hide
Free keywords:
-
Abstract:
Many machine learning problems (e.g. training SVMs) have a
mathematical programming (MP) formulation. An important advantage
of SVMs is elegant non-linear modelling via kernel functions;
however, a proper choice of kernel is crucial for accurate predictions.
Multiple kernel learning (MKL) is an extension to SVMs
that allows to optimize a linear combination of kernels during training.
We review MKL, present its different MP formulations, and
investigate their time complexity. Finally we discuss what is key
for tuning mathematical programming to achieve the required computational
efficiency.