Stacking and blending in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - Stacking and blending

Which metric matters for Stacking and Blending and WHY

Stacking and blending combine multiple models to improve predictions. The key metric depends on the task: for classification, accuracy, precision, recall, and F1 score matter. For regression, mean squared error or R-squared are important. We focus on metrics that show if the combined model predicts better than individual models. This helps us know if stacking or blending truly improves results.

Confusion Matrix Example for Classification

Suppose we stack two models to detect spam emails. After combining, the confusion matrix might look like this:

      | Predicted Spam | Predicted Not Spam |
      |----------------|--------------------|
      | True Positives (TP) = 85           |
      | False Positives (FP) = 15          |
      | False Negatives (FN) = 10          |
      | True Negatives (TN) = 90           |
    

Total samples = 85 + 15 + 10 + 90 = 200

From this, we calculate:

• Precision = TP / (TP + FP) = 85 / (85 + 15) = 0.85
• Recall = TP / (TP + FN) = 85 / (85 + 10) = 0.8947
• F1 Score = 2 * (Precision * Recall) / (Precision + Recall) ≈ 0.871

Precision vs Recall Tradeoff in Stacking and Blending

Stacking and blending aim to balance precision and recall better than single models. For example:

• If a spam filter has high precision but low recall, it misses many spam emails (bad for users).
• If it has high recall but low precision, many good emails are marked spam (annoying).

Stacking can combine models that are good at precision with those good at recall to get a better balance. This tradeoff depends on the problem's needs.

What Good vs Bad Metrics Look Like for Stacking and Blending

Good stacking/blending results show:

• Higher accuracy or F1 score than any single model alone.
• Balanced precision and recall suitable for the task.
• Stable performance on new data (not just training data).

Bad results show:

• No improvement or worse metrics compared to best single model.
• Overfitting signs: very high training accuracy but low test accuracy.
• Unbalanced precision or recall causing practical problems.

Common Metrics Pitfalls in Stacking and Blending

• Data leakage: Using test data in training the stacking model inflates metrics falsely.
• Overfitting: The meta-model may memorize training data, showing high training but poor test metrics.
• Ignoring metric tradeoffs: Focusing only on accuracy can hide poor recall or precision.
• Confusion matrix mismatch: Not verifying that TP, FP, TN, FN add up correctly can cause wrong metric calculations.

Self-Check Question

Your stacking model has 98% accuracy but only 12% recall on fraud detection. Is it good for production? Why or why not?

Answer: No, it is not good. High accuracy can be misleading if the fraud class is rare. The very low recall means the model misses most fraud cases, which is dangerous. For fraud detection, high recall is critical to catch as many frauds as possible.

Key Result

Stacking and blending improve model performance by balancing key metrics like precision and recall, but careful evaluation is needed to avoid overfitting and data leakage.

Practice

1. What is the main goal of stacking and blending in machine learning?

easy

A. To combine multiple models to improve prediction accuracy

B. To reduce the size of the dataset

C. To speed up training by using fewer models

D. To replace all base models with a single model

Solution

1. Step 1: Understand the purpose of stacking and blending

   Stacking and blending are ensemble techniques that combine predictions from multiple models.

2. Step 2: Identify the goal of combining models

   The goal is to improve prediction accuracy by leveraging strengths of different models.

3. Final Answer:

   To combine multiple models to improve prediction accuracy -> Option A

4. Quick Check:

   Stacking and blending = combine models for better accuracy [OK]

Hint: Stacking and blending combine models to boost accuracy [OK]

Common Mistakes:

• Thinking stacking reduces dataset size
• Believing stacking replaces base models
• Confusing speed with accuracy improvement

2. Which of the following correctly describes how stacking trains its final model?

easy

A. Using random subsets of features

B. Using cross-validation predictions from base models

C. Using a separate holdout set only

D. Using the entire training data without splitting

Solution

1. Step 1: Recall stacking training method

   Stacking trains the final model on predictions generated by base models using cross-validation.

2. Step 2: Compare options to stacking method

   Only Using cross-validation predictions from base models mentions cross-validation predictions, which is key to stacking.

3. Final Answer:

   Using cross-validation predictions from base models -> Option B

4. Quick Check:

   Stacking uses cross-validation predictions [OK]

Hint: Stacking uses cross-validation predictions for final model [OK]

Common Mistakes:

• Confusing stacking with blending's holdout set
• Thinking stacking uses entire data without splits
• Assuming random feature subsets are used

3. Given the following code snippet for blending, what will be the shape of X_blend_train if X_train has shape (1000, 10) and holdout_ratio=0.2?

from sklearn.model_selection import train_test_split
X_train_full, X_holdout, y_train_full, y_holdout = train_test_split(X_train, y_train, test_size=holdout_ratio, random_state=42)
# Base model predictions on holdout
base_pred_holdout = base_model.predict(X_holdout)
# Blending training data
X_blend_train = base_pred_holdout.reshape(-1, 1)

medium

A. (200, 1)

B. (800, 1)

C. (1000, 1)

D. (200, 10)

Solution

1. Step 1: Calculate holdout set size

   With 1000 samples and 0.2 holdout ratio, holdout size = 1000 * 0.2 = 200 samples.

2. Step 2: Determine shape of base model predictions

   Base model predicts on holdout set, so predictions have shape (200,). Reshaping to (-1, 1) makes it (200, 1).

3. Final Answer:

   (200, 1) -> Option A

4. Quick Check:

   Holdout size 200, reshape to (200,1) [OK]

Hint: Holdout size = total * ratio; reshape predictions accordingly [OK]

Common Mistakes:

• Using full training size instead of holdout size
• Confusing reshape dimensions
• Assuming predictions keep original feature count

4. You wrote this stacking code but get an error: ValueError: Found input variables with inconsistent numbers of samples. What is the likely cause?

from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_predict

base1 = LogisticRegression()
base2 = RandomForestClassifier()

pred1 = cross_val_predict(base1, X_train, y_train, cv=5)
pred2 = cross_val_predict(base2, X_train, y_train, cv=5)

X_meta = np.column_stack((pred1, pred2))
meta_model = LogisticRegression()
meta_model.fit(X_meta, y_train)

medium

A. Meta model cannot be logistic regression

B. Base models are not fitted before predictions

C. Using cross_val_predict with cv=5 is invalid

D. Base model predictions have different lengths than y_train

Solution

1. Step 1: Understand cross_val_predict output

   cross_val_predict returns predictions for each sample in X_train, so pred1 and pred2 should have length equal to X_train.

2. Step 2: Identify cause of inconsistent sample sizes

   If pred1 or pred2 have different lengths than y_train, stacking fails due to mismatch in input sizes.

3. Final Answer:

   Base model predictions have different lengths than y_train -> Option D

4. Quick Check:

   Prediction length mismatch causes ValueError [OK]

Hint: Check prediction and label lengths match before stacking [OK]

Common Mistakes:

• Assuming models must be pre-fitted before cross_val_predict
• Thinking cv=5 is invalid for cross_val_predict
• Believing meta model type causes this error

5. You want to blend three base models using a holdout set. Which approach correctly prepares the training data for the blender model?

hard

A. Train blender on base model predictions from full training data without holdout

B. Train base models on holdout set, predict on full training data, then train blender on full predictions

C. Train base models on full training data, predict on holdout, then train blender on holdout predictions

D. Train blender on random subsets of base model predictions without holdout or cross-validation

Solution

1. Step 1: Understand blending process

   Blending trains base models on full training data, then uses their predictions on a separate holdout set to train the blender model.

2. Step 2: Evaluate options against blending steps

   Only Train base models on full training data, predict on holdout, then train blender on holdout predictions correctly describes training base models on full data, predicting on holdout, and training blender on those predictions.

3. Final Answer:

   Train base models on full training data, predict on holdout, then train blender on holdout predictions -> Option C

4. Quick Check:

   Blending uses holdout predictions for blender training [OK]

Hint: Blending trains blender on holdout predictions from full-trained base models [OK]

Common Mistakes:

• Training base models on holdout instead of full data
• Training blender without holdout predictions
• Ignoring holdout set in blending