Feature Interaction Strength

A global, model-agnostic metric that measures the proportion of variance in a model's predictions explained by interactions between features. A value of 0 indicates a purely additive model with no interactions, while a value approaching 1 signifies that the prediction function is dominated by interaction effects. For a pair of features (j, k), the H-statistic quantifies how much of the joint effect cannot be explained by the sum of their individual partial dependence functions. In fraud detection, a high H-statistic between transaction_amount and time_since_last_transaction reveals that large transactions are only suspicious when they occur shortly after a previous one.

A diagnostic method that compares the two-dimensional partial dependence plot (PDP) of a feature pair against the sum of their individual one-dimensional PDPs. If the joint PDP surface is not flat—meaning it cannot be decomposed into additive components—an interaction is present. For example, the joint PDP of device_age and account_age might show a sharp spike in fraud probability only when both features are low, a multiplicative effect invisible in univariate analysis. This visual decomposition helps compliance officers understand why a model flags specific combinations of attributes.

An extension of SHAP (SHapley Additive exPlanations) that decomposes a prediction into main effects and pairwise interaction effects. The SHAP interaction value for features i and j represents the additional contribution that arises from their combination, beyond what each contributes independently. This is computed using the Shapley interaction index from cooperative game theory. In an explainable fraud detection workflow, these values can be visualized as a matrix, revealing that ip_country and shipping_country have a strong negative interaction when they match but a strong positive interaction when they diverge.

For differentiable models like neural networks, the interaction strength between two input features can be estimated by computing the second-order partial derivative of the model's output with respect to those features. A large, consistent second derivative indicates that the effect of one feature is highly dependent on the value of the other. This method is computationally efficient for deep learning fraud models and can be integrated directly into the training loop for real-time monitoring. Integrated Hessians extends this concept by accumulating second-order gradients along a path from a baseline to the input.

A statistical algorithm designed to detect interaction effects in regression and classification trees without the selection bias that plagues standard methods like CART. GUIDE uses chi-squared tests on residuals to determine if a feature's effect on the target is modified by another feature before making a split. This is particularly valuable in financial fraud modeling where categorical features like merchant_category_code and card_present_flag often interact. GUIDE provides unbiased, interpretable interaction rules that can be directly translated into investigation procedures.

Unlike PDPs, ALE plots correctly handle correlated features by calculating differences in predictions over conditional distributions. The second-order ALE plot visualizes the joint effect of two features after accounting for their individual main effects. If the second-order ALE surface shows significant variation, a strong interaction is present. This is crucial in fraud detection where features like transaction_frequency and average_transaction_value are often highly correlated, and standard PDP-based interaction measures would produce biased estimates due to extrapolation into unrealistic feature regions.