Interaction features are derived features in machine learning models that capture the combined effect of two or more variables, often leading to improved prediction accuracy by revealing complex relationships. They are particularly useful in models where the relationship between variables is not purely additive, such as in non-linear or high-dimensional data contexts.