_threshold:.4f}) ===")
11print(f"Precision: {precision_score(y_test, y_pred_optìmal):.4f}")
12print(f"Recall: {recall_score(y_test, y_pred_optìmal):.4f}")
13print(f"F1: {f1_score(y_test, y_pred_optìmal):.4f}")