Figure 8.

Comparing mixture model and Bayes networks using two different datasets

(Left) Accuracy and computation speed of two models in generating synthetic data (Carat study). For low-dimensional discrete data, Bayes networks are good, but as dimensionality grows, their computation time becomes intolerable and mixture models become more accurate. The solid lines denote the mean Frobenius norm (see Equation 8) between the original and the synthetic covariance matrices, with error bars denoting standard error of the mean from 10 independent runs of the algorithm. The dashed lines show the run times. Privacy budget was fixed to $\left(\epsilon =1.0,\delta ={10}^{-5}\right)$. (Right) Accuracy of data synthesized with two models (ARD study). Mixture models preserve regression coefficients better than the Bayes network. The curves show mean absolute error between the original and the learned coefficients. The average over 100 runs is shown. Error bars indicate the standard error of the mean.