Metodologías de detección de outliers en datos espaciales, temporales y espacio-temporales

Abstract

While you are recording datasets it is possible to notice values which are substantially larger or smaller than usual. Such registers, named outliers hereinafter, might be correct, or might arise after a data recording error or a processing error. Research on outlier detection might provide useful information about the dataset, while being helpful in the database cleansing operation. In this paper we describe different methodologies recently proposed for outlier detection, intended for spatial and spatio-temporal dataset, as well as case results. Among them, we distinguish between those purely spatial, purely temporal, and mixed ones that have proven its value in controlled experiments. Among others we will discuss: 1) use an algorithm that considers temporal data while combines the advantages of clustering and approximations based upon kernel densities 2) compare the measured value with its expected value calculated in a incremental fashion considering temporal correlation, thus showing a spatial correlation in the recent past 3) propose a new algorithm for outlier detection for large spatio-temporal databases using spatial and aspatial information as well as temporal information 4) detect oultiers in temporal databases using association rules extracted from the normal behaviour and linking temporal evolution of the atributes 5) detect the outliers based upon the global properties of the dataset unlike most of the prevailing methods which have a local span and 6) assign, for the specific case of spatial outliers, different weights for different neighbors in order the estimate the central value, and estimating the weights as functions of the distance. After analyzing the different methods described, the importance of selecting one method or another will be shown. It is necessary to prosecute the research on the reliability and speed of the outlier detection algorithms.