Introduction
The newly released Synthetic Data Generation feature included in the latest version of Analytic Solver Data Science allows users to generate synthetic data by automated Metalog probability distribution selection and parameter fitting, Rank Correlation or Copula fitting, and random sampling. This can be beneficial for several reasons such as when the actual training data is limited or when the data owner is unwilling to release the actual, full dataset but agrees to supply a limited copy or a synthetic version that statistically resembles the properties of the actual dataset.
This process consists of three main steps.
- Fit and select a marginal probability distribution to each feature – by automated and semi-automated search within the family of bounded, semi-bounded or unbounded Metalog distributions.
- Identify correlations among features, by using Rank Correlation or one of available Copulas – Clayton, Gumbel, Frank, Student or Gauss.
- Generate the random sample consistent with the best-fit probability distributions and correlations.
Additional to the generated synthetic data, Analytic Solver Data Science can optionally provide the details of the fitting process – fitted coefficients and goodness-of-fit metrics for all fitted candidate Metalog distributions, selected distribution for each feature and fitted correlation matrix.
To further explore the original/synthetic data and compare them, one may easily compute basic and advanced statistics for original and/or synthetic data, including but not limited to percentiles and Six Sigma metrics.