Stacking ModelsStacking Models

Stacking, also known as stacked generalization, is an ensemble learning technique that combines multiple base models (learners) with a meta-model to improve predictive performance. It is a more advanced form of ensemble learning than boosting and bagging.
Stacking aims to leverage the strengths of different base models and learn how to combine their predictions through the meta-model best.

How does stacking work?

Here's how stacking works:

  1. Base Models (Level 0 model): Several diverse base models are trained on the training data. These base models can be different machine learning algorithms or variations of the same algorithm with different hyperparameters. Each base model makes predictions on the test data.
  2. Meta-Model (Level 1 model): A meta-model, also known as a blender or a second-level model, is then trained using the predictions made by the base models as its input features. The meta-model learns to combine these predictions and generate the final ensemble prediction.
  3. Training and Validation: The training data is typically split into multiple folds (or subsets). The base models are trained on different folds, and the meta-model is trained on the predictions made by the base models on the remaining data (out-of-fold predictions). This helps to avoid overfitting during the stacking process.
  4. Prediction: After training the base models and the meta-model, the final prediction is made by passing the test data through the base models to obtain their predictions and then using these predictions as input for the meta-model to produce the ensemble's final prediction.


    Pay attention that in stacking ensembles, we have the ability to use multiple models together as base models. However, in other ensemble techniques, we are limited to using just one particular base model for training each ensemble.

You can see the principle of making predictions by the stacking ensemble in the image below:

By leveraging the strengths of different base models and learning how to optimally combine their predictions, stacking can lead to improved predictive performance compared to using any single model in isolation. However, it requires careful tuning and validation to prevent overfitting and achieve the best results.


Can you explain the concept of level 0 and level 1 models in stacking?

Select the correct answer

Everything was clear?

Section 1. Chapter 4