Certifying Robustness to Programmable Data Bias in Decision Trees

By Anna P. Meyer et al

Read the original document by opening this link in a new tab.

Table of Contents

Abstract
1 Introduction
2 Related work
3 Defining data bias programmatically
4 Certifying robustness for decision-tree learning
4.1 Certifying bias robustness with abstraction

Summary

Datasets can be biased due to societal inequities, human biases, under-representation of minorities, etc. The goal is to certify that models produced by a learning algorithm are pointwise-robust to potential dataset biases. Decision-tree learning is focused on due to the interpretable nature of the models. A novel symbolic technique is used to evaluate decision-tree learner on datasets. The study formalizes the bias-robustness certification problem and presents a language to define bias models programmatically. Techniques for certifying poisoning robustness and statistical defenses are discussed. The study highlights the importance of certifying robustness for decision trees in the context of bias in training data. The approach is evaluated on various bias models and datasets from the fairness literature.