A hybrid error-driven approach to data stream classification

Abstract

Challenges in the field of data stream mining include the vast volume of data being mined, the speed at which data arrives, and the presence of concept drifts. Traditionally, data classification has always involved the assumption of prior knowledge of the data sets, a method which is not particularly suitable when dealing with high-speed data streams. As such, various methods have been developed for the specific use-case of stream data mining, which are able to handle concept drifts during the data mining process with varying degrees of accuracy. Here, a probabilistic queuing model - based on an existing ‘SyncStream’ algorithm - is used in order to passively detect and account for the presence of abrupt concept drifts. In addition, other aspects of the system are tuned for better classification accuracy and throughput.