| The Ontario Research Centre for Computer Algebra
The UWO ORCCA Reading Room
Many practical problems reduce to classifying curves among multiple
classes, for example on-line recognition of handwritten mathematical
symbols. By treating a curve as an algebraic object and computing
truncated expansions of its parametric coordinate functions in an
orthogonal functional basis, we obtain an accurate, compact, and
geometrically intuitive representation of the curve as a point in a
low-dimensional vector space. Previous work has shown that, with this
representation, high top-k classification rates can be achieved using
support vector machines. However, the gap between the top-1 and top-2
classification accuracies remained large.
We report on a variation of nearest neighbor classification using the distance to the convex hull of several nearest neighbors. This reduces the tie-breaking errors between the top two classes by about 20% and the overall error rates by about 10%. The technique is well adapted to classification among hundreds of classes, with the number of training samples ranging from ten to several thousands, and with strict requirements on the speed and memory use.
If you have any questions or comments regarding this page please send mail to