Feature transformation by function decomposition

Abstract

The function decomposition described can identify subsets of existing features and discover nongiven functions that map these subsets to a new feature, also, it can organize the existing and new features into a hierarchy. The authors demonstrate their Hierarchy Induction Tool (HINT) system on a housing loan-allocation application. Methods for switching circuit design often implicitly deal with feature transformation. Such methods construct a circuit to implement a given or partially given tabulated Boolean function. The authors' function-decomposition method can discover and construct a hierarchy of new features that one can add to the original dataset or transform into a hierarchy of less complex datasets. The method allows the decomposition to deal with nominal-feature (that is, not necessarily binary) functions.