Discovering spatial contrast and common sets with statistically significant co-location patterns

Abstract

Co-location pattern mining is a spatial data mining technique which can be used to find associations among spatial features. Our work is motivated by an application in environmental health where the goal is to investigate whether the maternal exposure during pregnancy to air pollutants could be potentially associated with adverse birth outcomes. Discovering such relationships can be defined as finding spatial associations (i.e. co-location patterns) between adverse birth outcomes and air pollutant emissions. In particular, our application problem requires to find specific co-location patterns which are common to many spatial groups and co-location patterns which can discriminate one spatial group from the others. Traditional co-location pattern mining methods are not capable of finding such specific patterns. Hence, to achieve the spatial group comparison task, we introduce two new spatial patterns: spatial contrast sets and spatial common sets, and techniques to efficiently mine them based on co-location pattern mining. Traditional co-location pattern mining methods rely on frequency based thresholds which discard rare patterns and find exaggerated noisy patterns which may not be equally prevalent in unseen data. Addressing these limitations, we propose to use statistical significance tests instead of frequency to quantify the strength of a pattern. Towards this end, we propose to apply Fisher's exact test to efficiently find statistically significant co-location rules and use them to discover spatial contrast and common sets. Our experiments reveal that the Fisher's test based method could indeed help in finding co-location patterns with a better statistical significance leading to find valid spatial contrast and common sets. With the proposed methods we discovered that air pollutants such as heavy metals, NO2 and PM are significantly associated with adverse birth outcomes conforming to the existing domain knowledge thus validating our approach. We also evaluated our methods with synthetic datasets which confirmed that our methods indeed extract the patterns we seek to find.