Is there a rule-of-thumb for how to divide a dataset into training and validation sets?

Question

How can I divide data into training and validation sets, Should I divide it 50%-50% for both or is there another criteria to divide it in training data and validation data or this thing depends over application? Currently I am using 80% training data and 20% validation data, Is there anyone who's experienced in machine learning advice me on this? 

Answer 1

@kavita,There are two main concerns regarding the division-

1.With Less training data,your parameter estimates have greater variance.

2.With Less testing data, your performance statistic will have greater variance.

It should be divided in such a way that neither variance is too high.According to Pareto principle, 80/20 is the common occurring ratio.

Let’s assume you have enough data for a proper split, following are some instructive ways to get a handle on variances:

1. split the data into training and testing.
2. Then slit the training data into validation and training.
3. Subsample random selections of training data, train the classify and then record a performance on the validation set.
4. Try a different type of splits, you will notice greater performance with more data.
5. To get a handle on variance follows the same procedure but in reverse.

If you are a beginner and want to know more about Machine Learning, then check out this course by Intellipaat which will teach you ML from basics: Machine Learning Course