Tailoring data source distributions for fairness-aware data integration

Data scientists often develop data sets for analysis by drawing upon sources of available to them. A major challenge is ensure that the set used has an appropriate representation relevant (demographic) groups: it meets desired distribution requirements. Whether collected through some experiment or obtained from provider, any single source may not meet Therefore, a union multiple required. In this paper, we study how acquire such in most cost effective manner, typical functions observed practice. We present optimal solution binary groups when underlying distributions are known and all have equal costs. For generic case with unequal costs, design approximation algorithm performs well When unknown, exploration-exploitation based strategy reward function captures approximations group each source. Besides theoretical analysis, conduct comprehensive experiments confirm effectiveness our algorithms.