Kernel-Based Machine Learning with Multiple Sources of Information / Kernbasiertes Maschinelles Lernen mit mehreren Informationsquellen
Marius Kloft · it - Information Technology · 2013
Summary We present a new methodology for fusing information from multiple data sources - or kernels - in machine learning. Previous approaches promoted sparse combinations of kernels, which, however, may discard important information. We present a flexible approach based on ℓp-norm regularization, allowing for non-sparse solutions. In a theoretical analysis we show lower and upper generalization bounds of order up to O(M/n), overcoming the best previously known upper bounds for the problem, which achieved O(√M/n). The computational experiments indicate that the novel algorithms are up to two orders of magnitude faster than previous approaches. Applications to computational biology and computer vision show accuracies that go beyond the stateof-the-art.