SGR: A New Efficient Kernel for Outlier Detection in Sensor Data Minimizing Mise

In sensor network, collected data is error prone due to errors during sensors and transmission. Sometimes, the sensed data may appear to be erroneous due to large deviation from normal data distribution. Such data points termed as outliers may contain some important pattern. Outliers, if neglected as erroneous data, may result in failure to detect important phenomenon. Hence, it is necessary to not only detect such data points but analyze them further to establish the reason behind such data values. The presence of outliers may distort contained information. To ensure that the information is correctly extracted, it is necessary to identify the outliers and isolate them during knowledge extraction phase. In this paper, we propose a novel unsupervised algorithm for detecting outliers based on density by coupling two principles: first, kernel density estimation and second assigning an outlier score to each object. A new kernel function building a smoother version of density estimate is proposed. An outlier score is assigned to each object by comparing local density estimate of each object to its neighbors. The two steps provide a framework for outlier detection that can be easily applied to discover new or unusual types of outliers. Experiments performed on synthetic and real datasets suggest that the proposed approach can detect outliers precisely and achieve high recall rates which in turn demonstrate the potency of the proposed approach. Copyright © 2015 IFSA Publishing, S. L.