Mining cross-patterning in large areal aggregated spatial datasets

Intelligent crime analysis allows for a greater understanding of the dynamics of unlawful activities, providing possible answers to where, when and why certain crimes are likely to happen. With the growth of geo-referenced data and the sophistication and complexity of spatial databases, data mining and knowledge discovery techniques have become essential tools for the successful analysis of large spatial datasets. Crime analysis requires a combination of heterogeneous data, such as socio-economic and socio-demographic factors, geospatial features and crime datasets so that interesting patterns can be discovered.\ud \ud The Queensland Police Service (QPS) and Australian Bureau of Statistics (ABS) record crime and census information in areal aggregated datasets. The primary reason for this is to protect the privacy of individuals. To enable intelligent crime analysis with these areal aggregated spatial datasets, the cross-patterning relationship between different spatial layers across locations needs to be modelled and quantified.\ud \ud This thesis focuses on developing a framework for the discovery and visualisation of crosspatterning in areal aggregated spatial datasets. We show that cross-patterning can be modelled in three ways. Cross-association patterns that model the relationship between multiple datasets for each region while ignoring the effect of neighbouring regions, cross-varying patterns that take into consideration the effect of all local neighbouring regions and cross-distribution patterns that consider one local neighbouring region and the global distribution of the dataset. With areal aggregated datasets, local neighbours are defined as those that share a boundary.\ud \ud The most suited type of cross-patterning is dependent on the application and dataset. We present a generic framework that can discover all types of cross-patterning from areal aggregated spatial datasets. The user can either select a specific type of pattern they wish to discover, or can use the framework to discover all cross-patterning and then use the visualisation environment to highlight patterns of interest.\ud \ud We develop algorithms to discover each type of cross-patterning. To model cross-association we propose an Association Rules Mining (ARM) based approach. To model cross-varying we extend Lee's bivariate spatial correlation approach [51], and to model cross-distribution we propose two new techniques based on the spatial distribution of the dataset. The framework also includes a visualisation environment that easily allows the user to interpret discovered patterns and highlight patterns that show overlapping cross-patterning (for example both crossassociation and cross-distribution).\ud \ud The algorithms developed for discovering cross-patterning relationships are designed so they can be applied to areal aggregated spatial datasets from a wide variety of disciplines. We test the framework with synthetic datasets to verify their correctness and then present a case study using real crime data from Brisbane, Australia. The computational performance of the algorithms is analysed and is acceptable for large datasets