Identification of Key Feature Interactions via PDP Decomposition

Understanding feature interactions is essential for interpreting complex machine learning models. Global interpretation methods, such as Partial Dependence Plots (PDPs), are commonly used to visualize the marginal effects of features on model predictions. However, PDPs average feature effects across all other features, which can obscure critical interaction patterns and fail to identify important features influenced by these interactions. While Individual Conditional Expectation plots reveal variations in a feature’s effects across individual data points, they do not provide insights into the specific interactions causing these differences. To address these limitations, we propose a method that combines functional decomposition with PDP analysis, enabling the isolation and interpretation of feature interactions. High variance indicates significant interaction effects, while low variance suggests a constant contribution to the prediction. We evaluate this approach on synthetic and real-world datasets, showing that it effectively identifies and interprets feature interactions, offering deeper insights into model behavior.