5.5 Splitting the data

At this point we might choose to split the data into the test and the training partitions. The reason for this is that we need an independent test we did not train on. This will become clearer in the following sections, but without having a separate test set, we cannot assess the performance of our model or tune it properly.

5.5.1 Holdout test dataset

There are multiple data split strategies. For starters, we will split 30% of the data as the test. This method is the gold standard for testing performance of our model. By doing this, we have a separate data set that the model has never seen. First, we create a single data frame with predictors and response variables.


# push sample ids back to the row names

Now that the response variable or the class label is merged with our dataset, we can split it into test and training sets with the caret::createPartition() function.

set.seed(3031) # set the random number seed for reproducibility 

# get indices for 70% of the data set
intrain <- createDataPartition(y = tgexp[,1], p= 0.7)[[1]]

# seperate test and training sets
training <- tgexp[intrain,]
testing <- tgexp[-intrain,]

5.5.2 Cross-validation

In some cases, we might have too few data points and it might be too costly to set aside a significant portion of the data set as a holdout test set. In these cases a resampling-based technique such as cross-validation may be useful.

Cross-validation works by splitting the data into randomly sampled \(k\) subsets, called k-folds. So, for example, in the case of 5-fold cross-validation with 100 data points, we would create 5 folds, each containing 20 data points. We would then build models and estimate errors 5 times. Each time, four of the groups are combined (resulting in 80 data points) and used to train your model. Then the 5th group of 20 points that was not used to construct the model is used to estimate the test error. In the case of 5-fold cross-validation, we would have 5 error estimates that could be averaged to obtain a more robust estimate of the test error.

An extreme case of k-fold cross-validation, is to equalize the \(k\) to the number of data points or in our case, the number of tumor samples. This is called leave-one-out cross-validation (LOOCV). This could be better than k-fold cross-validation but it takes too much time to train that many models if the number of data points is large.

The caret package has built-in cross-validation functionality for all the machine learning methods and we will be using that in the later sections.

5.5.3 Bootstrap resampling

Another method to estimate the prediction error is to use bootstrap resampling. This is a general method we have already introduced in Chapter 3. It can be used to estimate variability of any statistical parameter. In this case, that parameter is the test error or test accuracy.

The training set is drawn from the original set with replacement (same size as the original set), then we build a model with this bootstrap resampled set. Next, we take the data points that are not selected for the random sample and predict labels for them. These data points are called the “out-of-the-bag (OOB) sample”. We repeat this process many times and record the error for the OOB samples. We can take the average of the OOB errors to estimate the real test error. This is a powerful method that is not only used to estimate test error but incorporated into the training part of some machine learning methods such as random forests. Normally, we should repeat the process hundreds or up to a thousand times to get good estimates. However, the limiting factor would be the time it takes to construct and test that many models. Twenty to 30 repetitions might be enough if the time cost of training is too high. Again, the caret package provides the bootstrap interface for many machine learning models for sampling before training and estimating the error on OOB samples.