Massive Spatial Interaction Data Mining And Visualization

Big and dynamic spatial-temporal data have been, and continue to be, collected with modern data acquisition techniques. In this research, we focus on the spatial interaction data, a specific type of spatial data that concerns the connection or movement between origin and destination, also refers as origin-destination flow data. This data is widely available in the ordinary life, for instance, human migration, vehicle trajectory, animal movement, and disease spreading. Understanding large volume of origin-destination flow data benefits us in different domains such as urban planning, intelligent transportation, demography, and emergency management. For example, we can learn how disease spread over space and time by studying connections accompany by human movement.Current methods can't handle large volume of complex spatial interaction data. It is difficult to extract the rich information lurking in large and complex spatial interaction data, and it is also hard to mapping, visualization and representing this dataset. The complexity of the dataset involve:(1) Big data volume: a moderate dataset normally has hundreds or thousands of locations, and easily has millions of origin-destination flows.(2) Multiple data spaces, such as spatial space, temporal space, multivariate space and graph space.(3) The modifiable unit problem(MAUP): the size of the units dramatically different, it may cause misleading result.(4) Multiple scales problem: Different patterns appears at varies spatial scales.Any unique method can't solve all the problems for such complex dataset. To cope with the problems mentioned above, we proposed a group of methods, which list below, to analyze and visualize large volume of spatial interaction data.Shared nearest neighbor similarity based spatial point clustering method, which group the spatial points into clusters to recognize natural gap within the dataset, and then extraction and mapping of the flow measures of clusters to understand the spatial distribution and temporal trends of movements. This is a data driven method, which can adapt to the skewed data distribution and form clusters with different size according to data density. This method can cope with data volume problem and MAUP.Shared nearest neighbor similarity based spatial interaction hierarchical clustering method extends the traditional agglomerative hierarchical clustering method to group spatial interaction data, which considers both origin and destination in determining the similarity of two spatial interactions. This method adopts shared nearest neighbor similarity to adapt skewed data distribution. Then we further investigate the temporal patterns based on the clustering results.Kernel density estimation based spatial interaction estimation model extends the spatial points' density estimation to spatial interactions, and select representative data according to the data distribution. This method can reduce the data size through highly abstraction. In addition, we can select adaptive bandwidth to cope with MAUP.Multi-scale spatial interaction mapping with kernel smoothing and selection improves the spatial interaction kernel density estimation model, which can extract multi-scale spatial interaction structures and facilitate the analysis and visualization of big spatial interaction data. It can discover inherent flow patterns for a given scale and naturally supports interactive and multi-scale flow mapping, where the pattern scale is related to parameter selection.These separate methods analyze the spatial interaction data from varies perspectives, and enhance our understanding towards complex dataset. This dissertation develops an integrated approach to examining large volume of spatial interaction from different perspectives, and achieves a holistic understanding by synthesizing different perspectives.