genetic programming - Is that make sense to construct a learning Model using only one feature? -

To improve the accuracy of the Adobost classifier (for image classification), I am using genetic programming so that new statistical measures can be achieved.

. Every time a new feature is produced, I evaluate its fitness by training an Adobost Classifier and testing its performance. But I want to know that this process is right; I mean use a single feature to train a learning model.

You can create a model on one feature. I think "one feature" means you only have a number in R (otherwise, this would be completely "traditional" usage). However this means, that you are making a classifier in a one-dimensional space, and like - many classifier files will be redundant (since this is actually a simple problem). What is more important - whether you can classify the objects correctly using a special dimension, is checking that does not mean that after using it one Good / Bad feature This can be especially a case of:

• Many attributes can "search" the same event in the data, and therefore - each of them can get good results separately. , But once combined - each one of them gets better (as they only capture similar information ),
• the facilities waste until used in combination Maybe. Some events can only be described in multi-dimensional space, and if you are analyzing only one-dimensional data - then you will never be able to find your real value, because a simple example should consider four points (0,0), (0,1), (1,0), (1,1) such as (0,0), (1,1) one There are elements of the class, and the rest is another. If you look differently on each dimension - then the best possible accuracy is 0.5 (as you always have two different class digits at the same point - 0 and 1). Once combined - you can easily separate them, because this is a xor problem.

  In short - OK to create a classifier in one dimensional space, but:

  • Such a problem can be solved without "heavy machinery" May be.
  • The result should not be as the basis of feature selection to be more stringent - it can be very deceptive).