Impact of Data Reduction Techniques on Classification

Although large datasets are often desirable in machine learning, they slow down training and require large storage. Data reduction can be used without affecting prediction accuracy. Many studies have examined feature selection, but did not consider instance reduction nor concern themselves with reducing training or testing time. Our goal is to reduce the training and testing time, by reducing both the number of features and the number of instances, with little or no drop in accuracy. We examined many methods of feature and instance reduction, and combinations thereof. Our results show that by integrating Information-Gain Feature Selection with Simple Random Sample without Replacement, we achieve our goal.