An Efficient Approach to Detect Concept Drifts in Data Streams

Abstract

Due to the presence of data streams in many applications like banking, sensor networks, and telecommunication, data stream mining has gained increased attention. Data stream is continuous, ordered sequence of data instances arriving at a rapid rate. One of the key challenges while learning from data streams is the detection of concept drift, i.e., changes in data distribution underlying data streams, observed over time. Drifts being either gradual or sudden, several algorithms have been put forward for detection of different kinds of drift. However, most of them work on only one of these kinds of drift. These algorithms show hampered output if different types of drift are mixed. To solve this issue, there is need of a single system that can handle all drifts simultaneously. In this paper, we propose a system that detects both kinds of drift efficiently. Our system combines features of the online classifier as well as a blockbased classifier to achieve the goal. We further analyzed drifts to find out missing values of attributes to be the root cause. Our system handles missing values in different ways for more improved performance.