Creating Acceptable Take a look at Units and Sleeping Soundly.
In this post, I want to look at a topic that is frequently disregarded by both the person asking and the person responding: “How do you split a dataset into training and test sets?”
It is standard procedure to divide the dataset into the training set and the test set, or at least, two halves, before taking on a supervised task. The test set is used to confirm whether the knowledge gained can be applied to “unknown” data—that is, data that was not included in the previous phase—while the training set is utilized to investigate the phenomenon.
Many people usually reach this conclusion by using common sense and straightforward methods. “I randomly partition the available data, reserving 20% to 30% for the test set” is the typical, uninteresting response.
Those who go beyond include the idea of stratified random sampling, which is sampling at random with one or more variables while keeping fixed proportions. Let us say we have a target variable with a prior probability of 5% and we are operating in a binary classification environment. Getting a training set and a test set that preserve the 5% proportion on the target variable’s prior is known as stratified random sampling. Such reasoning is occasionally required, such as when classifying in an extremely unbalanced environment, but it does not really excite the