Output details

Kernel choice and classifiability for RKHS embeddings of probability distributions

<13>Two-sample testing can be accomplished in high dimensions and on non-Euclidean spaces (distributions over documents or graphs) by representing these distributions in a high dimensional reproducing kernel Hilbert space (feature space). There are infinitely many possible feature spaces, however, and one needs to choose the one which best distinguishes the distributions. This paper shows a good kernel for hypothesis testing optimizes classifiability of the samples, in the process linking for the first time relatively new results on kernel hypothesis testing with classical results on classification. NIPS 2009 talk (awarded to 2% of submissions), runner up for best student paper award.