Towards Creating Fair Classification Models in Reality Mining Using Data Resampling Techniques

Multiple recent efforts in human-centered computing have used mobile and ubiquitous data to infer propensities of individuals (e.g. to trust others or behave altruistically). Often studied under the umbrella of “Reality Mining” multiple such efforts have reported high accuracies at the considered prediction tasks. However, there has been little work at quantifying the “fairness” of such algorithms in terms of how the quality of the predictions varies over different demographic groups (e.g. across gender). This work takes inspiration from data resampling techniques to create fair classification models. Empirical results suggest that a combination of over and under sampling technique (SmoteTomek) to the sensitive (protected) attribute (e.g. gender) yields improved model’s performance while reducing disparity across genders.