|
Combining graph Laplacians for semi--supervised learning
Andreas Argyriou, Mark Herbster and Massimiliano Pontil
In: NIPS 2005, December, 2005, Vancouver, CA.
AbstractWe build on recent methods for graph regularization in semi-supervised
learning. A foundational problem for these methods is the construction
of the underlying graph. We propose to use a method which can optimally combine a number
of differently constructed graphs.
For each of these graphs we associate a basic graph kernel.
We then compute an optimal {\em combined} kernel.
This kernel solves an extended regularization problem which requires
a joint minimization over both the data and the set of graph kernels.
We present encouraging results on a variety of OCR tasks where the optimal combined kernel
is computed from graphs constructed with a variety of distance
functions and the `$k$' in nearest neighbors.
[Edit]
|
