| dc.description.abstract |
Nonlinear regression models are often used in statistics and machine learning due
to greater accuracy than linear models. In this work, we present a novel modeling
framework that is both computationally efficient for high-dimensional datasets, and
predicts more accurately than most of the classic state-of-the-art predictive models.
Here, we couple a nonlinear random Fourier feature data transformation with an intrinsically
fast learning algorithm called Vowpal Wabbit or VW. The key idea we develop is that
by introducing nonlinear structure to an otherwise linear framework, we are able to
consider all possible higher-order interactions between entries in a string. The utility
of our nonlinear VW extension is examined, in some detail, under an important problem
in statistical genetics: genomic selection (i.e.~the prediction of phenotype from
genotype). We illustrate the benefits of our method and its robustness to underlying
genetic architecture on a real dataset, which includes 129 quantitative heterogeneous
stock mice traits from the Wellcome Trust Centre for Human Genetics.
|
|
