Andrew Ng 's machine learning lecture note (12)

Which feature should we choose?

Well, sometimes it's really hard to tell which feature we should choose just by feeling, but we can use number to guide our decision about which feature we should choose. Andrew recommends we follow the following steps.

(1) We can implement a quick and dirty model for our job, (maybe less than 24 hours? )

(2)We use the cross validation set to see our model's performance, From spotting the trend most errors were made we can decide which feature we should choose.

(3)We can also plot the learning curve to know that whether we have a high bias or high variance problem, then we can decide to add more data or add the hidden units and so on.

(4)Consider that we are the expert then ask ourselves questions about whether these features are enough or not. 

Also, it's necessary to compare several methods'(Like in email spam problem whether we should distinguish mom and Mon and so on) performance, we can use a number from error analysis to tell which method is better.

Large data is useful?

We can have a sophisticated features or many hidden units in the neuron model and so on likely to cause a high variance problem . In this situation, more data is often more useful.