we are looking at PLS as a means to reduce brittleness

traditionally has been used as a means to reduce training data sets
data mining algorithms have also focused on scalability because of the increase in data available
aim has been to reduce data while maintaining comparable results

back - many instance selectors, but none measure their brittleness.
aim - present a framework for measuring brittleness in IS and knn
method - you know
results - graphs - measure area under curve

As already noted elsewhere [32], training-set reduction
algorithms can be characterized by their storage reduction,
classification speed increase, generalization accuracy, noise
tolerance, and learning speed. Among these criteria,
learning speed is usually neglected. However, in order to
be practicable on large training sets or in knowledge
discovery applications requiring a learning step in their
cycle, the method should exhibit good learning behavior.

Lots of classification models, you measure classification acc, speed etc.
what about the brittleness of a result. [define]

Usually classifiers are characterized by...