Manual feature engineering lets you observe the data and make conclusions by yourself.
Automatic feature engineering (NN) tries to find patterns between variables for yourself.
Since NN enginner features for you, I would always prefer to use this approach as it makes my life easier. However, my experience so far showed me that if you can afford the time to make a great analysis over the features and make some deductions out of it, then you should always engineer the features on your own.
Do you agree with me and what’s your personal experience?
Hello @popaqy
As your post is a general discussion which is not coming under MLS course2 /Week 2 assignment, you can post this is in General discussion forum. You will get many answers.
Regards
Jenitta
Agree with you that engineer our own features when we can. Relying on NN to find good features for us takes us many neurons, a lot of computational time, and good regularization. We need many randomly initialized neurons so that such diversity is large enough to give us the luck that they may come up with at least something useful. No matter how magical NN looks, the calculations are deterministic, and without enough neurons we don’t have a good chance.
While engineering the features manually definitely has its pros, it might not be practical with a real dataset.
Real datasets can have large number of features. Also, finding the exact equation that models the relationship between input and output is a far more complicated and rigorous task than finding a piece-wise linear model that approximates the original relationship between the inputs and output