Multi Threshold FRPS: A New Approach to Fuzzy Rough Set Prototype Selection

Prototype Selection (PS) is the preprocessing technique for K nearest neighbor classification that selects a subset of instances before classification takes place. The most accurate state-of-the-art PS method is Fuzzy Rough Prototype Selection (FRPS), which assesses the quality of the instances by means of the fuzzy rough positive region and automatically selects a good threshold to decide if instances should be retained in the prototype subset. In this paper we introduce a new PS method based on FRPS, called Multi Threshold FRPS (MT-FRPS). Instead of determining one threshold against which the quality of every instance is compared, we consider one threshold for each class. We evaluate MT-FRPS on 40 standard classification datasets and compare it against MT-FRPS and the state-of-the-art PS methods and show that MT-FRPS improves the accuracy of the state-of-the-art PS methods.