Categorizing Concepts for Detecting Drifts in Stream

Mining evolving data streams for concept drifts has gained importance in applications like customer behavior analysis, network intrusion detection, credit card fraud detection. Several approaches have been proposed for detection of concept drifts in the context of supervised learning in data streams. Recently, researchers have been looking into the problem of identifying concept drifts in unlabeled data streams. Prevalent approaches study the evolution of streaming clusters using intrinsic and extrinsic characteristics of the discovered clusters, where each cluster is considered a concept. In this paper we model an unlabeled, uniform data stream as a stochastic poisson process and study the arrival pattern of data points to analyse the nature of an evolving concept (cluster). Each concept is modeled as stochastic poisson process and is individually observed for arrival rates of the incoming data points. A random sample of arrival rates is collected for each concept and appropriate non-parametric tests are applied to infer the nature of evolution for the concept. Concept drift in the stream can be inferred by the overall behavior of the concepts. We also propose a taxonomy of various types of concept behaviors and inter-relation among them. Experiments have been performed to demonstrate feasibility, validity and scalability of the proposed method.