I have been taking an online course on data science, and was recently introduced the ideas of overfitting and underfitting, that is splitting the dataset we have into two parts into training(80%-90%) and testing(10%-20%) dataset. The instructor says we prefer to shuffle our data before the split, what is the reason for this? Isnt the whole idea of the approach of Splitting into Testing/Training Set is based on the assumption, that the observations are Independent and identically distributed random variables. Then there is no meaning to shuffling our data. Did i get it correct?
It may depend on where the data came from and how it was exported. It's not uncommon that real world data is sorted in some manner. For example it could be sorted by:
- user id
- timestamp of the observation
- outcome of interest
In each of these cases if you do a test/train split on the data without shuffling then you may have a different data distribution in your splits! For example if the data is sorted by user_id, then most users will appear in either test or training set, but not both.
How much this matters probably depends on how you intend to use your model. For example in lots of real world ML applications you train on historical data, and make predictions on future unseen data. In that case having the data sorted by timestamp before creating your splits might actually be desirable, since it matches the way you'll apply the model in the real world.
In general, splits are random, (e.g. train_test_split) which is equivalent to shuffling and selecting the first X % of the data. When the splitting is random, you don't have to shuffle it beforehand.
If you don't split randomly, your train and test splits might end up being biased. For example, if you have 100 samples with two classes and your first 80 samples are from class 1, and remaining are from class 0, a 80/20 split would leave all class 0's in the training set and all class 1's in the test set. This way, in the training, you wouldn't see any examples from class 1 and cannot learn.
Sometimes, it's even helpful to shuffle after the splits, e.g. in neural nets, to keep the parameters inside a reasonable subset.