CrowdMI: Multiple Imputation via Crowdsourcing

Missing data is universal and methods to deal with it far ranging from simply ignoring it to using complex modelling strategies such as multiple imputation and maximum likelihood estimation.Missing data has only been effectively imputed by machines via statistical/machine learning models. In this paper we set to answer an important question ”Can humans perform reasonably well to fill in missing data, given information about the dataset?”. We do so in a crowdsourcing framework, where we first translate our missing data problem to a survey question, which then can be easily completed by crowdworkers. We address challenges that are inherent to crowdsourcing in our context and present the evaluation on a real dataset. We compare human powered multiple imputation outcomes with state-of-the-art model based imputation.