Multi-class
Generalized Linear Discriminant for Race Classification
Perhaps a
multi-class generalized linear discriminant
(multi-GLD) will yield better performance.
Capable
of approximating arbitrary decision boundaries.
May be
able to make more efficient use of the sparse training data.
Recall
that
a c-class multi-GLD is essentially a linear
machine; partitions the input feature space into c
decision regions.
So long
as the training data for the non-White classes is sufficiently
representative of the variation in the training set, it
should be possible to define these partitions with relatively few
examples.