Generalized Groves of Neural Additive Models - Pursuing Transparent and Accurate Machine Learning Models in Finance

Aug 3, 2026

Upload by Oliver

Machine learning improves predictive performance but often acts as a black box, making interpretation difficult. In highly regulated finance, transparency is as critical as accuracy, since adoption depends on understanding model behavior. This work introduces generalized groves of neural additive models, which partition features into linear terms, individual nonlinear terms, and locally interacted nonlinear groups. A stepwise selection separates linear from nonlinear components, while verified additive separation criteria ensure controlled interaction structure, yielding models that remain accurate yet largely interpretable.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

AccessFree to download

Introduction

Explainability and transparency in regulated finance

Two explainability directions

Transparent neural additive model framework

Problem motivation and design goals

Generalized Groves of Neural Additive Models: Pursuing Transparent Machine Learning Models in Finance Dangxing Chen 􀀃 and Weicheng Ye Zu Chongzhi Center for Mathematics and Computational Sciences Duke Kunshan University, Kunshan, Jiangsu, China Email: [dangxing.chen@dukekunshan.edu.cn and weicheng.ye@dukekunshan.edu.cn](dangxing.chen@dukekunshan.edu.cn and weicheng.ye@dukekunshan.edu.cn) arXiv :2209 . 10082v2 [ cs .LG] 29 Jul 2024 Abstract—While machine learning methods have signi􀀂cantly improved model performance over traditional methods, their black-box structure makes it dif􀀂cult for researchers to interpret results. For highly regulated 􀀂nancial industries, model transparency is equally important to accuracy. Without understanding how models work, even highly accurate machine learning methods are unlikely to be accepted. We address this issue by introducing a novel class of transparent machine learning models known as generalized groves of neural additive models. The generalized groves of neural additive models separate features into three categories: linear features, individual nonlinear features, and interacted nonlinear features. Additionally, interactions in the last category are only local. A stepwise selection algorithm distinguishes the linear and nonlinear components, and interacted groups are carefully veri􀀂ed by applying additive separation criteria. Through some empirical examples in 􀀂nance, we demonstrate that generalized grove of neural additive models exhibit high accuracy and transparency with predominantly linear terms and only sparse nonlinear ones. Index Terms—Neural Network, Transparency, Interpretability I. INTRODUCTION Machine learning (ML) models have been proven extremely successful in analyzing complex and high-dimensional datasets, with improved accuracy over traditional methods, such as linear and logistic regressions (LaLRs) . On the other hand, there has been an increase in public concern about the use of ML methods without enhanced regulation. As of April 2021, the European Commission (EC) has proposed the Arti􀀂cial Intelligence Act (AIA) [1], which marks a historic 􀀂rst step towards 􀀂lling the regulatory gap. Additionally, thereview article [2] explains why regulators are obliged to require ML methods to be transparent and explainable. In the highly regulated 􀀂nancial sector, transparency and explainability are equally important to the model accuracy. In the handbook on model risk management of the US Of􀀂ce of the Comptroller of the Currency (OCC) published in August 2021, it stressed the importance of evaluation transparency and explainability for risk management when using complex models [3] . More recently in May 2022, the Consumer Financial Protection Bureau (CFPB) con􀀂rmed that anti-discrimination laws require institutions to provide a detailed explanation 􀀃 Corresponding author. to consumers when denying a credit application using ML methods [4] . Researchers are investigating explainable ML tools in light of the growing regulatory requirements [5]–[7] . Speci􀀂cally, two directions have been extensively explored by researchers in order to provide explainability. In the 􀀂rst direction, model-agnostic approaches are provided to disentangle a trained black-box model. Several popular methods have been developed including locally interpretable modelagnostic explanations (LIME) [8], SHapley Additive Explanations (SHAP) [9], and sensitivity-based analysis [10] . Despite these successes, it is important to note that ML methods maybe intrinsically opaque, as opposed to LaLRs. Therefore, while such explainability may meet explanation requirements for applications in 􀀂elds of text and image analysis, they may not be adequate enough in 􀀂nancial applications. Furthermore, universal explanations do not exist, and there has been criticism of blindly adopting them [6], [11]–[13] . In the second approach, it simpli􀀂es model architecture by enhancing its transparency, see [5],

FAQs

Why is transparency especially important for machine learning in finance?

Because regulators and institutions require explanations that users can understand, and adoption depends on being able to interpret how models work in addition to achieving high accuracy.

How do generalized groves of neural additive models structure features?

They separate features into three categories: linear features, individual nonlinear features, and interacted nonlinear features, where interactions are restricted to be local.

What method is used to obtain the linear and nonlinear components and verify interactions?

A forward stepwise selection algorithm distinguishes linear and nonlinear components, and interacted groups are carefully verified using additive separation criteria to keep interactions controlled.

Aug 3, 2026

Upload by Oliver