Preserving Clusters in Synthetic Data Sets Based on Correlations and Distributions

The rising popularity of machine learning has resulted in quality data becoming increasingly valuable. However, in some cases, the data are too sparse to effectively train an algorithm or the data cannot be disclosed to unaffiliated researchers due to privacy concerns. The sparsity of data may also affect various data analyses that require a certain volume of data to be accurate. One possible solution to the aforementioned problems is data generation. However, to be a viable solution, data generation must simulate real-life data well. To this end, this paper tests whether a previously presented iterative data generation method that generates synthetic data sets based on the attribute distributions and correlations of a real-life data set can faithfully reproduce a clustered data set. The approach is shown to be ineffective for the proposed application, and consequently, a new method is introduced that might preserve the clusters present in the real-life data set. The new method is demonstrated to not only preserve the clusters within the synthetic data set, but also improve the similarity of the attribute correlations of the synthetic data set and the real-life data set.