Using Optimization-Based Classification Methodfor Massive Datasets

摘要

Optimization-based algorithms, such as Multi-Criteria Linear programming (MCLP), have shown their effectiveness inclassification. Nevertheless, due to the limitation of computation power and memory, it is difficult to apply MCLP, or similaroptimization methods, to huge datasets. As the size of today’s databases is continuously increasing, it is highly important thatdata mining algorithms are able to perform their functions regardless of dataset sizes. The objectives of this paper are: (1) topropose a new stratified random sampling and majority-vote ensemble approach, and (2) to compare this approach with theplain MCLP approach (which uses only part of the training set), and See5 (which is a decision-tree-based classification tooldesigned to analyze substantial datasets), on KDD99 and KDD2004 datasets. The results indicate that this new approach notonly has the potential to handle arbitrary-size of datasets, but also outperforms the plain MCLP approach and achievescomparable classification accuracy to See5.