Abstract

Many learning systems implicitly use the fit-andsplit learning method to create a comprehensive hypothesis from a set of partial hypotheses. At the core of the fit-and-split method is the assignment of examples to partial hypotheses. To date, however, this core has been neglected. This paper provides the first definition and model of the fit-and-split assignment problem. Extant systems perform assignment nearly arbitrarily, implicitly using, for example, greedy set covering. This paper also presents Conceptual Set Covering (CSC), a new assignment algorithm. An extensive empirical evaluation over a wide range of learning problems suggests that CSC can improve any fit-and-split learning system. 1 Introduction One way to solve a complex learning problem is to decompose it into simpler learning problems. The fit-andsplit learning method is perhaps the most direct way to do such a decomposition. FIT-AND-SPLIT LEARNING METHOD Input: Examples and learning problem called the...

Citations