An Empirical Analysis of KDE-based Generative Models on Small Datasets

One of the approaches to deal with the small dataset problem is synthetic data generation. Kernel density estimation is a common method to approximate the underlying probability distribution of a small dataset. The present paper aims to analyze the generation capability of KDE-based models by evaluating their samples. For this purpose, we introduce a framework for synthetic dataset quality estimation which also accounts for the overfitting of a generative model. The performance of KDE is analyzed on samples from theoretical distributions and real datasets. The results state that KDE generates synthetic samples of a good quality and outperforms its competitors on small datasets.