Accumulated Local Effects (ALE) Plots

A global, model-agnostic visualization showing the marginal effect of one or two features on the predicted outcome, averaged over the distribution of all other features. Unlike ALE plots, PDPs assume feature independence, which can produce misleading results when predictors are highly correlated—a common scenario in financial transaction data where variables like transaction amount and account balance move together. The averaging process can also obscure heterogeneous effects across different subpopulations of cardholders.

A disaggregated visualization that plots the relationship between a feature and the model's prediction for each individual instance, revealing heterogeneous relationships hidden by PDP averaging.

• Each line represents one transaction or account
• Divergent slopes indicate interaction effects with other features
• Clustered patterns can identify distinct behavioral segments
• Essential for fraud models where a feature like transaction velocity may affect high-net-worth and low-balance accounts differently

A game-theoretic framework that assigns each feature an importance value for a particular prediction, unifying several existing feature attribution methods. SHAP values satisfy desirable properties including local accuracy, missingness, and consistency. For fraud detection, SHAP provides per-transaction explanations that answer: 'Why was this specific payment flagged?'—critical for generating adverse action reason codes required by regulators when blocking a customer's transaction.

A quantitative measure of the degree to which the effect of one feature on a model's prediction depends on the value of another feature. The Friedman's H-statistic uses partial dependence decomposition to identify non-additive behavior.

• Value of 0 indicates no interaction
• Value of 1 indicates pure interaction with no additive component
• Critical for fraud models where interactions—like high amount + unusual merchant category—are the primary fraud signal
• Pairs naturally with ALE plots to validate that accumulated local effects correctly capture interaction structure

A model inspection technique that measures the decrease in a model's performance score when the values of a single feature are randomly shuffled, breaking the relationship between the feature and the true outcome. Unlike ALE plots which show the direction and shape of a feature's effect, permutation importance quantifies the magnitude of a feature's contribution to predictive accuracy. When used together, these methods provide complementary views: ALE reveals how a feature influences predictions, while permutation importance confirms whether that influence is actually useful.

A local explanation method that identifies the minimal changes to an input instance's features that would alter the prediction to a predefined, desired outcome. For a flagged fraudulent transaction, a counterfactual might state: 'If the transaction amount were below $500 and the merchant category were grocery instead of electronics, this transaction would not have been blocked.' This approach is highly intuitive for compliance officers and investigators who need to understand decision boundaries, and it complements ALE's global perspective with actionable, instance-specific guidance.