Large Protein function prediction via faster graph kernels
Karsten Borgwardt, S V N Vishwanathan, Nicol Schraudolph and Hans-Peter Kriegel
In: NIPS workshop on Bio-informatics, 05 - 11 December 2005, Vancouver, Canada.
Kernel functions on graphs have been defined over recent
years. In earlier work, we have employed random walk graph kernels
for predicting protein function from graph representations that
integrate both protein sequence and structure. While yielding good
protein function prediction results, random walk graph kernels suffer
from a high computational complexity of $O(n^6)$ where $n$ is the
number of nodes in the input graphs. In this paper, we present an
approach for speeding up graph kernels. It is based on the
observation that random walks on a graph can be regarded as a
dynamical system and makes use of conjugate gradient.