Instance Selection by using Polar Grids

Instance selection is about algorithms that search for a representative portion of data that can fulfill a data mining task as if the whole data is used. It is a very important data reduction technique, which can spare much memory and running time for data mining algorithms. This paper proposes a new method for Instance Selection by using Polar Grids (ISPG). The main idea is to search for a subset of instances located close to decision boundary by using a method based on polar grids. Original training instances are mapped into Polar reference frame, and the data space is partitioned as a set of polar grids. Then a special search algorithm is designed for determining which instances locate close to decision boundary. The method can also handle noisy instances and smooth data boundaries. The classical k-Nearest Neighbors (kNN) classification algorithm is employed to test the proposed method. Experiments show that the proposed method can reduce datasets effectively and achieve reasonable generalization accuracy. Moreover, the method achieves prominent learning speed, which can be used to process large spatial datasets.