Fairness in machine learning with tractable models

Abstract Machine Learning techniques have become pervasive across a range of different applications, and are now widely used in areas as disparate recidivism prediction, consumer credit–risk analysis insurance pricing. The prevalence machine learning has raised concerns about the potential for learned algorithms to biased against certain groups. Many definitions been proposed literature, but fundamental task reasoning probabilistic events is challenging one, owing intractability inference. focus this paper taking steps towards application tractable models fairness learning. Tractable recently emerged that guarantee conditional marginal can be computed time linear size model. In particular, we show sum product networks (SPNs) enable an effective technique determining statistical relationships between protected attributes other training variables. We will also motivate concept “fairness through percentile equivalence”, new definition predicated on notion individuals at same their respective distributions should treated equivalently, prevents unfair penalisation those who lie extremities distributions. compare efficacy pre-processing with alternative approach assumes additive contribution. It was found when these two approaches were compared data set containing results law school applicants, equivalence method reduced average underestimation exam score ethnic minority applicants black bottom end distribution by fifth. conclude outlining improvements our existing methodology suggest opportunities further work field.