Code for the paper "An Empirical Analysis of KDE-based Generative Models on Small Datasets" // Procedia Computer Science Journal - 2021 by Plesovskaya E., Ivanov S. (In press).
This paper presents a framework for synthetic dataset similarity estimation based on two-sample tests. It also specifically accounts for the model overfitting detection. The framework is applied to samples, obtained from a range of KDE-based generative models trained on samples from multivariate_normal_datasets and empirical_datasets. In addition to that, we compare the generative capability of KDE with other algorithms used to generate tabular data: gaussian mixture models, copulas, and deep learning models.
tools/generative_model_estimation.pyandtools/generative_model_sampling.pyimplement generative model estimation and sampling.- Code for synthetic dataset similarity evaluation is in
tools/dataset_similarity_metrics.py. - We also use graphical analysis methods to analyse the similarity of samples which include pairwise plots, Quantile-Quantile and Probability-Probability plots
tools/plot_functions.py.