Parallel R Model Prediction Building and Analytics

[This article was first published on R, Ruby, and Finance, and kindly contributed to R-bloggers]. (You can report issue about the content on this page here)
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

Modifying R code to run in parallel can lead to huge performance gains. Although a significant amount of code can easily be run in parallel, there are some learning techniques, such as the Support Vector Machine, that cannot be easily parallelized. However, there is an often overlooked way to speed up these and other models. It involves executing the code that generates predictions and other analytics in parallel, instead of executing the model building phase in parallel, which is sometimes impossible. I will show you how this can be done in this post.

First, we will set up our variables. The setup is fairly similar to the one I have used in other posts, but note that the length of the vectors has been increased by a magnitude of 100 to more easily show how much time can be saved by parallelizing the prediction building phase.

positions <- sample(nrow(all_data),size=floor((nrow(all_data)/4)*3))
training<- all_data[positions,]
testing<- all_data[-positions,]
We now have a testing set of 25,000 rows, and a training set of 75,000 rows, which is somewhat linear. We will train an SVM on this data. Note that it may take more than 10 minutes to train an SVM on this data, particularly if you have an older computer. If you do have an older computer, feel free to reduce the number of rows in the data frame as needed.
After we have built the model, we generate predictions for it, which yields an error of 13045.9 for me(although this may be different for your data). Our next step is timing the prediction phase to see how long it takes.
On my system, this took 35.1 seconds.

We are now ready to set up our parallel infrastructure. I run Windows, and will use foreach and doSNOW here, although you can certainly use other parallel packages here if you prefer. You can read this post if you want an introduction to foreach, doSNOW, and doMC. If you do not elect to use doSNOW, you will not need to use the stopCluster() function that appears in some of the code below.
cl<-makeCluster(4) #change the 4 to your number of CPU cores
Now, we have the groundwork for our parallel foreach loop, but we need to find a way to split the data up in order to perform predictions on small sets of data in parallel.
This will create a numeric vector that can be used to split the testing data frame into 4 parts. I suggest setting num_splits to some multiple of your number of CPU cores in order to execute the below foreach loop as quickly as possible. Now that we have a way to split the data up, we can go ahead and create a loop that will generate predictions in parallel.
.combine=c,.packages=c("e1071")) %dopar% {
It is very important that the .packages argument be used to load the package that corresponds to the prediction function you are going to use in the loop, or R will get confused about which prediction function to use and generate an error. The .combine argument tells the foreach loop to combine the outputs of the foreach loop into a vector. A hidden argument that defaults to true ensures that all the outputs remain in order.

Now, we test to make sure that everything is okay by checking what the error value is:
I got 13045.9, which matches the value I got before, and confirms that both the parallel and non-parallel prediction routines return the exact same results.

Now, we can create a function and time it to see how fast the parallel technique is:
.combine=c,.packages=c("e1071")) %dopar% {

This takes 12.76 seconds on my system, which is significantly faster than the non-parallel implementation.

This technique can be extended to other analytics functions that can be run after the model is built, and it can generate predictions for any model, not just for the svm that the example uses. While creating the model can take up much more time than generating predictions, it is not always feasible to parallelize model creation. Running the prediction phase in parallel, particularly on high dimensional data, can save significant time.

To leave a comment for the author, please follow the link and comment on their blog: R, Ruby, and Finance. offers daily e-mail updates about R news and tutorials about learning R and many other topics. Click here if you're looking to post or find an R/data-science job.
Want to share your content on R-bloggers? click here if you have a blog, or here if you don't.

Never miss an update!
Subscribe to R-bloggers to receive
e-mails with the latest R posts.
(You will not see this message again.)

Click here to close (This popup will not appear again)