Fair Outlier Detection Based on Adversarial Representation Learning

Shu Li,Yi Lu,Rui Qiu,Xusheng Du,Jiong Yu

doi:10.3390/sym14020347

Abstract

Outlier detection aims to identify rare, minority objects in a dataset that are significantly different from the majority. When a minority group (defined by sensitive attributes, such as gender, race, age, etc.) does not represent the target group for outlier detection, outlier detection methods are likely to propagate statistical biases in the data and generate unfair results. Our work focuses on studying the fairness of outlier detection. We characterize the properties of fair outlier detection and propose an appropriate outlier detection method that combines adversarial representation learning and the LOF algorithm (AFLOF). Unlike the FairLOF method that adds fairness constraints to the LOF algorithm, AFLOF uses adversarial networks to learn the optimal representation of the original data while hiding the sensitive attribute in the data. We introduce a dynamic weighting module that assigns lower weight values to data objects with higher local outlier factors to eliminate the influence of outliers on representation learning. Lastly, we conduct comparative experiments on six publicly available datasets. The results demonstrate that compared to the density-based LOF method and the recently proposed FairLOF method, our proposed AFLOF method has a significant advantage in both the outlier detection performance and fairness.

Highlights

With the development of machine learning technology, more and more decisionmaking problems have been replaced by algorithms
Aiming at the insufficiency of the FairLOF algorithm in outlier detection performance and fairness, this paper proposes an outlier detection method based on adversarial fair representation learning
We explore whether the LOF algorithm can produce fair detection results in two cases

Summary

Introduction

With the development of machine learning technology, more and more decisionmaking problems have been replaced by algorithms. Machine learning is a data-driven approach to automated decision-making that has a high potential to introduce or even perpetuate discriminatory issues already present in the data [1]. Existing research results suggest that algorithms trained using unbalanced datasets may reflect or even reinforce the social biases present in the data, such as the bias of facial analysis algorithms against skin color [2], Word2Vec algorithms against gender [3], and advertising recommendation systems against gender [4]. Research work in fairness machine learning aims to eliminate potential discrimination of algorithms. Most work in fairness machine learning has focused on supervised learning, especially on classification problems [5,6]. The latest research work has been on fairness research in unsupervised directions, such as clustering algorithm [7] and recommendation systems [8]

Objectives

Methods

Results

Conclusion