WeightLOFCC: A heuristic weight-setting strategy of lof applied to outlier detection in time series data

Abstract

Finding outliers is more interesting than finding common patterns in many KDD applications. The local outlier factor method (LOF) is a popular approach to detect outliers, in which a degree of being an outlier will be assigned to each object. In this paper, we present a modification method called WeightLOFCC to better handle outliers in time series data. Differing from the traditional LOF algorithm, the proposed WeightLOFCC method utilizes the idea of semi-supervised learning and weight factor to model data, and makes use of the cross correlation to measure the similarity. We evaluated the proposed algorithm on a large variety of data sets, and the experiment results show that for most of the data sets, our solution for outlier detection can achieve the best performance compared with other classical techniques. © 2010 Springer-Verlag.