BiasFairnessAccountabilitycs.LG
Underestimation Bias and Underfitting in Machine Learning
Padraig Cunningham, Sarah Jane Delany
May 18, 2020
arXiv.org
4 citations
1 influential
Key Takeaways
-
Algorithmic bias has multiple sources: The paper distinguishes between bias inherited from historical training data and bias that algorithms themselves can introduce or worsen—a critical distinction often overlooked in bias mitigation discussions.
-
Overfitting prevention can paradoxically increase bias: Regularization techniques used to prevent overfitting (when models memorize training data too closely) can actually accentuate underestimation bias, creating an unintended trade-off that practitioners need to understand.
-
Current research focuses too narrowly on elimination: Most bias research targets removing bias regardless of its source, but the paper argues researchers should first understand how algorithms contribute to bias before attempting solutions.
-
Underestimation bias deserves more attention: The paper highlights that underestimation bias—a specific type of algorithmic bias—has received insufficient research focus despite being "inextricably tied" to common machine learning practices like regularization.
Abstract
Often, what is termed algorithmic bias in machine learning will be due to historic bias in the training data. But sometimes the bias may be introduced (or at least exacerbated) by the algorithm itself. The ways in which algorithms can actually accentuate bias has not received a lot of attention with researchers focusing directly on methods to eliminate bias - no matter the source. In this paper we report on initial research to understand the factors that contribute to bias in classification algorithms. We believe this is important because underestimation bias is inextricably tied to regularization, i.e. measures to address overfitting can accentuate bias.