In: Statistics and Probability
Which dimension reduction method (Network multidimensional scaling, Multidimensional scaling or Optimized multidimensional scaling) should be chosen to study outliers in the lower dimensional embedding? Why?
Solution: The dimension reduction method is often used to address problems by decreasing the number of variables in the data and looking for shorter representations. Dimension reduction method is often aimed at normal daily data, and applying it to events deviating from this daily data (so-called outliers) can affect such events negatively. It can indeed have a large impact on outliers.
Multidimensional scaling is used to
study outliers in the lower dimensional embedding. Multidimensional
Scaling is the name for a family of dimensionality reduction
techniques based on preserving distances in the data set. The
classical version of Multidimensional Scaling finds points
in a low dimensional space that minimizes
Here
are the high-dimensional points, and
is the Euclidean distance in the respective space. The classical
version of MDS is equivalent to PCA. Other members of the MDS
family use a different distance measure or a different quantity to
optimize than Eq.(1).