Scalability of Learning Arbiter and Combiner Trees from Partitioned Data

Much of the research in inductive learning concentrates on problems with relatively small amounts of data residing at one location. In this paper we explore the scalability of learning arbiter and combiner trees from partitioned data. Arbiter and combiner trees integrate classiiers trained in parallel from small disjoint subsets. Previous work demonstrated their eecacy in terms of accuracy, this paper discusses their performance in terms of speedup and scalability. The performance of serial learning algorithms is evaluated. The performance of the algorithms used to construct combiner and arbiter trees in parallel is then analyzed. Our empirical results indicate that the techniques can eeectively scale up to large datasets with millions of records.