I have reviewed some questions already posted in the forum regarding whether it is necessary to normalize the target variable if you normalized the feature variables. Most of the answers stated that you don’t need to also normalize the target variable.
However, in the “C1_W2_Lab03_Feature_Scaling_and_Learning_Rate_Soln” lab, shown above, Dr. Ng states that “Any predictions using the parameters learned from a normalized training set must also be normalized”. Does this mean that you should be normalizing your target variable?
I understand that this question of whether to also normalize the target variable usually depends on the model being used. But in terms of learning algorithms using multiple regression is this necessary?
It means if you normalize the features of the training set, you also need to apply the same normalization when you use the model to make new predictions.
Also, this is an answer where the student basically asked my previous question but it was stated that you never normalize the target variable for classification algorithms but you can when using regression. Do you have any insight on when you should do this or should you generally not and if you ever need to, it is most likely a rare exception?
No. Typically your entire pile of labeled data will be split into three subsets. Perhaps 60% is the training data, 20% is the validation data, and 20% is the test data. These are only guidelines, not firm requirements.
I’ve only found extreme cases where normalizing the target data provides any benefit. If the labels span a very wide range of values (like many orders of magnitude), then normalizing the labels might be useful. Generally this is very rare, and at this point in an introductory course, you don’t need to worry about it.
