Table of Contents

Enum RegularizationType

Namespace
AiDotNet.Enums
Assembly
AiDotNet.dll

Specifies the type of regularization to apply to a machine learning model.

public enum RegularizationType

Fields

ElasticNet = 3

A combination of L1 and L2 regularization that balances their properties.

For Beginners: ElasticNet combines the best of both L1 and L2 regularization.

It works by applying both types of penalties at the same time, with adjustable weights to control how much of each to use.

Think of it like a flexible teacher who says: "Let's mostly keep all features but with limited influence, while still completely removing the least useful ones."

Benefits:

  • Can eliminate irrelevant features (like L1)
  • Handles groups of correlated features well (like L2)
  • Provides more flexibility through adjustable balance between L1 and L2

Example: For house price prediction, ElasticNet might eliminate 30 truly irrelevant features while keeping the remaining 70 with appropriately controlled influence.

This is often the best choice when you're not sure which regularization to use.

L1 = 1

L1 regularization (also known as Lasso regularization) that encourages sparsity in the model parameters.

For Beginners: L1 regularization encourages the model to completely ignore less important features.

It works by penalizing the absolute size of the model's parameters, which often results in many parameters becoming exactly zero.

Think of it like a strict teacher who says: "If a feature isn't clearly helpful, don't use it at all."

Benefits:

  • Automatically selects the most important features
  • Creates simpler models that are easier to interpret
  • Works well when you suspect many features aren't relevant

Example: If you're predicting house prices with 100 features, L1 might decide that only 20 features (like size, location, and age) actually matter and ignore the rest.

L2 = 2

L2 regularization (also known as Ridge regularization) that discourages large parameter values.

For Beginners: L2 regularization encourages the model to use all features, but keep their influence small.

It works by penalizing the squared size of the model's parameters, which results in all parameters becoming smaller but rarely exactly zero.

Think of it like a balanced teacher who says: "Use all the information available, but don't rely too much on any single piece."

Benefits:

  • Handles correlated features well
  • Generally prevents overfitting without eliminating features
  • Usually the safest default choice for regularization

Example: For house price prediction, L2 might keep all 100 features but ensure that no single feature (like having a swimming pool) has an excessively large impact on the prediction.

None = 0

No regularization is applied to the model.

For Beginners: This option turns off regularization completely.

Use this when:

  • You have lots of training data compared to model complexity
  • Your model is already simple and unlikely to overfit
  • You want to see how the model performs without any restrictions

It's like removing the training wheels - sometimes it works fine, but there's a higher risk the model might become too specialized to your training data.

Remarks

For Beginners: Regularization is like adding training wheels to your AI model.

When models learn too much from their training data, they might become too specialized (this is called "overfitting"). Regularization helps prevent this by encouraging the model to keep things simple.

Think of it like this:

  • Without regularization: The model might create very complex rules that work perfectly for training data but fail on new data.
  • With regularization: The model is encouraged to create simpler rules that work well enough for training data and are more likely to work on new data too.

Different regularization types use different approaches to encourage simplicity.

Edit this page