Machine Learning Academy · Lección

Pesos de clase y ajuste del umbral

Pesos de clase y ajuste del umbral es una lección gratuita de Machine Learning Academy en CoddyKit. Esta es la lección 4 de 4. Puedes leer la lección completa abajo gratuitamente — luego la practicas en el navegador con un editor de código integrado y un tutor de IA 24/7. Forma parte de la ruta de aprendizaje de Machine Learning Academy, y tu progreso se sincroniza en la web y la app de CoddyKit. El curso de Machine Learning Academy incluye 4 lecciones en total.

Partes de esta lección aún no han sido traducidas y se muestran en inglés.

Two Alternatives to Resampling

Resampling (SMOTE, undersampling) physically changes the training data. Two alternative approaches work directly with the original data: class weighting penalises misclassifying minority samples more heavily during training, and threshold moving adjusts the decision boundary after training. Both are simpler, faster, and avoid the information loss or synthetic-noise risks of resampling.

Class Weights: Penalise Minority Errors More

Most sklearn classifiers accept a class_weight parameter. Setting class_weight='balanced' automatically computes weights inversely proportional to class frequencies: weight[c] = n_samples / (n_classes * count[c]). Misclassifying a rare positive sample is then penalised much more than misclassifying a common negative sample, pushing the model to learn the minority class better.

from sklearn.linear_model import LogisticRegression
from sklearn.metrics import classification_report
from sklearn.preprocessing import StandardScaler
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split

X, y = make_classification(n_samples=1000, weights=[0.95, 0.05], random_state=42)
X_train, X_test, y_train, y_test = train_test_split(X, y, stratify=y, random_state=42)

sc = StandardScaler()
X_tr = sc.fit_transform(X_train)
X_te = sc.transform(X_test)

# With balanced class weights
lr = LogisticRegression(class_weight='balanced').fit(X_tr, y_train)
print('--- class_weight=balanced ---')
print(classification_report(y_test, lr.predict(X_te)))

Computing Balanced Weights Manually

You can also pass a dictionary of custom weights if you want finer control than 'balanced'. For example, if your business says a missed fraud is 20x more costly than a false alarm, set class_weight={0: 1, 1: 20}. This directly encodes the misclassification cost ratio into training.

from sklearn.linear_model import LogisticRegression
from sklearn.utils.class_weight import compute_class_weight
import numpy as np

# Compute balanced weights automatically
y_train = np.array([0] * 475 + [1] * 25)
classes = np.unique(y_train)
weights = compute_class_weight('balanced', classes=classes, y=y_train)
weight_dict = dict(zip(classes, weights))
print('Auto-balanced weights:', weight_dict)

# Custom: penalty 10x higher for missing positive
custom_weights = {0: 1, 1: 10}
print('Custom weights:', custom_weights)

lr = LogisticRegression(class_weight=custom_weights)
# lr.fit(X_train, y_train)

Which Algorithms Support class_weight?

In scikit-learn, these estimators accept class_weight: LogisticRegression, LinearSVC, SVC, SGDClassifier, DecisionTreeClassifier, RandomForestClassifier. Gradient boosting models (XGBoost, LightGBM) use a scale_pos_weight parameter that plays the same role. Neural networks handle it through sample_weight in the loss function.

from sklearn.ensemble import RandomForestClassifier
from sklearn.tree import DecisionTreeClassifier
from sklearn.svm import SVC
from sklearn.linear_model import LogisticRegression

# All support class_weight='balanced'
models = [
    LogisticRegression(class_weight='balanced'),
    DecisionTreeClassifier(class_weight='balanced'),
    RandomForestClassifier(class_weight='balanced'),
    SVC(class_weight='balanced', probability=True)
]

print('All these models support class_weight:')
for m in models:
    print(' ', m.__class__.__name__)

Threshold Moving: Adjusting the Decision Boundary

Classifiers output a probability score; the default decision threshold is 0.5 — predict positive if proba >= 0.5. For imbalanced problems, lowering the threshold (e.g., to 0.3) increases recall (catches more positives) at the cost of more false positives. Raising it increases precision at the cost of missing more positives. Threshold moving is applied after training and does not require retraining the model.

from sklearn.linear_model import LogisticRegression
from sklearn.metrics import precision_score, recall_score, f1_score
from sklearn.preprocessing import StandardScaler
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
import numpy as np

X, y = make_classification(n_samples=1000, weights=[0.95, 0.05], random_state=42)
X_tr, X_te, y_tr, y_te = train_test_split(X, y, stratify=y, random_state=42)
sc = StandardScaler()
X_tr_s = sc.fit_transform(X_tr)
X_te_s = sc.transform(X_te)

lr = LogisticRegression(class_weight='balanced').fit(X_tr_s, y_tr)
proba = lr.predict_proba(X_te_s)[:, 1]

print(f'{'Threshold':>10}  {'Precision':>10}  {'Recall':>8}  {'F1':>6}')
for t in [0.2, 0.3, 0.4, 0.5, 0.6]:
    preds = (proba >= t).astype(int)
    p = precision_score(y_te, preds, zero_division=0)
    r = recall_score(y_te, preds)
    f = f1_score(y_te, preds, zero_division=0)
    print(f'{t:>10.1f}  {p:>10.4f}  {r:>8.4f}  {f:>6.4f}')

Finding the Optimal Threshold

Use the precision_recall_curve and roc_curve functions to sweep all possible thresholds and compute corresponding precision and recall. Then choose the threshold that maximises F1 (or any other business metric). This is more principled than guessing a threshold value.

from sklearn.metrics import precision_recall_curve, f1_score
import numpy as np

# proba and y_te from previous step
precisions, recalls, thresholds = precision_recall_curve(y_te, proba)

# F1 at each threshold
f1_scores = 2 * (precisions * recalls) / (precisions + recalls + 1e-8)
best_idx = np.argmax(f1_scores)
best_threshold = thresholds[best_idx]
best_f1 = f1_scores[best_idx]

print(f'Optimal threshold: {best_threshold:.4f}')
print(f'Best F1 at that threshold: {best_f1:.4f}')

# Apply the optimal threshold
y_pred_opt = (proba >= best_threshold).astype(int)
print('Minority class recall:', recall_score(y_te, y_pred_opt).round(4))

class_weight vs Resampling: Trade-Offs

class_weight advantages: no data modification, no risk of overfitting to synthetic samples, works inside a standard sklearn Pipeline, faster. Resampling advantages: works with algorithms that do not support class_weight (e.g., some boosting variants), can help with very severe imbalance (>100:1) where class weighting alone is not enough. In practice, try class weighting first — it is simpler and often sufficient.

Threshold Moving With ROC Curve

The ROC curve shows the trade-off between true-positive rate (recall) and false-positive rate at every threshold. The Youden's J statistic (TPR - FPR) is maximised at the optimal threshold for balanced sensitivity and specificity. Use this when you want equal consideration of both error types.

from sklearn.metrics import roc_curve
import numpy as np

fpr, tpr, thresholds = roc_curve(y_te, proba)

# Youden's J: maximise TPR - FPR
j_scores = tpr - fpr
best_idx = np.argmax(j_scores)
optimal_threshold = thresholds[best_idx]

print(f'Optimal threshold (Youden J): {optimal_threshold:.4f}')
print(f'TPR: {tpr[best_idx]:.4f}  FPR: {fpr[best_idx]:.4f}')

y_pred_youden = (proba >= optimal_threshold).astype(int)
from sklearn.metrics import classification_report
print(classification_report(y_te, y_pred_youden))

Class Weights in XGBoost and LightGBM

XGBoost uses scale_pos_weight — the ratio of negative to positive samples in the training set — to upweight positive examples. For 95:5 imbalance, set scale_pos_weight=19 (95/5). LightGBM uses is_unbalance=True or scale_pos_weight similarly. Both are equivalent to sklearn's class_weight='balanced' but use a different parameter name convention.

import xgboost as xgb
import numpy as np
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearn.metrics import roc_auc_score

X, y = make_classification(n_samples=1000, weights=[0.95, 0.05], random_state=42)
X_tr, X_te, y_tr, y_te = train_test_split(X, y, stratify=y, random_state=42)

neg, pos = np.bincount(y_tr)
scale = neg / pos
print(f'scale_pos_weight = {scale:.1f}')

clf = xgb.XGBClassifier(scale_pos_weight=scale, random_state=42, eval_metric='logloss')
clf.fit(X_tr, y_tr)
print('AUC:', roc_auc_score(y_te, clf.predict_proba(X_te)[:, 1]).round(4))

Comparing All Imbalance Strategies

A systematic comparison across strategies on the same dataset and test set reveals which approach works best for your specific problem and algorithm. Run this comparison and report results in a table to justify your final strategy choice.

from imblearn.over_sampling import SMOTE
from imblearn.under_sampling import RandomUnderSampler
from imblearn.pipeline import Pipeline as ImbPipeline
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import roc_auc_score
from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
import numpy as np

X, y = make_classification(n_samples=1000, weights=[0.95, 0.05], random_state=42)
X_tr, X_te, y_tr, y_te = train_test_split(X, y, stratify=y, random_state=0)
sc = StandardScaler()
X_tr_s = sc.fit_transform(X_tr)
X_te_s = sc.transform(X_te)

strategies = [
    ('Baseline', None, {}),
    ('class_weight', None, {'class_weight': 'balanced'}),
    ('RUS', RandomUnderSampler(random_state=0), {}),
    ('SMOTE', SMOTE(random_state=0), {})
]

for name, sampler, kwargs in strategies:
    Xr, yr = (sampler.fit_resample(X_tr_s, y_tr) if sampler else (X_tr_s, y_tr))
    lr = LogisticRegression(**kwargs).fit(Xr, yr)
    auc = roc_auc_score(y_te, lr.predict_proba(X_te_s)[:, 1])
    print(f'{name:15s}: AUC={auc:.4f}')

Calibrating Probability Outputs

Class weights and threshold moving both rely on the model's probability scores. If the classifier is poorly calibrated (e.g., SVM with probability=True using Platt scaling), threshold tuning may not work well. Use sklearn.calibration.CalibratedClassifierCV or compare calibration curves with calibration_curve to ensure probabilities are reliable before threshold optimisation.

Quick Check

Test your understanding of class weights and threshold moving from this lesson.

Lesson Recap

In this lesson you learned: class_weight='balanced' penalises minority-class misclassifications proportionally and works without modifying training data, threshold moving adjusts the decision boundary post-training to shift the recall-precision trade-off, and the optimal threshold can be found by maximising F1 over the precision-recall curve or Youden's J over the ROC curve. Next up we save trained models with joblib and pickle for production deployment.

Preguntas frecuentes

¿La lección «Pesos de clase y ajuste del umbral» es gratis?

Sí — el texto completo de «Pesos de clase y ajuste del umbral» es gratis para leer aquí en la web. Para practicarla de forma interactiva (editor de código integrado y tutor de IA 24/7) y desbloquear el resto del curso de Machine Learning Academy, actualiza a CoddyKit PRO. El curso de Machine Learning Academy incluye 4 lecciones en total.

¿Qué aprenderé en «Pesos de clase y ajuste del umbral»?

Establecerá class_weight='balanced' en clasificadores de sklearn y después cambiará el umbral de decisión de las probabilidades para optimizar aún más la exhaustividad en eventos poco frecuentes. Practicas Machine Learning Academy con código real que ejecutas directamente en el navegador, y un tutor de IA 24/7 responde tus preguntas mientras trabajas en la lección.

¿Necesito experiencia previa para empezar Machine Learning Academy?

No se requiere experiencia previa. Machine Learning Academy en CoddyKit está estructurado para principiantes hasta estudiantes avanzados, así que puedes empezar aquí o desde el inicio y avanzar a tu ritmo. Esta es la lección 4 de 4.

¿Cuánto tiempo toma la lección «Pesos de clase y ajuste del umbral»?

La mayoría de las lecciones de CoddyKit toman alrededor de 5–10 minutos. Cada una es compacta e interactiva, así que avanzas constantemente y retomas exactamente por donde dejaste en la web y la app.

¿Puedo escribir y ejecutar código en esta lección de Machine Learning Academy?

Sí. Cada lección de Machine Learning Academy incluye un editor de código integrado, así que escribes y ejecutas código real directamente en tu navegador y obtienes retroalimentación instantánea de IA — sin configuración local necesaria.

Todas las lecciones de este curso

Detección del desequilibrio: distribución de clases y problemas de las líneas base
Sobremuestreo aleatorio y SMOTE
Submuestreo aleatorio y centroides de clústeres
Pesos de clase y ajuste del umbral

← Volver a Machine Learning Academy