Boundary effect correction ink-nearest-neighbor estimation

Abstract

The problem of the boundary effect for the k-nearest-neighbor (kNN) estimation is addressed, and a correction method is suggested. The correction is proposed for bounded distributions, but it can be used for any set of bounded samples. We apply the proposed correction to entropy estimation of multidimensional distributions and time series, and this correction reduces considerably the bias and statistical errors in the estimation. For a small sample size or high-dimensional data, the corrected estimator outperforms the uncorrected estimator significantly. This advantage makes the kNN method applicable to more real-life situations, e.g., the analysis of biological and molecular data.

Comprehensive review of association estimators for the inference of gene networks

“…These approaches directly estimate the MI by using the K-nearest neighborhood (KNN) method [98][99][100][101][102][103][104][105]. They compared those two approaches with the KNN entropy estimator [97,99,[106][107][108][109][110][111][112][113], which estimates the MI indirectly from the entropies by using yeast expression dataset with 6000 genes and 300 samples. The bias caused by the separately estimation of H(X) , H( Y ), and H( X , Y ) is decreased in the MI (1) and MI (2) methods.…”