FAE: A Fairness-Aware Ensemble Framework

Abstract

Automated decision making based on big data and machine learning (ML) algorithms can result in discriminatory decisions against certain protected groups defined upon personal data like gender, race, sexual orientation etc. Such algorithms designed to discover patterns in big data might not only pick up any encoded societal biases in the training data, but even worse, they might reinforce such biases resulting in more severe discrimination. The majority of thus far proposed fairness-aware machine learning approaches focus solely on the pre-, in-or post-processing steps of the machine learning process, that is, input data, learning algorithms or derived models, respectively. However, the fairness problem cannot be isolated to a single step of the ML process. Rather, discrimination is often a result of complex interactions between big data and algorithms, and therefore, a more holistic approach is required.The proposed FAE (Fairness-Aware Ensemble) framework combines fairness-related interventions at both pre-and postprocessing steps of the data analysis process. In the preprocessing step, we tackle the problems of under-representation of the protected group (group imbalance) and of class-imbalance by generating balanced training samples. In the post-processing step, we tackle the problem of class overlapping by shifting the decision boundary in the direction of fairness.

Online Fairness-Aware Learning with Imbalanced Data Streams

“…In reality though the low discrimination scores are just an artifact of the low prediction rates for the minority class. This observation has been made in [20,21,22] but for the static case. We observe the same issue for the streaming case and therefore, we propose an imbalance monitoring mechanism based on which we adapt the weighted training distribution.…”

“…Finally, CSMOTE is amplifying unfairness (and causes reverse discrimination in NYPD dataset) by re-sampling the instances. A reason for such behavior can be the amplification of existing encoded biases in the data through instance re-sampling (also reported in [20]).…”

“…In-processing approaches reformulate the classification problem by explicitly incorporating the model's discrimination behavior in the objective function through regularization or constraints, or by training on latent target labels [22,28]. Postprocessing approaches [12,18,20] alter a model's predictions or adjust a model's decision boundary to reduce unfairness. Stream Learning: In stream learning, data arrive sequentially and their distributions can change over time, the so-called concept drifts [15].…”

“…As a result of the ever-increasing interest in issues of fairness and responsibility of datadriven learning algorithms, a large body of work already exists in the domain of fairnessaware learning [19,20,21,22,26,27,28,32]. Only a few recent works, however, investigate the problem of fair learning in non-stationary environments [24,38].…”

Multi-fairness under class-imbalance

“…Recent studies [12,14] also showed that many datasets in the domain are imbalanced and typically, the class-imbalance problem is more severe for the protected group (e.g female, black, etc. ), which is often underrepresented in the important minority class [13]. Despite high overall accuracy and fairness, methods ignoring imbalance may still perform poorly on the minority class, thus amplifying prevalent biases.…”