Seppo Virtanen (University of Leeds) – Balancing diversity and fidelity of probabilistic synthetic data generation

In a probabilistic framework, synthetic data corresponds to data generated from the posterior distribution. The main task is to assess how similar the synthetic data is to the real data: this corresponds to model checking and assessing model fit. We present an approach that allows the user to balance between diversity and fidelity of the synthetic data regularising the model via a single parameter between zero and one. In one extreme, the model focuses on diversity covering the data distribution, and in the other the model distribution focuses increasingly close to the mode(s) of the data distribution. We demonstrate the proposed approach for latent factor and mixture models as well as regression and classification.