Deep learning-based Data Alternation to Remove Bias

This project is motivated by a unique problem identified by the ABS. ABS has a huge collection of administrative datasets collected over the last few years. Over the years, the dataset has become biased and lost its robustness to be used for any future downstream tasks. They also have a recently collected survey dataset with the same attributes as that of the historically biased dataset. While the survey dataset is unbiased, it is a small collection. This research develops a transformer-based model to alter the biased data based on the available small unbiased data.