It should be noted that--contrary to boosting, where the individual
runs depend on each other and have to be performed in
succession--pairwise classification can be easily parallelized by
assigning the binary classification problems to different processors,
as already noted by Friedman (1996) and
Lu and Ito (1999). As each binary task will be smaller than the
original task, the total training time of a multi-class problem of
size *n* will be significantly below that of a binary problem of the
same size, if each binary classifier can be trained on a separate
processor. Naturally, a parallel implementation would also provide a
trivial solution to the problem with classification efficiency
discussed above.

Johannes Fürnkranz 2002-03-11