[Paper Review] Neural Additive Models: Interpretable Machine Learning with Neural Nets

Rishabh Agarwal, Melnick, Levi+5 more|arXiv (Cornell University)|Apr 29, 2020

Explainable Artificial Intelligence (XAI)Computer Science37 references121 citations

TL;DR

NAMs combine additive model interpretability with neural networks, achieving competitive accuracy to GAMs and strong performance versus intelligible baselines, while enabling multitask and differentiable, modular interpretations.

ABSTRACT

Deep neural networks (DNNs) are powerful black-box predictors that have achieved impressive performance on a wide variety of tasks. However, their accuracy comes at the cost of intelligibility: it is usually unclear how they make their decisions. This hinders their applicability to high stakes decision-making domains such as healthcare. We propose Neural Additive Models (NAMs) which combine some of the expressivity of DNNs with the inherent intelligibility of generalized additive models. NAMs learn a linear combination of neural networks that each attend to a single input feature. These networks are trained jointly and can learn arbitrarily complex relationships between their input feature and the output. Our experiments on regression and classification datasets show that NAMs are more accurate than widely used intelligible models such as logistic regression and shallow decision trees. They perform similarly to existing state-of-the-art generalized additive models in accuracy, but are more flexible because they are based on neural nets instead of boosted trees. To demonstrate this, we show how NAMs can be used for multitask learning on synthetic data and on the COMPAS recidivism data due to their composability, and demonstrate that the differentiability of NAMs allows them to train more complex interpretable models for COVID-19.

Motivation & Objective

• Motivate the need for interpretable high-performance models in high-stakes domains (healthcare, finance, justice).
• Propose Neural Additive Models (NAMs) that learn a linear combination of feature-specific neural networks to form a Generalized Additive Model (GAM).
• Show that NAMs achieve accuracy comparable to state-of-the-art GAMs and beat common intelligible baselines on tabular datasets.
• Demonstrate the extensibility of NAMs to multitask learning and differentiable, interpretable model components for applications like COVID-19 treatment modeling.

Proposed method

• NAMs learn a linear combination of neural networks, each attending to a single input feature, forming g(E[y]) = β + sum_i f_i(x_i) with f_i modeled by neural nets.
• Each f_i is trained jointly via backpropagation and can model arbitrarily complex shapes.
• ExU hidden units are introduced to learn jagged, jumpy 1D functions by computing h(x) = f(e^w * (x - b)).
• ReLU-_n activations and regularization (dropout, weight decay, etc.) are used to control overfitting and encourage realistic shape functions.
• NAMs are differentiable and modular, enabling multitask extension and composability across tasks.
• The method emphasizes visualizable shape functions f_i(x_i) as exact descriptions of how NAMs compute predictions.

Experimental results

Research questions

• RQ1 Can NAMs achieve competitive predictive accuracy while preserving interpretability for tabular data?
• RQ2 How do NAMs compare to logistic/linear models, decision trees, EBMs, XGBoost, and DNNs on standard datasets?
• RQ3 Can NAMs be extended to multitask learning while maintaining interpretability?
• RQ4 What is the role of ExU units in modeling jagged, real-world shape functions, and how should they be regularized?

Key findings

• NAMs achieve competitive accuracy, often outperforming logistic/linear regression and CART, and approaching state-of-the-art GAMs and EBMs on benchmark datasets.
• NAMs deliver interpretable, per-feature shape plots that exactly describe how predictions are computed, enabling clear bias detection and debiasing (e.g., race in COMPAS).
• ExU units enable NAMs to model jagged, jumpy functions more effectively than standard units; appropriate initialization and ReLU-n activations help capture sharp changes.
• Multitask NAMs improve predictive performance and allow task-specific shape functions, enabling sharing of subnets while preserving interpretability.
• On COMPAS, multitask NAMs reveal gender-specific relationships and offer a path to debiasing by inspecting and adjusting feature contributions while maintaining overall accuracy.
• NAMs are differentiable, enabling their use as parameter generators for personalized treatment policies (e.g., COVID-19) and integration into differentiable pipelines.