Abstract
The classification based schemes, especially One class Support Vector Machines (SVM) have received a great interest in Machine learning community based on their Outlier & Event Detection applications in Wireless Sensor Networks (WSNs). The quarter-sphere (QS) formulation of One-Class SVM, known as Spatio-temporal-attribute (STA-QS-SVM) has a high efficiency, but has a significant communication overhead. Since communication is much more expensive than computation is WSNs, therefore STA-QS-SVM is unsuitable for energy constrained WSNs. This work presents three partially online novel approaches based on STA-QS-SVM, which lead to a significant (upto 95%) reduction in the communication overheads. Most appropriate scheme of the three can be selected depending on the application. The results indicate that the proposed schemes give a performance comparable to that of STA-QS-SVM, with an added advantage of a significant reduction in communication cost. The results also suggest that the proposed schemes retain their performance even for very high fraction of outliers and events in the data sets. Thus the algorithms can be used for event detection in energy constrained WSNs deployed in harsh environments.