| |
Abstract:
This paper introduces the
Relevance Vector Machine,
a generalised linear model for regression and classification whose
output, like the Support Vector Machine (SVM), is a weighted sum of
kernel functions associated with a subset of the training examples.
A high degree of sparsity is achieved through a Bayesian treatment
by introducing a prior distribution over the model weights, and
maximising the marginal likelihood over the values of the
hyperparameters. Examples in both regression and classification
settings illustrate generalisation at least as good as a comparable
SVM, while utilising dramatically fewer kernel functions.
|
