An efficient approach to detecting concept-evolution in network data streams

An important challenge in network management and intrusion detection is the problem of data stream classification to identify new and abnormal traffic flows. An open research issue in this context is concept-evolution, which involves the emergence of a new class in the data stream. Most traditional data classification techniques are based on the assumption that the number of classes does not change over time. However, that is not the case in real world networks, and existing methods generally do not have the capability of identifying the evolution of a new class in the data stream. In this paper, we present a novel approach to the detection of novel classes in data streams that exhibit concept-evolution. In particular, our approach is able to improve both accuracy and computational efficiency by eliminating “noise” clusters in the analysis of concept evolution. Through an evaluation on simulated and benchmark data sets, we demonstrate that our approach achieves comparable accuracy to an existing scheme from the literature with a significant reduction in computational complexity.