Information Gain and Mutual Information for Machine Learning

Information Gain and Mutual Information are used to measure how much knowledge one variable provides about another. They help optimize feature selection, split decision boundaries and improve model accuracy by reducing uncertainty in predictions.

Need for Information-Based Metrics

Some common reasons why these measures are essential are:

• Identify attributes that most effectively reduce uncertainty in target predictions.
• Improve decision boundaries in tree-based models by choosing meaningful splits.
• Support feature selection in high-dimensional datasets such as text or genomics.
• Enhance model interpretability by showing which features influence outcomes most.
• Prevent overfitting by avoiding noisy or irrelevant feature usage.

Information Gain

Information Gain quantifies the reduction in entropy after splitting the data on a particular feature. Higher gain means a more useful split.

Formula

Where:

• H(S) is the entropy before the split,
• H(S_v) is the entropy of each subset after splitting.

Properties

• Prefers attributes that produce pure child nodes.
• Works efficiently in decision tree algorithms like ID3 and C4.5.
• Slightly biased toward multi-valued attributes.

Implementation

Calculating information gain using iris dataset.

Output:

Information Gain for each feature: [0.50576572 0.27875048 0.98425371 0.99358193]

Petal length and petal width are highly informative features compared to sepal length and sepal width for predicting the species of Iris flowers.

Mutual Information

Mutual Information measures how strongly two variables depend on each other. It works well for both linear and non-linear relationships.

Formula

Where:

• p(x, y) is the joint probability,
• p(x) and p(y) are marginal probabilities.

Properties

• Captures any form of dependency, not just linear.
• Stable in high-dimensional feature spaces.
• Robust against noise compared to correlation alone.

Implementation

Implementing Mutual Information using mutual_info_regression and generating data synthetically.

Output:

Mutual Information for each feature: [0.42283584 0.54090791]

Higher Mutual Information values suggest a stronger relationship or dependency between the features and the target variable.

Information Gain Ratio

Information Gain Ratio normalizes Information Gain to reduce bias toward attributes with many values.

Formula

Where:

• IG(S,A) is the Information Gain from splitting on attribute
• H(A) is the Entropy of attribute A, indicating how widely it splits the data.
• IGR is Normalized value to avoid bias toward attributes with many distinct values.

Properties

• Reduces multi-value splitting bias.
• Used in C4.5 algorithm for balanced decisions.
• Improves generalization across unseen data.

Implementation

Calculating IGR.

Output:

IGR: 0.11232501392736335

An IGR of 0.1123 is considered low, the feature is not very informative for splitting the data compared to other possible features.

Information Gain Vs Mutual Information

Difference between information gain and mutual information: