Some of the limitations of simple minimum-Euclidean distance classifiers
can be overcome by using a Mahalanobis metric.
In particular, this can often solve problems caused by poorly scaled and/or
highly correlated features.
We will develop the Mahalanobis metric indirectly by considering the effects
of scaling and linear transformations on data for which the Euclidean metric
is appropriate. The following topics are covered:
- Mean and variance
- Linear transformations
- Covariance matrix
- Mahalanobis metric
- Linear discriminants again
Back to Limitations
On to Learning
Up to PR for HCI