Boosting of local experts is a classification algorithm that facilitates
complex non-linear functions by using a divide-and-conquer and
margin-maximization approach. A local expert is a linear support vector
machine (SVM) that is trained by a subset of the input data lying in a
sub-region of the input space. Each local expert’s output is distance
weighted and then presented to a linear program (LP) that chooses a
linear combination of the few best hypotheses by maximizing the L1-norm.
When using this approach recursively, we obtain a set of trees that
order the data hierarchically. This structure is similar to a committee
of neural networks with several small hidden layers. First tests show a
competitively viable performance when compared to Gaussian- and
polynomial kernel SVMs. Moreover, the method holds further potential due
to its semi-supervised training strategy.
Ort: INB Seminarraum
