Performs resampling (both under- and over-sampling) of observations in a multivariate data frame
Consider the following scenario: you have a multivariate data set containing different numbers of observations between different levels of a factor, e.g., different number of samples or repetitions for different participants in a study. You need to make sure that you have the same number of observations for each of these levels.
You could under-sample the entire data set via random sampling, so that the number of observations in each level is equal to a small number; this isn't optimal because you throw away a lot of data in the process.
You could over-sample the entire data set via random sampling, so that the number of observations in each level is equal to a large number; this isn't optimal because you replicate feature vectors in the process, which is detrimental to, e.g., accurate training of neural networks.
The R function multivar_resamp allows for resampling of a data set to a target number of observations, either under-sampling or over-sampling as appropriate. In the case of over-sampling, new observations are created in a bootstrap fashion by creating weighted averages of feature vectors of randomly selected nearest neighbors in a multidimensional space. Verification that no duplicated observations are created is included in the over-sampling process.
Examples are shown below of a multivariate data set that was over-sampled at 250% (from 2338 to 5845 observations).
Feature with a uni-modal distribution:
Feature with a bi-modal distribution:
Feature with a multi-modal distribution:
Feature with a skewed distribution:
