[Paper Review] Interpretability is Harder in the Multiclass Setting: Axiomatic Interpretability for Multiclass Additive Models.

Xuezhou Zhang, Sarah Tan+4 more|arXiv (Cornell University)|Oct 22, 2018

Explainable Artificial Intelligence (XAI)28 references3 citations

TL;DR

This paper proposes a generalized additive model (GAM) framework for multiclass classification that maintains interpretability through a post-hoc calibration technique called API (Axiomatic Post-Processing Interpretability). It extends boosted tree-based GAM learning to multiclass settings and introduces two axioms ensuring visual interpretability, which API enforces on any pretrained multiclass additive model without accuracy loss, validated on a 12-class infant-mortality dataset.

ABSTRACT

Generalized additive models (GAMs) are favored in many regression and binary classification problems because they are able to fit complex, nonlinear functions while still remaining interpretable. In the first part of this paper, we generalize a state-of-the-art GAM learning algorithm based on boosted trees to the multiclass setting, and show that this multiclass algorithm outperforms existing GAM fitting algorithms and sometimes matches the performance of full complex models. In the second part, we turn our attention to the interpretability of GAMs in the multiclass setting. Surprisingly, the natural interpretability of GAMs breaks down when there are more than two classes. Drawing inspiration from binary GAMs, we identify two axioms that any additive model must satisfy to not be visually misleading. We then develop a post-processing technique (API) that provably transforms pretrained additive models to satisfy the interpretability axioms without sacrificing accuracy. The technique works not just on models trained with our algorithm, but on any multiclass additive model. We demonstrate API on a 12-class infant-mortality dataset.

Motivation & Objective

• To extend state-of-the-art boosted tree-based GAM learning to the multiclass classification setting.
• To identify fundamental interpretability axioms that prevent visual misleadingness in multiclass additive models.
• To develop a post-processing technique (API) that enforces these axioms on any pretrained multiclass additive model.
• To preserve model accuracy while ensuring interpretability in multiclass settings.
• To validate the method on a real-world 12-class infant-mortality dataset.

Proposed method

• Generalizes a state-of-the-art boosted tree algorithm to the multiclass setting for training multiclass GAMs.
• Defines two axioms—local consistency and global monotonicity—that ensure additive models are not visually misleading in multiclass scenarios.
• Develops a post-processing algorithm (API) that reparameterizes the output logits of any pretrained multiclass additive model to satisfy the axioms.
• API operates independently of the training algorithm, making it universally applicable to any multiclass additive model.
• The method ensures that feature contributions are visually interpretable and consistent across classes.
• Employs a constrained optimization framework to adjust model outputs while preserving original predictive performance.

Experimental results

Research questions

• RQ1 How can generalized additive models be effectively extended to multiclass classification while maintaining interpretability?
• RQ2 What axioms are necessary to ensure that multiclass additive models are not visually misleading?
• RQ3 Can a post-processing technique be designed to enforce interpretability axioms without degrading model accuracy?
• RQ4 How generalizable is the proposed API method across different multiclass additive models and datasets?
• RQ5 What is the empirical performance of the API-enhanced model on a real-world multiclass benchmark?

Key findings

• The proposed multiclass GAM training algorithm outperforms existing GAM fitting methods and matches the performance of complex models in some cases.
• The natural interpretability of GAMs breaks down in multiclass settings due to inconsistent feature contribution patterns across classes.
• The two proposed interpretability axioms—local consistency and global monotonicity—prevent misleading visualizations of feature effects.
• The API post-processing technique successfully enforces these axioms on any pretrained multiclass additive model without sacrificing accuracy.
• On a 12-class infant-mortality dataset, API-transformed models exhibit visually consistent and interpretable feature contributions across all classes.
• The method is universally applicable, working across different training algorithms and model architectures in the multiclass additive model space.