A heuristic for tuning the kernel in dynamic clustering

Kernel methods are a class of methods for data analysis that
generalize existing techniques by implicitly mapping the data into a
high dimensional feature space. In this talk we focus on kernel
clustering in a dynamic context where the groups in the cluster and
the features evolve over time. As in any kernel method, in kernel
clustering, the choice of the kernel is crucial for the results. We
explore different definitions of the performance of a kernel in the
context of dynamic clustering and develop a heuristic to tune the
kernel in order to maximize such performance. Kernel models that
are very flexible allow us to capture important information in the
data, at the expense of a need to tune many parameters. When the
number of parameters is large, it is difficult for traditional
metaheuristics to find good solutions. Our algorithm takes
advantage of the fact that complex kernel models can be seen as
generalization of simpler ones, yielding a nested sequence of models
of increasing complexity.