Online Data Stream Learning and Classification with Limited Labels

Loo Hui Ru (Faculty of Electrical Engineering, Universiti Teknologi Malaysia)
Trias Andromeda (Department of Electrical Engineering, Diponegoro University, Semarang)
M. N. Marsono (Faculty of Electrical Engineering, Universiti Teknologi Malaysia)

Proceeding of the Electrical Engineering Computer Science and Informatics
Vol 1: EECSI 2014

Mining data streams such as Internet traffic andnetwork security is complex. Due to the difficulty of storage, datastreams analytics need to be done in one scan. This limits thetime to observe stream feature and hence, further complicatesthe data mining processes. Traditional supervised data miningwith batch training natural is not suitable to mine data streams.This paper proposes an algorithm for online data streamclassification and learning with limited labels using selective selftrainingsemi-supervised classification. The experimental resultsshow it is able to achieve up to 99.6% average accuracy for 10%labeled data and 98.6% average accuracy for 1% labeled data. Itcan classify up to 34K instances per second.