Abstract
This paper describes the design and implementation of new naive Bayes and k-Nearest Neighbour methods that are highly scalable and efficient for document classification. Three methods for improving scalability are analysed: a change in the data representation and therefore in the algorithms' implementation, a partitioning mechanism that breaks down the problem into smaller parts, and a buffering mechanism to improve memory efficiency for large datasets. The classifiers were tested over two Reuters datasets: ModApte a popular but small benchmark, and RCV1 a new large collection of news stories, and compared to more standard implementations of these methods, both experimentally and analitically.