Fair Active Learning
share
Bias in training data and proxy attributes are probably the main reasons for bias in machine learning. ML models are trained on historical data that are biased due to the inherent societal bias. This causes unfairness in model outcomes. On the other hand, collecting labeled data in societal applications is challenging and costly. Hence, other attributes are usually used as proxies for labels that are already biased and result in model unfairness. In this paper, we introduce fair active learning (FAL) for mitigating machine bias. Considering a limited labeling budget, FAL carefully selects data points to be labeled in order to balance between the model performance and its fairness quality. Our comprehensive experiments, comparing traditional active learning with FAL on real datasets, confirm a significant improvement of fairness for models trained using FAL, while maintaining the model performance.
READ FULL TEXT
