A transductive extension of maximum entropy/iterative scaling for decision aggregation in distributed classification

Abstract

Many ensemble classification systems apply supervised learning to design a function for combining classifier decisions, which requires common labeled training samples across the classifier ensemble. Without such data, fixed rules (voting, Bayes rule) are usually applied. [1] alternatively proposed a transductive constraint-based learning strategy to learn how to fuse decisions even without labeled examples. There, decisions on test samples were chosen to satisfy constraints measured by each local classifier. There are two main limitations of that work. First, feasibility of the constraints was not guaranteed. Second, heuristic learning was applied. Here we overcome both problems via a transductive extension of maximum entropy/improved iterative scaling for aggregation in distributed classification. This method is shown to achieve improved decision accuracy over the earlier transductive approach on a number of UC Irvine data sets.