I thought I understood Adaboost, until code analysis made me realize that sample_weight is not an array of the feature weights... and after further investigation I am left confused as to how multidimensional features are even utilized in Adaboost without stumps or tree.
Is the implicit definition of a weak learner the inability to calculate the specific feature weights of a sample? (ie if a learner knew each feature weight, it could simply attempt the classification on its own?) Therefore, an average weight-sum over a set of weak learners gives a sign that can (hopefully) approximate away any mistakes...
If so, after delving into the code and a little theory, I am left confused as to how multidimensional features are utilized by weak learners in pure Adaboost. I am interested in knowing the feature weight dynamics as they change over the iterative set of weak learners. I do(believe) I understand Gini and entropy measures method to split/feature selection for trees. I am also under the impression that each feature is utilized in the iterations of Adaboost, perhaps because they are distribution(sum=1)... or are features selected iteratively as well... if so I see no evidence or mention in theory/code code applications I have reviewed. Perhaps there are some helpful resources?
So to recap, is my intuition about feature weight complexity vs. approximation of an averaged sample_weight threshold correct? How does sample_weight relate to feature weight? When exponential loss is described as a population minimizer, is this in reference to features, or samples with negative labels?
Finally, I apologize if these questions are naive, or have been explained elsewhere. I am really interested in weight dynamics more so than classification, and do not wish to go down the tree path... how are feature weights obtained/tracked using adaboost with exponential loss?
Thankyou
