Table of Contents

Enum SplitCriterion

Namespace
AiDotNet.Enums
Assembly
AiDotNet.dll

Specifies the criterion used to determine the best way to split data in decision trees and other tree-based models.

public enum SplitCriterion

Fields

FriedmanMSE = 3

Selects splits using Friedman's improvement to MSE, which accounts for the potential improvement from further splits.

For Beginners: This is an advanced version of MSE that looks ahead to future possibilities.

Named after statistician Jerome Friedman, this criterion:

  • Uses the standard MSE calculation but with an additional factor
  • Considers not just the current split but potential future improvements
  • Often leads to better tree structures in gradient boosting models

Think of it like a chess player who doesn't just consider the current move, but also thinks about how that move sets up future opportunities.

This criterion is particularly useful in gradient boosting decision trees (like XGBoost or LightGBM) and can lead to better overall model performance.

MeanAbsoluteError = 2

Selects splits that minimize the mean absolute error between the actual and predicted values.

For Beginners: This criterion focuses on reducing the absolute (positive) difference between predictions and actual values.

Mean Absolute Error (MAE) works by:

  1. Taking the difference between each actual value and the predicted value
  2. Converting each difference to a positive number (absolute value)
  3. Calculating the average of these absolute differences

This criterion:

  • Treats all sizes of errors more equally than MSE
  • Is more robust to outliers (unusual values)
  • May be better when your data has some extreme values

Think of it like measuring the average distance between your guesses and the correct answers, regardless of whether you guessed too high or too low.

MeanSquaredError = 1

Selects splits that minimize the mean squared error between the actual and predicted values.

For Beginners: This criterion focuses on reducing the squared difference between predictions and actual values.

Mean Squared Error (MSE) works by:

  1. Taking the difference between each actual value and the predicted value
  2. Squaring each difference (to make negatives positive and emphasize large errors)
  3. Calculating the average of these squared differences

This criterion:

  • Penalizes large errors more heavily than small ones
  • Is sensitive to outliers (unusual values far from the average)
  • Often produces similar results to VarianceReduction

Think of it like trying to minimize the "penalty points" you get when your guess is wrong, with bigger mistakes costing exponentially more points.

VarianceReduction = 0

Selects splits that maximize the reduction in variance of the target variable.

For Beginners: This criterion tries to group similar values together.

Variance measures how spread out numbers are. High variance means the numbers are all over the place, while low variance means they're clustered together.

Example: If we have house prices [100K, 105K, 110K] in one group and [500K, 510K, 490K] in another, each group has low variance (the prices are similar within each group).

This criterion:

  • Tries to create groups where the values within each group are as similar as possible
  • Is the traditional approach for regression trees
  • Works well for most regression problems

Think of it like sorting books by height, so each shelf has books of similar height.

Remarks

For Beginners: Split criteria are like the "rules" that help decision trees decide how to divide data.

Imagine you're organizing books on shelves. You could sort them by:

  • Size (big books vs. small books)
  • Color (red books vs. blue books)
  • Topic (fiction vs. non-fiction)

But which way of sorting is best? Split criteria help the AI decide which way of dividing the data will lead to the most accurate predictions. Different criteria measure "best" in different ways, each with their own advantages.

These criteria are primarily used for regression problems (predicting numeric values like house prices or temperatures) rather than classification problems (predicting categories).

Edit this page