PARA: A positive-region based attribute reduction accelerator.

Attribute reduction, also known as feature selection, is a common problem by selecting a subset of relevant attributes (e.g. features) to reach efficient learning/mining. Many attribute reduction methods have been proposed however, quite often, these methods are still computationally time-consuming while handling large-scale data. To overcome this shortcoming, we present a novel accelerator based on the positive region, by deleting the learned/discernible instance pairs in the process of attribute reduction, which can avoid redundant computation and accelerate attribute reduction. Our experiments numerically demonstrate that the proposed accelerator can reach drastically faster computation than previous methods, especially on the datasets with a large number of instances.