The k nearest neighbour method (kNN) can be used not only on an entire data set, but also after a selection of instances is performed. Selection of instances should select prototypes which well represent the knowledge about a given problem. We propose a new algorithm of prototype selection. The algorithm is based on selection of instances which represent the borders between classes and additionally they are trustworthy instances. Moreover, our algorithm was optimized with a forest of dedicated locality sensitive hashing (LSH) trees to speed up the prototype selection and the classification process. The algorithm's final expected complexity is O(m log m). Additionally, results show that the new algorithm lays ground for accurate classification.
展开▼