FairOD: Fairness-aware Outlier Detection

Abstract

Fairness and Outlier Detection (OD) are closely related, as it is exactly the goal of OD to spot rare, minority samples in a given population. When being a minority (as defined by protected variables, e.g. race/ethnicity/sex/age) does not reflect positive-class membership (e.g. criminal/fraud), however, OD produces unjust outcomes. Surprisingly, fairness-aware OD has been almost untouched in prior work, as fair machine learning literature mainly focus on supervised settings. Our work aims to bridge this gap. Specifically, we develop desiderata capturing well-motivated fairness criteria for OD, and systematically formalize the fair OD problem. Further, guided by our desiderata, we propose FairOD, a fairness-aware outlier detector, which has the following, desirable properties: FairOD (1) does not employ disparate treatment at test time, (2) aims to flag equal proportions of samples from all groups (i.e. obtain group fairness, via statistical parity), and (3) strives to flag truly high-risk fraction of samples within each group. Extensive experiments on a diverse set of synthetic and real world datasets show that FairOD produces outcomes that are fair with respect to protected variables, while performing comparable to (and in some cases, even better than) fairness-agnostic detectors in terms of detection performance.

Deep Clustering based Fair Outlier Detection

“…Competitive Methods. We compare our model with the fair outlier detectors FairLOF [9], FairOD [46], and 15 conventional unsupervised outlier detection methods, including linear models: Principal Component Analysis (PCA) [47], One-class Support Vector Machine (OCSVM) [44]; proximity-based models: Local Outlier Factor (LOF) [2], Connectivity-Based Outlier Factor (COF) [48], Clustering Based Local Outlier Factor (CBLOF) [18]; probabilitybased models: Fast angle-based Outlier Detector (FABOD) [25], Copula Based Outlier Detector (COPOD) [33]; ensemble-based models: Feature Bagging (FB) [28], iForest [34], Lightweight On-line Detector of Anomalies (LODA) [43], Clustering with Outlier Removal (COR) [35]; neural networks: AutoEncoder (AE) [1], Variational Auto Encoder (VAE) [22], Random Distance Prediction (RDP) [51].…”

“…FairLOF [9] is the first paper to address the fair outlier detection problem, which incorporates a corrective term on the baseline LOF algorithm, in regards to local sensitive subgroup diversity and global outlier alignment with the baseline. FairOD [46] is another pioneering work, which has recently been released on arXiv. This method targets an equal outlier rate on the majority and minority sensitive subgroups.…”

“…Existing methods [9,46] on fair outlier detection achieve group fairness by balancing out the protected subgroup diversity, which adds difficulty to outlier uncovering as it blurs the underlying data pattern. To alleviate that downside, they preserve a consistent outlier ranking with respect to a pre-obtained baseline result, yet this may also impose restrictions on detection validity.…”

“…Deepak and Abraham [9] propose the FairLOF algorithm, which improves classical LOF [2] in regards to three heuristic principles: neighborhood diversity, apriori distribution, and attribute asymmetry, to prevent an unfair sensitive subgroup distribution among top-ranked outlier candidates. Shekhar et al [46] design FairOD, a deep model using a standard autoencoder component for self-reconstruction with the statistical parity and group fidelity as fairness constraints to ensure the majority protected subgroup share the same outlier rate as the minority one.…”

Automated Self-Supervised Learning for Graphs

“…Several works have illustrated the potential for representations learned in a self-supervised way to encode bias unintentionally, for example in language modeling [57,58], image representation learning [59] and outlier detection [60].…”