A method to detect data outliers from smart urban spaces via tensor analysis

An Online Method to Detect Urban Computing Outliers via Higher-Order Singular Value Decomposition

“…In the offline modeling step, we model the outlier detection problem as a system. The following scheme summarizes our proposed method in a diagram based on that presented by Aquino et al [25] and also used by Souza et al [18] in the modeling of the offline detection. However, for an online monitoring model, we adapt the modeling using a sliding window that allows us to employ fixed-length sliding windows with well-defined time intervals.…”

“…The data collection used in this study was obtained through an environmental monitoring platform called Smart Citizen [19,26,27]. The data collection, processing and modeling procedure in this article was performed similarly to that done by Souza et al [18]. In our proposal, we adapted the method proposed by Souza et al [18] of the sampling Ψi of three functions, to a new method composed of two functions since we did not use the clustering analysis function.…”

“…The data collection, processing and modeling procedure in this article was performed similarly to that done by Souza et al [18]. In our proposal, we adapted the method proposed by Souza et al [18] of the sampling Ψi of three functions, to a new method composed of two functions since we did not use the clustering analysis function. Therefore, sampling was composed of the functions a HOSVD reduction of dimensionality (ΨH) and online outlier detection function (ΨO) through the sliding window:Ψi=ψH∘ψO,…”

“…In particular, from an environmental perspective, outlier detection from environmental monitoring gains strength in the literature since the detection of noise in cities, water and air pollution, forest conditions, and so on, can provide for a sustainable and intelligent development for cities [10]. In a previous paper, we proposed an offline method to explore the data of a multiway nature of urban spaces in the detection of outliers [18]. This method executes through three stages: in the first one, the data is modeled as a third order data tensor to achieve the reduction of dimensionality in order to obtain a better fit, the second stage is comprised of a classification step, and finally the third step generates a model of identification of refined urban space standards.…”

“…Here we propose an online for an outlier detection method that combines multivariate and multidimensional approaches to characterize the behavior patterns of urban environmental monitoring data from high-dimensional structured datasets. We extend our previous method [18] to an online detection approach which: i. uses a sliding window which keeps track of the most recent data and all decomposition and detection tasks performed based on what is "visible" through the window; and ii. evaluates a performance of online outlier detection from both accuracy and receiver operational characteristic (ROC) curves.…”