Next: Methods
Up: Finding structure
Previous: Coarsening and Refining
New dimensions - Meta-dimensions
This technique enables us to add new useful dimensions to our table. This can be ``measures'' as
discussed in section 2.3.1, extracted features, labels of identified clusters,
or any kind
of function of other variables.
This is especially useful for getting other viewpoints of the data. It helps connecting
the results of other algorithms (e.g. clustering) with the structure-finding techniques
discussed previously. For continuous variables linear combination, logarithms, multiplication
of variables, etc. may be useful. For example regression and neural networks use these transformations for classification (see sections 3.1.2 and 3.1.3).
It should be also mentioned that new dimensions are almost always used for
dimensionality reduction. In linear regression n dimensions are replaced with
n-1, n-k or 1 dimension(s), expressed e.g. as linear combinations of the original
distribution. Several dimensions are replaced with fewer new
dimensions. This is especially important in context of the dramatic complexity increase
with the number of dimensions (section 1).
Furthermore, a primary function is that a new dimension represents by itself
dependencies and relations among the old dimensions.
Thomas Prang
1998-06-07