Generalized Groves of Neural Additive Models: Pursuing Transparent Machine Learning Models in Finance

Dangxing Chen, Weicheng Ye

TL;DR

This paper introduces Generalized Groves of Neural Additive Models (GGNAMs), a novel class of transparent machine learning models designed for the finance industry. GGNAMs achieve high accuracy comparable to complex black-box models while offering superior interpretability by explicitly separating features into linear, individual nonlinear, and locally interacted nonlinear components.

What Problem Does This Paper Address?

The finance industry, being highly regulated, demands transparency and explainability in its machine learning (ML) models, often as much as accuracy. Traditional ML models, while powerful, are frequently "black boxes," making it difficult to understand their decision-making processes. This lack of transparency hinders regulatory compliance and public trust, especially when models are used for critical decisions like credit scoring. Existing explainable AI (XAI) methods either try to explain black-box models post-hoc (which can be unreliable) or simplify model architectures, potentially sacrificing accuracy. This paper aims to bridge this gap by developing a model that is both highly accurate and inherently transparent, specifically tailored for financial applications.

Methods

The core contribution is the introduction of Generalized Groves of Neural Additive Models (GGNAMs). The GGNAM framework builds upon Neural Additive Models (NAMs) by incorporating several key innovations:

  • Feature Categorization: GGNAMs categorize input features into three types:
  • Linear Features: Features whose relationship with the output is modeled linearly.
  • Individual Nonlinear Features: Features whose relationship with the output is nonlinear but modeled independently of others.
  • Interacted Nonlinear Features: Features that interact with other features, but crucially, these interactions are local, meaning they only occur within predefined subsets of features. This avoids global, complex interactions that are hard to interpret.
  • Forward Stepwise Selection: A novel algorithm is proposed to distinguish between linear and nonlinear components. This algorithm iteratively builds a model, starting with a linear model and progressively adding nonlinear components, selecting the most impactful ones. This helps in identifying which features can be treated linearly and which require more complex modeling.
  • Additive Separability Theorem: The authors leverage and generalize the concept of additive separability to rigorously test for interactions between features. This allows them to determine if a complex interaction can be simplified into separate additive components, thus promoting transparency.
  • GGNAM Architecture: The final GGNAM architecture combines linear components (similar to linear regression) with individual nonlinear components (modeled by neural networks) and locally interacted nonlinear components (modeled by neural networks operating on subsets of features).

Datasets

The paper evaluates GGNAMs on three publicly available datasets:

  • Taiwan Credit Scoring (TCS) Dataset: Used for credit risk assessment, focusing on predicting the probability of default.
  • Polish Companies Bankruptcy Dataset: Used for predicting corporate bankruptcies.
  • Productivity Prediction of Garment Employees Dataset: Used for predicting worker productivity.

These datasets are standard benchmarks for classification and regression tasks in financial and economic contexts.

Key Insights and Ideas

  • Transparency Doesn't Necessarily Mean Sacrificing Accuracy: The GGNAM framework demonstrates that it's possible to achieve accuracy comparable to complex black-box models (like FCNNs) while maintaining high interpretability.
  • Local Interactions are Key for Parsimony: By restricting interactions to specific subsets of features, GGNAMs achieve a sparser, more interpretable model structure than models that allow arbitrary interactions.
  • Linearity is Often Sufficient: The forward stepwise selection algorithm reveals that for many real-world datasets, a significant portion of features can be modeled linearly, contributing to model simplicity and transparency.
  • Diminishing Marginal Effects (DMEs) are Important: The paper highlights how GGNAMs can effectively capture DMEs, which are common in financial and economic phenomena but are often missed by simpler linear models, leading to more accurate risk assessments.
  • GGNAMs Provide Actionable Insights: The visualizations generated by GGNAMs allow for a deeper understanding of feature relationships and interactions, providing valuable insights for regulators and decision-makers.

Key Results

  • Comparable Accuracy: GGNAMs achieved accuracy levels on par with or very close to Fully-Connected Neural Networks (FCNNs) across all tested datasets (Table I). For instance, on the TCS dataset, GGNAM achieved an AUC of 0.766, compared to FCNN's 0.759.
  • Simplified Architectures: The resulting GGNAM architectures were significantly simpler than FCNNs, characterized by a large number of linear components and a sparse set of interacted nonlinear components (Table II). For example, in the TCS dataset, only two features were identified as nonlinear, and no interactions were found.
  • Effective Feature Selection: The forward stepwise selection algorithm effectively identified a small number of nonlinear features, reducing the complexity of the models.
  • Capturing Complex Interactions: In the bankruptcy dataset, GGNAMs successfully identified and visualized a complex interaction between two features that was not apparent from simpler models, demonstrating their ability to uncover subtle relationships.

Limitations and Future Work

  • Domain Knowledge Integration: The authors acknowledge that the current approach does not explicitly incorporate domain knowledge. They suggest that future work should explore how to integrate domain expertise into the model building process to create more conceptually sound and fair models.
  • Computational Resources: While the forward stepwise selection aims to reduce computation, the process of calculating separability matrices and training multiple models can still be computationally intensive.
  • Choice of Metrics: The paper notes that the choice of metrics for evaluating separability (like the <€ threshold) can be subjective and depends on the specific application and desired trade-off between accuracy and interpretability.
  • Further Exploration of Interaction Types: While GGNAMs focus on local interactions, future work could explore other forms of structured interactions or more complex interaction patterns if justified by domain knowledge or empirical necessity.

Summary updated

Abstract

While machine learning methods have significantly improved model performance over traditional methods, their black-box structure makes it difficult for researchers to interpret results. For highly regulated financial 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 verified by applying additive separation criteria. Through some empirical examples in finance, we demonstrate that generalized grove of neural additive models exhibit high accuracy and transparency with predominantly linear terms and only sparse nonlinear ones.
BibTeX
@inproceedings{chen2024generalized,
  author = {Dangxing Chen and Weicheng Ye},
  title = {{Generalized Groves of Neural Additive Models: Pursuing Transparent Machine Learning Models in Finance}},
  booktitle = {CIFEr},
  year = {2024},
  doi = {10.1109/cifer62890.2024.10772966},
}
Figure