Performance Comparison of Cumulative and Class Indicator Approaches for Pollution Risk Assessment

Abstract

Cumulative indicator coding usually consists in setting the attribute value to one when this value is below a given threshold, and to zero otherwise. Another possibility is to use class indicator transform. Here, the indicator is one if the variable falls within a specified interval, and zero otherwise. The two types of indicator have been used on a real dataset, in order to estimate conditional distribution functions (ccdfs) using ordinary indicator cokriging. The results show that, although it is theoretically more appropriated to capture the class-specific patterns of spatial continuity, the class indicator approach shows less performance than its cumulative counterpart. Furthermore, it is more tedious to implement and produces a higher number of order relation deviations.