Extending Decision Tree to Handle Multiple Fairness Criteria

Alessandro CastelnovoIJCAI 2022 · DOI

TL;DR

This paper introduces FFTree, a novel decision tree classifier designed to simultaneously address multiple fairness criteria and sensitive attributes. By extending traditional decision trees with a flexible, discrimination-aware splitting mechanism, FFTree aims to provide transparent and interpretable AI systems that meet complex business and regulatory fairness requirements.

What Problem Does This Paper Address?

As AI systems become more integrated into decision-making processes, ensuring both transparency and fairness is paramount. While Explainable AI (XAI) helps with transparency, it doesn't inherently solve ethical issues like algorithmic bias. Biases in training data can lead to unfair outcomes for certain demographic groups (e.g., based on gender or race). The challenge is that "fairness" itself is context-dependent and lacks a universal definition. Existing approaches to mitigate bias often struggle with handling multiple fairness criteria simultaneously, accommodating multiple sensitive attributes, or providing interpretable models. This paper tackles the need for a more flexible and comprehensive approach to building fair AI systems.

Methods

The core contribution of this paper is FFTree, an extension of the standard decision tree algorithm. The novelty lies in its splitting criterion. Instead of solely optimizing for information gain (a measure of how well a feature splits data based on the target variable), FFTree's splitting criterion also incorporates fairness constraints. Specifically, it seeks to find a split that maximizes information gain while satisfying user-defined fairness criteria (FC) for sensitive attributes (S) up to a specified discrimination level ($\delta$).

FFTree allows users to:

  • Select and mitigate desired fairness criteria from a set of available options (specifically, observational fairness criteria like independence, separation, and sufficiency).
  • Implement more than one fairness criterion simultaneously.
  • Handle more than one sensitive attribute.
  • Specify the desired level of fairness ($\delta$) to meet specific needs.

The paper mentions that the fairness metrics are calculated by computing the local prediction of each split.

Datasets

The paper mentions evaluating FFTree on:

  • A well-known benchmark dataset from the literature.
  • A real-world private dataset from Intesa Sanpaolo containing approximately 250,000 records related to personal loan applications.

The specific names of the benchmark dataset are not provided.

Key Insights and Ideas

  • Flexibility is Key: The demand for AI systems that can handle diverse and evolving fairness requirements necessitates flexible mitigation strategies, which current methods often lack.
  • Simultaneous Fairness: The ability to enforce multiple fairness criteria and handle multiple sensitive attributes concurrently is a significant step towards more robust and responsible AI.
  • Interpretable Fairness: By extending decision trees, FFTree aims to provide transparent rules that are understandable to users, bridging the gap between performance and fairness.
  • Local Optimization Matters: Optimizing fairness at each local split level ensures that the resulting rules are consistent with the chosen fairness constraints, avoiding potential issues where globally optimized but locally unfair rules might emerge.

Key Results

The paper states that FFTree is competitive with state-of-the-art fair trees on common features. Its novel functionality allows it to support users in various application domains. The authors suggest that tuning the $\delta$ hyperparameter can help identify the optimal global trade-off between performance and fairness for a given application.

Limitations and Future Work

Limitations: * Meeting multiple fairness criteria simultaneously can be challenging, as these criteria can impose non-trivial constraints on the data distribution, potentially leading to over-constrained solutions or mutually exclusive criteria under certain assumptions.

Future Work: * Releasing the FFTree code as open source to encourage community utilization and improvement. * Adding new fairness criteria or different formalizations of existing ones. * Exploring the use of FFTree within tree-based ensemble methods like Random Forests. * Integrating FFTree into XAI tools that operate on surrogates (e.g., ContrXT) or for monitoring fairness over time.

Summary updated

BibTeX
@inproceedings{castelnovo2022extending,
  author = {Alessandro Castelnovo},
  title = {{Extending Decision Tree to Handle Multiple Fairness Criteria}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/822},
}
Figure