Reduction of Reference Set for Network Data Analyzing Using the Bubble Algorithm

Abstract

Data in modern networks should be analysed in real time. This is necessary to maintain security and enable management. The real problem is the size of analysed data and its classification. One of several solutions to these problems may be reduction the reference network data set. Most algorithms for the condensation of the reference set involve a lot of computation when processing a very large set which contains several dozens of objects. That was the basis for our attempt to develop a completely new classifier which would maintain the quality of classification on the levels obtained with the primary reference set as well as allow to accelerate computations considerably. The proposed solution consists in covering the primary reference set with disjoint hyperspheres; however, these hyperspheres may contain objects from one class only. Classification is completed when it has been determined that the classified point belongs to one of the mentioned spheres. If an object does not belong to any hypersphere, it is counted among the objects of the same class, to which the objects from the nearest hypersphere belong (the distance to the centre of the sphere minus the radius). As was shown in our tests, this algorithm was very effective with very large data sets. © Springer International Publishing Switzerland 2015.