Downsampled and Undersampled Datasets in Feature Selective Validation (FSV)

摘要

Feature selective validation (FSV) is a heuristic method for quantifying the (dis)similarity of two datasets. The computational burden of obtaining the FSV values might be unnecessarily high if datasets with large numbers of points are used. While this may not be an important issue per se it is an important issue for future developments in FSV such as real-time processing or where multidimensional FSV is needed. Coupled with the issue of dataset size, is the issue of datasets having “missing” values. This may come about because of a practical difficulty or because of noise or other confounding factors making some data points unreliable. These issues relate to the question “what is the effect on FSV quantification of reducing or removing data points from a comparison-i.e., down- or undersampling data?” This paper uses three strategies to achieve this from known datasets. This paper demonstrates, through a representative sample of 16 pairs of datasets, that FSV is robust to changes providing a minimum dataset size of approximately 200 points is maintained. It is robust also for up to approximately 10% “missing” data, providing this does not result in a continuous region of missed data.