Optimizing Stream Data Classification Using Improved Hoeffding Bound

Abstract

Classification of online stream data is must for network analysis and providing Quality of Service (QoS). Stream data has properties which requires the algorithm to be incremental and should handle concept drift. Traffic classification is the prominent solution to handle bulk data streams to provide services like packet filtering, routing policies, traffic shaping, limiting traffic, etc. Many stream-based classification algorithms exists in literature to meet the requirements like scanning the data only once, any time analysis and fast response, and limited memory utilization. Further, more accurate, fast, and limited memory supporting algorithms and concepts are required to handle everyday increasing data over Internet. This research work proposes an improvement in accuracy of the classification performed using lesser number of training instances to decide a split during induction of the decision tree (Hoeffding tree). Jensens’s inequality concept is used, and the Hoeffding bound reduces to minimize the bound for the bad events (i.e., it limits the margin of error of the algorithm). Number of examples reduced results in fast execution and decrease the memory used.