Generating Synthetic Data with The Nearest Neighbors Algorithm
The k nearest neighbor algorithm (kNN) is one of the most popular nonparametric methods used for various purposes, such as treatment effect estimation, missing value imputation, classification, and clustering. The main advantage of kNN is its simplicity of hyperparameter optimization. It often produces favorable results with minimal effort. This paper proposes a generic semiparametric (or nonparametric if required) approach named Local Resampler (LR). LR utilizes kNN to create subsamples from the original sample and then generates synthetic values that are drawn from locally estimated distributions. LR can accurately create synthetic samples, even if the original sample has a non-convex distribution. Moreover, LR shows better or similar performance to other popular synthetic data methods with minimal model optimization with parametric distributional assumptions.
READ FULL TEXT