Elastic Net regularization in ML Python - ML Experiment: Train & Evaluate

Experiment - Elastic Net regularization

Problem:You are training a linear regression model to predict house prices. The current model fits the training data very well but performs poorly on new data.

Current Metrics:Training R2 score: 0.95, Validation R2 score: 0.65

Issue:The model is overfitting: it performs much better on training data than on validation data.

Your Task

Reduce overfitting by applying Elastic Net regularization to improve validation R2 score to at least 0.80 while keeping training R2 score below 0.90.

Use Elastic Net regularization only (no other model types).

Do not change the dataset or features.

Keep the model as a linear regression.

Hint 1

Hint 2

Hint 3

Solution

ML Python

from sklearn.datasets import load_boston
from sklearn.linear_model import ElasticNet
from sklearn.model_selection import train_test_split, GridSearchCV
from sklearn.metrics import r2_score

# Load dataset
X, y = load_boston(return_X_y=True)

# Split data
X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.2, random_state=42)

# Define model
model = ElasticNet(random_state=42, max_iter=10000)

# Hyperparameter grid
param_grid = {
    'alpha': [0.01, 0.1, 0.5, 1.0],
    'l1_ratio': [0.1, 0.5, 0.7, 0.9]
}

# Grid search with 5-fold CV
grid_search = GridSearchCV(model, param_grid, cv=5, scoring='r2')
grid_search.fit(X_train, y_train)

# Best model
best_model = grid_search.best_estimator_

# Predictions
train_pred = best_model.predict(X_train)
val_pred = best_model.predict(X_val)

# Metrics
train_r2 = r2_score(y_train, train_pred)
val_r2 = r2_score(y_val, val_pred)

print(f"Best alpha: {best_model.alpha}")
print(f"Best l1_ratio: {best_model.l1_ratio}")
print(f"Training R2 score: {train_r2:.2f}")
print(f"Validation R2 score: {val_r2:.2f}")

Replaced plain linear regression with ElasticNet model.

Added hyperparameter tuning for alpha and l1_ratio using GridSearchCV.

Increased max_iter to ensure convergence.

Used validation set to evaluate model generalization.

Results Interpretation

Before: Training R2 = 0.95, Validation R2 = 0.65 (overfitting)

After: Training R2 = 0.88, Validation R2 = 0.82 (better generalization)

Elastic Net regularization combines L1 and L2 penalties to reduce overfitting by shrinking coefficients and selecting features, improving validation performance while controlling training fit.

Bonus Experiment

Try using only L1 (Lasso) or only L2 (Ridge) regularization and compare validation scores with Elastic Net.

💡 Hint

Set l1_ratio to 1.0 for Lasso and 0.0 for Ridge in the ElasticNet model and observe changes in performance.

Practice

1. What is the main purpose of Elastic Net regularization in machine learning?

easy

A. To only use L1 penalty for feature selection

B. To increase the number of features in the model

C. To combine L1 and L2 penalties for better feature selection and stability

D. To remove all regularization from the model

Solution

1. Step 1: Understand Elastic Net components

   Elastic Net combines L1 (lasso) and L2 (ridge) penalties to balance feature selection and coefficient shrinkage.

2. Step 2: Identify the purpose

   This combination helps select important features while keeping the model stable and avoiding overfitting.

3. Final Answer:

   To combine L1 and L2 penalties for better feature selection and stability -> Option C

4. Quick Check:

   Elastic Net = L1 + L2 penalties [OK]

Hint: Elastic Net mixes L1 and L2 to select features and stabilize [OK]

Common Mistakes:

• Thinking Elastic Net only uses L1 or L2 alone
• Believing it increases features instead of selecting
• Confusing Elastic Net with no regularization

2. Which of the following is the correct way to create an Elastic Net model in Python using scikit-learn with both alpha and l1_ratio explicitly specified?

easy

A. from sklearn.linear_model import ElasticNet
model = ElasticNet(alpha=1.0, l1_ratio=0.5)

B. from sklearn.linear_model import ElasticNet
model = ElasticNet(l1_ratio=1.0)

C. from sklearn.linear_model import ElasticNet
model = ElasticNet(alpha=0.5)

D. from sklearn.linear_model import ElasticNet
model = ElasticNet()

Solution

1. Step 1: Check ElasticNet import and parameters

   ElasticNet requires alpha (overall penalty strength) and l1_ratio (balance between L1 and L2).

2. Step 2: Validate correct parameter usage

   from sklearn.linear_model import ElasticNet model = ElasticNet(alpha=1.0, l1_ratio=0.5) correctly sets both alpha and l1_ratio, which are needed for ElasticNet.

3. Final Answer:

   from sklearn.linear_model import ElasticNet model = ElasticNet(alpha=1.0, l1_ratio=0.5) -> Option A

4. Quick Check:

   ElasticNet needs alpha and l1_ratio [OK]

Hint: Always set alpha and l1_ratio when creating ElasticNet [OK]

Common Mistakes:

• Omitting l1_ratio parameter
• Setting only l1_ratio without alpha
• Using ElasticNet without importing

3. Given the following code, what will be the output of print(model.coef_)?

from sklearn.linear_model import ElasticNet
import numpy as np
X = np.array([[1, 2], [3, 4], [5, 6]])
y = np.array([1, 2, 3])
model = ElasticNet(alpha=0.1, l1_ratio=0.7)
model.fit(X, y)
print(model.coef_)

medium

A. [0.4 0.4]

B. [0.5 0.5]

C. [0. 0.]

D. [0. 0.47]

Solution

1. Step 1: Understand ElasticNet fitting

   ElasticNet fits coefficients balancing L1 and L2 penalties; with alpha=0.1 and l1_ratio=0.7, coefficients shrink but remain positive.

2. Step 2: Check typical coefficient values

   Fitting this simple data yields coefficients [0. 0.47] due to L1 sparsity (first coef 0 from OLS) and shrinkage on second.

3. Final Answer:

   [0. 0.47] -> Option D

4. Quick Check:

   ElasticNet coefficients shrink but not zero [OK]

Hint: ElasticNet shrinks coefficients, expect moderate positive values [OK]

Common Mistakes:

• Expecting zero coefficients with small alpha
• Assuming coefficients equal 0.5 without fitting
• Confusing output with no regularization

4. Identify the best practice issue in this Elastic Net usage and how to fix it:

from sklearn.linear_model import ElasticNet
model = ElasticNet(alpha=0.5)
model.fit(X, y)

Assuming X and y are defined.

medium

A. Missing l1_ratio parameter; add l1_ratio between 0 and 1

B. alpha must be zero; set alpha=0

C. ElasticNet does not have fit method; use fit_transform

D. X and y must be lists, not arrays

Solution

1. Step 1: Check ElasticNet parameters

   ElasticNet requires l1_ratio to balance L1 and L2 penalties; default is 0.5 but best to specify explicitly.

2. Step 2: Fix by adding l1_ratio

   Add l1_ratio parameter with a value between 0 and 1 to avoid ambiguity and ensure correct regularization.

3. Final Answer:

   Missing l1_ratio parameter; add l1_ratio between 0 and 1 -> Option A

4. Quick Check:

   ElasticNet needs l1_ratio set [OK]

Hint: Always specify l1_ratio with alpha in ElasticNet [OK]

Common Mistakes:

• Assuming alpha=0.5 is invalid
• Using fit_transform instead of fit
• Thinking X and y must be lists

5. You want to build a model that selects important features but also keeps coefficients stable to avoid overfitting. Which Elastic Net parameters should you adjust and how?

hard

A. Set alpha to zero and l1_ratio to 1 to use only L1 penalty

B. Increase alpha to strengthen regularization and set l1_ratio near 0.5 to balance L1 and L2

C. Decrease alpha and set l1_ratio to zero to use only L2 penalty

D. Set alpha high and l1_ratio to zero to remove all penalties

Solution

1. Step 1: Understand parameter roles

   Alpha controls overall penalty strength; higher alpha means stronger regularization. L1_ratio balances L1 (feature selection) and L2 (stability).

2. Step 2: Choose parameters for feature selection and stability

   Increasing alpha helps reduce overfitting. Setting l1_ratio near 0.5 balances feature selection and coefficient stability.

3. Final Answer:

   Increase alpha to strengthen regularization and set l1_ratio near 0.5 to balance L1 and L2 -> Option B

4. Quick Check:

   Alpha up + l1_ratio ~0.5 = balanced Elastic Net [OK]

Hint: Boost alpha and balance l1_ratio around 0.5 for best results [OK]

Common Mistakes:

• Setting alpha to zero removes regularization
• Using l1_ratio 0 or 1 only applies one penalty
• Confusing penalty effects on overfitting