Is weight of evidence and information value a technique of dimension reduction I am trying to understand the concept of weight of evidence and information value. From what I understand, it is a variable reduction technique where we only use variables with IV > 0.5 in the model. If it is a dimension reduction technique, is it fair to say it is like principal component analysis or linear discriminant analysis?
How come the wikipedia page on dimension reduction does not include a discussion of WOE and IV?
 A: I think most people would classify what you are describing as feature selection instead. You want to pick the subset of features that have the most information about the predicted variable. In this case, you just pick the subset of individual features that have some threshold amount of information. 
A more sophisticated version of this approach is to consider the information value of an entire set of variables together. For instance, suppose I have three features, $f_1$, $f_2$, $f_3$ and I'm trying to predict $y$. Maybe I measure the information value, with, e.g., mutual information, and I find that $I(f_1; y) = 0.7, I(f_2; y) = 0.7, I(f_3; y) = 0.4$. Your strategy would be to pick $f_1$ and $f_2$. But, what if we look at these features and realize that $f_1 = f_2$! Keeping $f_2$ doesn't help us at all. If we had combined features we might have found that $I(f_1, f_2; y) = 0.7$ but $I(f_1, f_3; y) = 0.8$ (for instance). This is the intuition behind "maximal relevance minimum redundancy" methods (also discussed in the wikipedia article for feature selection).
