Hiperparámetros clave: learning rate, n_estimators y max_depth
Analice mediante experimentos sistemáticos el efecto del shrinkage (learning rate), el tamaño del ensemble y la profundidad en el equilibrio entre sesgo y varianza.
Hiperparámetros clave: learning rate, n_estimators y max_depth 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.
The Holy Trinity of Gradient Boosting
Gradient boosting has three hyperparameters that interact most strongly and have the biggest impact on model performance: learning_rate (how large each tree's contribution is), n_estimators (how many trees to build), and max_depth (how complex each individual tree can be). Understanding how these three interact is the key to effectively tuning any gradient boosting model — whether scikit-learn's GradientBoostingClassifier, XGBoost, or LightGBM.
Learning Rate (Shrinkage) in Depth
The learning rate η (eta) scales each tree's contribution: F_m(x) = F_{m-1}(x) + η × h_m(x). A smaller η means each tree corrects only a fraction of the residual, requiring more trees to converge but producing a smoother, more regularised prediction surface. Typical values: 0.01-0.3. The inverse relationship between learning_rate and n_estimators is crucial: halving the learning rate requires roughly doubling n_estimators to reach the same training loss. Always use early stopping to find the right n_estimators for your chosen learning rate.
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import cross_val_score
X, y = load_breast_cancer(return_X_y=True)
# Inverse relationship between lr and n_estimators
configs = [(0.3, 50), (0.1, 150), (0.05, 300), (0.01, 1000)]
for lr, n in configs:
model = GradientBoostingClassifier(learning_rate=lr, n_estimators=n, max_depth=3, random_state=42)
score = cross_val_score(model, X, y, cv=5).mean()
print(f'lr={lr:.2f}, n={n:4d}: CV={score:.4f}')n_estimators: More Trees, Not Always Better
n_estimators sets the total number of sequential trees in the ensemble. Unlike random forests (where more is always better), boosting models can overfit with too many trees, especially at high learning rates. The sweet spot is found by monitoring validation loss across rounds. With early stopping, you can safely set n_estimators=2000 — training will stop when validation stops improving. Without early stopping, use cross-validation on a range of values and pick the n_estimators at the elbow of the validation error curve.
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import cross_val_score
X, y = load_breast_cancer(return_X_y=True)
for n in [10, 50, 100, 200, 500]:
model = GradientBoostingClassifier(n_estimators=n, learning_rate=0.1, max_depth=3, random_state=42)
score = cross_val_score(model, X, y, cv=5).mean()
print(f'n_estimators={n:4d}: CV accuracy={score:.4f}')max_depth: Weak Learners Win in Boosting
Gradient boosting works best with weak learners — shallow trees (max_depth 2-6). A depth-1 tree (decision stump) captures only one feature interaction per round; a depth-3 tree captures three-way interactions. Deeper trees are more powerful individually, requiring fewer rounds but risking overfitting. Shallower trees need more rounds but generalise better. In practice: start with max_depth=3 for classification, max_depth=4-6 for complex tabular regression. Very deep trees (>8) are almost never optimal in boosting.
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import cross_val_score
X, y = load_breast_cancer(return_X_y=True)
for depth in [1, 2, 3, 5, 8, 15]:
model = GradientBoostingClassifier(n_estimators=200, learning_rate=0.1, max_depth=depth, random_state=42)
train_score = model.fit(X, y).train_score_[-1]
cv_score = cross_val_score(GradientBoostingClassifier(n_estimators=200, learning_rate=0.1, max_depth=depth, random_state=42), X, y, cv=5).mean()
print(f'max_depth={depth:2d}: train={train_score:.4f}, CV={cv_score:.4f}')The Learning Rate - n_estimators Trade-Off
The interaction between learning_rate and n_estimators is the most important trade-off to understand. Consider these scenarios:
- High lr (0.3) + few trees (50): fast training, but risks missing the optimal fit
- Medium lr (0.1) + medium trees (200): the standard starting point
- Low lr (0.01) + many trees (2000): most regularised, best generalisation but slow to train
Training vs Validation Loss Curves
One of the most informative diagnostics in gradient boosting is plotting training and validation loss across boosting rounds. Training loss always decreases monotonically. Validation loss decreases initially but starts to rise when the model overfits. The point where validation loss is minimised is the optimal n_estimators. This plot reveals whether you need more regularisation (validation loss rises quickly), less regularisation (validation loss still decreasing at the end), or whether you have found the sweet spot.
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
import numpy as np
X, y = load_breast_cancer(return_X_y=True)
X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.2, random_state=42)
model = GradientBoostingClassifier(n_estimators=300, learning_rate=0.1, max_depth=3, random_state=42)
model.fit(X_train, y_train)
val_scores = [model.estimators_[:i+1] and model.loss_(y_val, model.staged_predict_proba(X_val).__next__()) for i in range(5)]
# Simplified: use staged_predict to check improvement
val_staged = list(model.staged_predict(X_val))
val_acc = [np.mean(p == y_val) for p in val_staged]
print('Val acc at round 50:', round(val_acc[49], 4))
print('Val acc at round 300:', round(val_acc[299], 4))
print('Best round:', np.argmax(val_acc) + 1)Regularisation Beyond the Three Parameters
Beyond learning_rate, n_estimators, and max_depth, gradient boosting offers additional regularisation: min_samples_leaf (minimum examples per leaf, prevents overfitting on tiny leaf groups), subsample (stochastic training, reduces variance), max_features (feature subsampling per split, reduces correlation), min_impurity_decrease (only split if the gain exceeds a threshold). These secondary parameters typically matter less than the primary three but can make a meaningful difference on noisy datasets.
Systematic Hyperparameter Search Strategy
A practical 3-step strategy for gradient boosting tuning: (1) Fix learning_rate=0.1, set n_estimators=1000 with early stopping — find optimal number of rounds; (2) Grid search max_depth (try 3, 4, 5, 6) and min_samples_leaf (try 1, 5, 10) with the tree count from step 1; (3) Lower the learning rate to 0.05 or 0.01 and scale n_estimators accordingly to potentially squeeze out more performance. This avoids the curse of dimensionality from searching all parameters simultaneously.
subsample as a Stochastic Regulariser
Setting subsample < 1.0 (e.g., 0.8) in gradient boosting means each tree is trained on a random 80% of the training data (without replacement). This is called stochastic gradient boosting. The stochastic element introduces diversity among trees (similar to dropout in neural networks) and often improves validation performance by 0.5-2%. A typical value is 0.8. If you reduce it further (0.5), you may need more trees to compensate for the higher variance per tree.
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import cross_val_score
X, y = load_breast_cancer(return_X_y=True)
for ss in [1.0, 0.8, 0.6, 0.5]:
model = GradientBoostingClassifier(n_estimators=200, learning_rate=0.1, max_depth=3,
subsample=ss, random_state=42)
score = cross_val_score(model, X, y, cv=5).mean()
print(f'subsample={ss}: CV accuracy={score:.4f}')Balancing Training Time and Performance
In production, training time is a real constraint. Larger n_estimators and more CV folds multiply linearly. Practical shortcuts: (1) use n_jobs=-1 for cross-validation parallelism; (2) use LightGBM or XGBoost instead of sklearn's GradientBoosting for 5-10x speed gains; (3) use RandomizedSearchCV with 30-50 iterations instead of full grid search; (4) use early stopping within cross-validation folds to automatically set n_estimators. The combination of early stopping + LightGBM + RandomizedSearchCV dramatically reduces tuning time without sacrificing accuracy.
Gradient Boosting Default Hyperparameters
Good default starting points for gradient boosting hyperparameters: learning_rate=0.1, n_estimators=200 (with early stopping on hold-out), max_depth=3 for classification or 4-5 for regression, subsample=0.8 for stochastic regularisation, and min_samples_leaf=5-10 to prevent over-splitting on small groups. These defaults work surprisingly well across many datasets and should be your starting point before any tuning. Only tune if the baseline performance is insufficient for your use case.
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.datasets import load_breast_cancer
from sklearn.model_selection import cross_val_score
X, y = load_breast_cancer(return_X_y=True)
# Good defaults — try these before tuning
default_gb = GradientBoostingClassifier(
learning_rate=0.1, n_estimators=200, max_depth=3,
subsample=0.8, min_samples_leaf=5, random_state=42
)
scores = cross_val_score(default_gb, X, y, cv=5)
print('Default GBM CV:', scores.mean().round(4), '+/-', scores.std().round(4))Quick Check
Test your understanding of gradient boosting hyperparameters from this lesson.
Lesson Recap
In this lesson you learned: learning rate and n_estimators have an inverse relationship — lower lr requires more trees, shallow trees (max_depth 2-5) are preferable in boosting because they act as weak learners, and subsample and min_samples_leaf provide additional regularisation beyond the primary three parameters. Next up we explore K-Fold Cross-Validation and how to evaluate models without leaking test data.
Preguntas frecuentes
¿La lección «Hiperparámetros clave: learning rate, n_estimators y max_depth» es gratis?
Sí — el texto completo de «Hiperparámetros clave: learning rate, n_estimators y max_depth» 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 «Hiperparámetros clave: learning rate, n_estimators y max_depth»?
Analice mediante experimentos sistemáticos el efecto del shrinkage (learning rate), el tamaño del ensemble y la profundidad en el equilibrio entre sesgo y varianza. 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 «Hiperparámetros clave: learning rate, n_estimators y max_depth»?
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
- Intuición sobre Boosting: corrección secuencial de errores
- XGBoost: regularización, parada temprana e importancia de las características
- LightGBM: crecimiento hoja a hoja y ventajas de velocidad
- Hiperparámetros clave: learning rate, n_estimators y max_depth