Quarter-Sphere SVM: Attribute and Spatio-Temporal correlations based Outlier & Event Detection in wireless sensor networks

Abstract

Support-Vector Machines (SVM) have received a great interest in the machine learning community since their introduction, especially in Outlier Detection in Wireless Sensor Networks (WSN). The Quarter-Sphere formulation of One-Class SVM (QS-SVM), extends the main SVM ideas from supervised to unsupervised learning algorithms. The QS-SVM formulation is based only on Spatio-Temporal correlations between the sensor nodes (hence the name Spatio-Temporal Quarter-Sphere SVM, ST-QS-SVM). Thus, it has a non-ideal performance. This work presents a new One-Class Quarter-Sphere SVM formulation based on the novel concept of Attribute Correlations between the sensor nodes, hence the name, Spatio-Temporal-Attribute Quarter-sphere SVM (STA-QS-SVM) formulation. Online and partially online approaches to Outlier Detection in WSNs have been presented using this formulation. The results indicate a significant increase in the Outlier Detection rates and a remarkable reduction in the False Positive rates over the previous formulation (ST-QS-SVM). The results of this novel technique also suggest that the partially online approach is as efficient as the online approach, thereby conserving significant computational and communication complexity. Moreover very high Event Detection rates have been reported for STA-QS-SVM, which have not been reported by ST-QS-SVM.