Abstract
Based on a leave-one-out bound of Jaakola and Haussler a modification
of the original SV algorithm is devised in order to minimize the
bound directly. This formulation is essentially parameter free,
maintains sparsity of the solution, and can be solved by a linear
program. The novelty can be found in the fact that rather than maximizing
the overall minimum margin, the individual margin of patterns is
maximized adaptively. Experiments show that its classification performance
is very competitive with an optimally adjusted SV machine and comparable
to a $\nu$-SV classifier. Uniform convergence bounds are provided.
Users
Please
log in to take part in the discussion (add own reviews or comments).