Contextual Spatial Outlier Detection with Metric Learning
Guanjie Zheng (College of Information Sciences and Technology, Pennsylvania State University);Susan L. Brantley (Department of Geosciences, Pennsylvania State University);Zhenhui Li (College of Information Sciences and Technology, Pennsylvania State University)
Abstract
Hydraulic fracturing (or ``fracking’‘) is a revolutionary well stimulation technique for shale gas extraction, but has spawned controversy in environmental contamination. If methane from gas wells leaks extensively, this greenhouse gas can impact drinking water wells and enhance global warming. Our work is motivated by this heated debate on environmental issue and we propose data analytical techniques to detect anomalous water samples with potential leakages. We propose a spatial outlier detection method based on contextual neighbors. Different from existing work, our approach utilizes both spatial attributes and non-spatial contextual attributes to define neighbors. We use robust metric learning to combine different contextual attributes in order to find more precise neighbors. Our technique can be generalized to any spatial dataset. The extensive experimental results on six real-world datasets demonstrate the effectiveness of our proposed approach. We also show some interesting case studies, with one case linking to a gas well leakage.
