Neural Additive Models (NAM)

The model's prediction is the sum of univariate shape functions applied to each input feature, plus a bias term. This enforces the form:

g(E[y]) = f₁(x₁) + f₂(x₂) + ... + fₙ(xₙ) + β

• Each fᵢ is a neural network trained on a single feature
• No feature interactions are learned, making the contribution of each input completely independent
• The final prediction is a simple summation, enabling exact decomposition of the output

Each feature is processed by its own dedicated feedforward neural network that learns a non-linear transformation of that single input.

• A shape function fᵢ(xᵢ) can capture complex, jagged relationships that linear models miss
• The architecture typically uses ExU (Exp-centered) units in the first layer to better model sharp jumps common in tabular data
• Each network outputs a scalar contribution, which is added to the total prediction
• The learned function can be plotted directly as a graph of feature value vs. contribution

NAMs are intrinsically interpretable glass-box models, not black-boxes requiring post-hoc explanation.

• The contribution of any feature to a specific prediction is simply fᵢ(xᵢ) — a single, exact value
• No approximation methods like SHAP or LIME are needed; the explanation is the model itself
• Auditors can inspect every shape function to understand the model's global behavior across the entire feature range
• This satisfies regulatory requirements for adverse action reason codes and model documentation

The standard NAM can be extended to include learned pairwise feature interactions, creating a Neural Additive Model plus Interactions (NA²M).

• Adds terms of the form fᵢⱼ(xᵢ, xⱼ) for selected feature pairs
• Each interaction is a small neural network trained on two features
• The model remains interpretable because each interaction can be visualized as a heatmap
• Balances the accuracy gains of modeling interactions against the cost of added complexity

NAMs achieve a favorable position on the Pareto frontier between predictive performance and explainability.

• On tabular benchmarks, NAMs often match or approach the accuracy of gradient-boosted trees like XGBoost
• They consistently outperform logistic regression and other linear models on complex, non-linear problems
• The accuracy gap to fully-connected deep networks is typically small, while the interpretability gain is massive
• For fraud detection, this means catching sophisticated patterns while still providing auditable reason codes

Shape functions can be constrained to enforce desirable real-world properties that align with domain knowledge.

• Feature dropout randomly drops entire features during training, preventing over-reliance on any single input
• Smoothness penalties on the output of shape networks prevent erratic, high-frequency learned functions
• Monotonicity constraints can be imposed to ensure that, for example, a higher transaction amount never decreases risk
• These constraints produce shape functions that are not only accurate but also intuitively reasonable to human reviewers

The fundamental building block of a NAM. Each input feature is passed through its own independent neural network, called a shape function, which learns a non-linear transformation. The final prediction is the sum of these shape function outputs.

• Interpretability: Each shape function can be plotted as a graph showing the feature's effect on the prediction.
• Additivity: The model assumes no interactions between features, making the contribution of each feature completely modular and auditable.

A glass-box model that is a direct predecessor and close relative of NAMs. EBM uses cyclic gradient boosting to learn a set of interpretable feature functions, one per input.

• Key Difference: EBM also learns pairwise interaction terms between features, which NAMs typically omit for pure additivity.
• Use Case: Provides exact, modular explanations and is often used as a high-performance, interpretable benchmark for fraud models.

A global, model-agnostic visualization tool that shows the marginal effect of one or two features on a model's predicted outcome.

• How it works: Averages the model's predictions over the distribution of all other features to isolate the effect of the target feature.
• NAM Connection: The shape function plot of a NAM is essentially a perfect, exact PDP for that feature, free from the averaging bias that plagues PDPs on correlated features in black-box models.

An unbiased alternative to PDPs that correctly handles correlated features. ALE calculates the local differences in prediction when a feature value changes within a small interval, then accumulates these effects.

• Advantage over PDP: Avoids extrapolating into unrealistic regions of the feature space where data doesn't exist.
• Relevance: While NAMs provide exact feature effects, ALE is the go-to method for interpreting black-box models where features are correlated, such as transaction velocity and amount.

Concise, human-readable statements that provide the primary reasons for a model's specific decision, such as flagging a transaction as fraudulent.

• NAM Advantage: Because a NAM's prediction is a sum of independent feature effects, the top contributing features for any single prediction can be instantly identified and ranked to generate a reason code.
• Regulatory Need: These are often mandated as Adverse Action Reason Codes to explain automated decisions to customers and auditors.

A property of models that are inherently simple and understandable due to their structure, providing direct insight without needing post-hoc analysis.

• Examples: Linear regression, a single decision tree, Explainable Boosting Machines, and Neural Additive Models.
• Contrast: This is the opposite of post-hoc explainability, where complex methods like SHAP or LIME are applied to a black-box model after training to approximate its behavior.