README.md

In avery-kruger/kitchen: Convolutional Kitchen Sinks in R

kitchen: Convolutional Kitchen Sinks in R

kitchen is an R package that provides tools to implement convolutional kitchen sinks, also known as convolutional random features, in R. These tools allow you to generate random features from data, train models, identify what set of hyperparameters produce best-performing models, and predict values from data.

Installation

You can install the development version of kitchen from GitHub with:

# install.packages("devtools")
devtools::install_github("avery-kruger/kitchen")

Details

Details on various functions can be found in their respective documentation pages. For a complete list of functions, use library(help = kitchen).

Example

Here is a simple example of how to implement kitchen:

Let there be some data, x, that has a complicated relationship with an outcome, y.

library(kitchen)
x <- matrix(sample(1:10,10000,TRUE),2000,5)
y <- 5*x[,1] + 20*x[,1]*x[,2] + 3*x[,3]^2 - 10*x[,4] - 2*x[,5]

We can test CKS across a variety of feature and window sizes to determine the best performing hyperparameters by using kitchen_sweep( ). This presents R2 values of prediction-to-true value for each combination of feature and window sizes.

kitchen_sweep(x[1:1000,],y[1:1000],
   x[1001:2000,],y[1001:2000],
   2^(4:8),2:5, verbose = F)
#>             2         3         4         5
#> 16  0.6132109 0.8180464 0.9220285 0.9203979
#> 32  0.6122578 0.8409847 0.9225346 0.9524986
#> 64  0.6119458 0.8421398 0.9319825 0.9665591
#> 128 0.6118863 0.8556787 0.9473810 0.9762115
#> 256 0.6131870 0.8554886 0.9483183 0.9851318

Models with 5 features appear to perform well, and feature counts beyond 64 don’t appear to make a difference.

Now, we can make a prediction on some new data using kitchen_prediction( ). We’ll use the hyperparameters we determined work well and bootstrap to generate a distribution of predictions based on our sampling of the data.

a <- matrix(sample(1:10,100,TRUE),20,5)
b <- 5*a[,1] + 20*a[,1]*a[,2] + 3*a[,3]^2 - 10*a[,4] - 2*a[,5]

mybootstrap <- kitchen_prediction(x, y, a,
                                  features=64, windows=5,
                                  bootstrap=10, verbose = F)

Lets plot the mean of our predictions versus the true values alongside a 1:1 line to see how well our predictions did. This is of course only possible in simulated data.

plot(b, rowMeans(mybootstrap[[1]][[1]]),
     xlab = "True", ylab = "Predicted"); abline(0, 1)

Author(s)

Avery M. Kruger

Maintainer: Avery M. Kruger, [e-mail]

References

Kruger, A. M., Shankar, V., Davies, T.S. (2024) Inferring the Evolutionary Model of Community-Structuring Traits with Convolutional Kitchen Sinks. Systematic Biology (In Press).

avery-kruger/kitchen documentation built on May 7, 2024, 6:40 a.m.