Explainable Boosting Machine (EBM)

EBMs are an implementation of a Generalized Additive Model (GAM) with modern machine learning enhancements. The core equation is g(E[y]) = Σ f_i(x_i), where each feature function f_i is a non-linear shape learned independently. This additive structure prevents complex feature interactions from obscuring the model's logic, ensuring that the contribution of a single feature can be isolated and plotted as a graph. This directly contrasts with deep neural networks, where features are entangled across multiple layers.

While standard GAMs are purely additive, EBMs can optionally include pairwise interaction terms (f_ij(x_i, x_j)) to boost accuracy without sacrificing full interpretability. The model automatically detects and ranks important interactions using a smart, stage-wise selection process. These interactions are visualized as heatmaps, allowing analysts to inspect exactly how two features combine to influence a prediction. This bridges the gap between the simplicity of linear models and the complexity of gradient-boosted trees.

The defining characteristic of an EBM is the ability to visualize every feature's learned shape function. Each graph shows the exact contribution of a feature value to the final prediction, making the model's logic transparent. For example, in a credit risk model, you can see the precise score penalty applied as an applicant's debt-to-income ratio increases. This allows domain experts to validate the model's reasoning against established business rules and regulatory requirements before deployment.

EBMs train using a combination of bagging (bootstrap aggregation) and gradient boosting. The model builds a large ensemble of shallow trees for each feature in a round-robin fashion, using a very low learning rate. This technique dramatically reduces overfitting and variance, resulting in a model that is both highly accurate and robust to noise. The final shape function for a feature is the average of all the individual trees learned for that feature, providing a smooth, stable representation.

EBMs natively support both binary classification and regression tasks through different link functions (g). For classification, a logit link function is used, and the learned shape functions output values in log-odds space, which are then passed through a logistic function. For regression, an identity link is used, and the shape functions directly model the target value. This flexibility makes EBMs a drop-in replacement for black-box models like XGBoost or deep neural networks in a wide range of enterprise applications.

EBMs are a glass-box model, meaning their structure is the explanation. A global understanding is gained by viewing the entire set of feature shape functions. A local explanation for a single prediction is generated by simply looking up the score contribution of each feature value on its respective graph and summing them. This eliminates the need for post-hoc approximation tools like LIME or SHAP, providing exact, faithful, and instantaneous explanations for every prediction without any additional computation.

The statistical foundation upon which EBMs are built. A GAM represents the prediction as a sum of smooth functions of individual features: g(E[y]) = f₁(x₁) + f₂(x₂) + ... + fₚ(xₚ). Unlike linear models, each shape function can capture non-linear relationships while remaining fully separable and inspectable. Traditional GAMs use splines or local regression, whereas EBMs use gradient-boosted bagged trees to learn these functions, dramatically improving accuracy without sacrificing interpretability.

A deep learning counterpart to EBMs that learns shape functions using neural networks instead of boosted trees. Each feature gets its own small feedforward network, and predictions are summed linearly. Key properties:

• Visualizable: Each feature's network output can be plotted as a shape function
• Regularization: ExU (exp-centered) units help capture jagged, high-frequency patterns
• Trade-off: NAMs can model more complex per-feature patterns than EBMs but lose the automatic interaction detection that EBMs provide through pairwise terms

EBMs belong to the glass-box category—models whose internal logic is fully transparent and auditable by design. This contrasts with black-box models like deep neural networks or XGBoost ensembles, which require post-hoc explanation tools (SHAP, LIME) to approximate their reasoning.

Glass-box properties:

• Direct inspection: Every feature contribution is visible without approximation
• Editability: Shape functions can be manually adjusted by domain experts
• No explanation fidelity gap: The explanation is the model, eliminating the risk of misleading post-hoc rationalizations

A distinctive capability of EBMs that automatically identifies and models two-way feature interactions while maintaining interpretability. The algorithm:

1. Trains main effects for all features independently
2. Scans residual predictions to detect pairs of features with remaining signal
3. Fits low-rank interaction terms fᵢⱼ(xᵢ, xⱼ) only for significant pairs
4. Renders these interactions as heatmaps for visual inspection

This balances the additive structure's simplicity with the reality that some features genuinely interact, without resorting to opaque higher-order combinations.

The formal name for EBMs that include pairwise interaction terms. GA²M extends the standard GAM formula to: g(E[y]) = Σ fᵢ(xᵢ) + Σ fᵢⱼ(xᵢ, xⱼ). This architecture was proven by Yin Lou et al. (2013) to achieve accuracy competitive with the best black-box models on structured data while remaining fully intelligible. The key insight: most real-world patterns can be captured by individual features and their first-order interactions, making higher-order terms unnecessary for strong performance.

The primary interpretability mechanism of EBMs. Each feature's learned function f(x) is plotted as a curve showing exactly how the model's prediction changes as that feature varies, holding others constant. Key characteristics:

• Non-linear patterns: Thresholds, saturation points, and U-shaped relationships are directly visible
• Per-feature scoring: The contribution of any feature value can be read directly from the plot
• Auditability: Regulators can verify that protected attributes (race, gender) have zero contribution
• Comparison: Unlike PDPs, these plots show the actual model logic, not an approximation