constructFastKRRLearner: Fast implementation for Kernel Ridge Regression.

In gdkrmr/DRR: Dimensionality Reduction via Regression

Fast implementation for Kernel Ridge Regression.

Description

Constructs a learner for the divide and conquer version of KRR.

Usage

constructFastKRRLearner()

Details

This function is to be used with the CVST package as a drop in replacement for constructKRRLearner. The implementation approximates the inversion of the kernel Matrix using the divide an conquer scheme, lowering computational and memory complexity from O(n^3) and O(n^2) to O(n^3/m^2) and O(n^2/m^2) respectively, where m are the number of blocks to be used (parameter nblocks). Theoretically safe values for m are < n^{1/3}, but practically m may be a little bit larger. The function will issue a warning, if the value for m is too large.

Value

Returns a learner similar to constructKRRLearner suitable for the use with CV and fastCV.

References

Zhang, Y., Duchi, J.C., Wainwright, M.J., 2013. Divide and Conquer Kernel Ridge Regression: A Distributed Algorithm with Minimax Optimal Rates. arXiv:1305.5029 [cs, math, stat].

See Also

constructLearner

Examples

ns <- noisySinc(1000)
nsTest <- noisySinc(1000)

fast.krr <- constructFastKRRLearner()
fast.p <- list(kernel="rbfdot", sigma=100, lambda=.1/getN(ns), nblocks = 4)
system.time(fast.m <- fast.krr$learn(ns, fast.p))
fast.pred <- fast.krr$predict(fast.m, nsTest)
sum((fast.pred - nsTest$y)^2) / getN(nsTest)

## Not run: 
krr <- CVST::constructKRRLearner()
p <- list(kernel="rbfdot", sigma=100, lambda=.1/getN(ns))
system.time(m <- krr$learn(ns, p))
pred <- krr$predict(m, nsTest)
sum((pred - nsTest$y)^2) / getN(nsTest)

plot(ns, col = '#00000030', pch = 19)
lines(sort(nsTest$x), fast.pred[order(nsTest$x)], col = '#00C000', lty = 2)
lines(sort(nsTest$x), pred[order(nsTest$x)], col = '#0000C0', lty = 2)
legend('topleft', legend = c('fast KRR', 'KRR'),
       col = c('#00C000', '#0000C0'), lty = 2)

## End(Not run)

gdkrmr/DRR documentation built on March 23, 2023, 5:42 a.m.