Using information gain to build meaningful decision forests for multilabel classification

Abstract

“Gain-Based Separation” is a novel heuristic that modifies the standard multiclass decision tree learning algorithm to produce forests that can describe an example or object with multiple classifications. When the information gain at a node would be higher if all examples of a particular classification were removed, those examples are reserved for another tree. In this way, the algorithm performs some automated separation of classes into categories; classes are mutually exclusive within trees but not across trees. The algorithm was tested on naive subjects' descriptions of objects to a robot, using YUV color space and basic size and distance features. The new method outperforms the common strategy of separating multilabel problems into L binary outcome decision trees, and also outperforms RAkEL, a recent method for producing random multilabel forests.