Explainable Boosting Machine (EBM)

EBMs are modern implementations of Generalized Additive Models (GAMs) where the prediction is the sum of individual feature functions:

• Each feature x_i learns a shape function f_i(x_i)
• The final prediction is g(E[y]) = f_1(x_1) + f_2(x_2) + ... + f_n(x_n)
• This additive structure means no complex feature interactions are hidden
• Every feature's contribution is isolated and independently auditable

For fraud detection, this allows analysts to see exactly how much a transaction amount or time-of-day contributed to a risk score.

While EBMs are additive by default, they can explicitly model specified pairwise interactions without sacrificing interpretability:

• The model learns f_ij(x_i, x_j) for selected feature pairs
• These interactions are rendered as heatmap visualizations
• Analysts can inspect exactly how two features combine to influence predictions
• Unlike black-box models, interactions are opt-in and transparent

Example: In financial fraud, an EBM can show how transaction_amount and merchant_category together affect risk, with the full interaction surface available for audit.

EBMs use a specialized training procedure that ensures both accuracy and stability:

• Features are trained in a round-robin, cyclic fashion
• Each round updates one feature's shape function using gradient boosting
• The algorithm incorporates bagging (bootstrap aggregation) to reduce variance
• Training continues until convergence across all features

This cyclic approach prevents any single feature from dominating early training rounds, producing well-calibrated, stable shape functions that generalize reliably to new fraud patterns.

Every feature in an EBM produces a fully inspectable score graph:

• Each graph shows the exact contribution of a feature value to the prediction
• Graphs reveal non-linear relationships that linear models would miss
• Thresholds, step changes, and saturation points are immediately visible
• Compliance teams can validate that features behave monotonically or as expected

For a fraud model, the account_age graph might show risk decreasing sharply in the first 30 days, then plateauing — a pattern directly visible to regulators without any approximation.

EBMs handle both binary classification and regression tasks natively:

• Classification: Uses logit link function for fraud/no-fraud decisions
• Regression: Uses identity or log link for continuous risk scoring
• Supports multi-class classification for categorizing fraud types
• Outputs are calibrated probabilities suitable for risk thresholds

This flexibility means the same glass-box architecture can power both a real-time fraud classifier and a continuous money laundering risk score, with identical interpretability guarantees across use cases.

EBMs provide both global model understanding and local prediction explanations without separate tooling:

• Global: Feature score graphs show overall model behavior across the entire dataset
• Local: For any single transaction, the contribution of each feature is the exact value read from its graph
• No approximations, sampling, or surrogate models are needed
• Reason codes are simply the top contributing features for that prediction

This eliminates the fidelity gap between global explanations and local justifications that plagues post-hoc methods like LIME or SHAP when applied to black-box models.

The foundational statistical framework upon which EBM is built. A GAM makes predictions by summing the output of independent shape functions, one for each feature. The formula is g(E[y]) = Σ f_i(x_i), where each f_i is a smooth, interpretable curve. Unlike traditional GAMs that use splines, EBM uses bagged and boosted trees to learn each f_i, automatically capturing complex non-linear relationships and pairwise interactions while preserving full modular additivity.

A deep learning counterpart to EBM that also learns a linear combination of independent shape functions. NAMs use exponential-centered neural networks as the shape functions instead of boosted trees. Key distinctions:

• NAMs are trained with stochastic gradient descent, while EBMs use gradient boosting
• EBMs automatically detect and include pairwise interaction terms (f_ij(x_i, x_j)); standard NAMs do not
• EBMs provide exact intelligibility; NAMs require the shape function network to be visualized and interpreted

A property of models that are transparent by design, not explained after the fact. EBM is a prime example of an intrinsically interpretable model because its additive structure allows every component to be inspected directly. This contrasts with post-hoc explainability methods like SHAP or LIME, which approximate the behavior of an opaque model. For regulated fraud detection, intrinsic interpretability provides an unimpeachable audit trail because the explanation is the model's exact computation, not a surrogate approximation.

A critical capability of EBM that automatically identifies and models non-additive effects between pairs of features. While the base model learns independent shape functions f_i(x_i), EBM can also learn 2D interaction functions f_ij(x_i, x_j) for high-signal pairs. For fraud detection, this captures critical logic like:

• High transaction amount and unusual merchant category code
• Rapid session velocity and new device fingerprint These interactions are rendered as heatmap visualizations, remaining fully interpretable.

The human-readable output of an EBM prediction that lists the top contributing features and their specific impact on a decision. For a flagged transaction, an EBM generates exact reason codes by simply ranking the additive contributions from each shape function. Example output:

• Transaction Amount: +0.42 (high risk)
• Time Since Last Login: +0.31 (high risk)
• Device Reputation Score: -0.55 (low risk) This satisfies Adverse Action Reason Code requirements under regulations like FCRA and ECOA.