An Investigation in Efficient Spatial Patterns Mining

The technical progress in computerized spatial data acquisition and storage results
in the growth of vast spatial databases. Faced with large amounts of increasing spatial
data, a terminal user has more difficulty in understanding them without the helpful
knowledge from spatial databases. Thus, spatial data mining has been brought under
the umbrella of data mining and is attracting more attention.
Spatial data mining presents challenges. Differing from usual data, spatial data includes
not only positional data and attribute data, but also spatial relationships among
spatial events. Further, the instances of spatial events are embedded in a continuous
space and share a variety of spatial relationships, so the mining of spatial patterns demands
new techniques.
In this thesis, several contributions were made. Some new techniques were proposed,
i.e., fuzzy co-location mining, CPI-tree (Co-location Pattern Instance Tree),
maximal co-location patterns mining, AOI-ags (Attribute-Oriented Induction based on Attributes’
Generalization Sequences), and fuzzy association prediction. Three algorithms
were put forward on co-location patterns mining: the fuzzy co-location mining algorithm,
the CPI-tree based co-location mining algorithm (CPI-tree algorithm) and the orderclique-
based maximal prevalence co-location mining algorithm (order-clique-based algorithm).
An attribute-oriented induction algorithm based on attributes’ generalization sequences
(AOI-ags algorithm) is further given, which unified the attribute thresholds and
the tuple thresholds. On the two real-world databases with time-series data, a fuzzy association
prediction algorithm is designed. Also a cell-based spatial object fusion algorithm
is proposed. Two fuzzy clustering methods using domain knowledge were proposed:
Natural Method and Graph-Based Method, both of which were controlled by a
threshold. The threshold was confirmed by polynomial regression. Finally, a prototype
system on spatial co-location patterns’ mining was developed, and shows the relative
efficiencies of the co-location techniques proposed
The techniques presented in the thesis focus on improving the feasibility, usefulness,
effectiveness, and scalability of related algorithm. In the design of fuzzy co-location
Abstract
mining algorithm, a new data structure, the binary partition tree, used to improve the
process of fuzzy equivalence partitioning, was proposed. A prefix-based approach to
partition the prevalent event set search space into subsets, where each sub-problem can
be solved in main-memory, was also presented. The scalability of CPI-tree algorithm is
guaranteed since it does not require expensive spatial joins or instance joins for identifying
co-location table instances. In the order-clique-based algorithm, the co-location table
instances do not need be stored after computing the Pi value of corresponding colocation,
which dramatically reduces the executive time and space of mining maximal colocations.
Some technologies, for example, partitions, equivalence partition trees, prune
optimization strategies and interestingness, were used to improve the efficiency of the
AOI-ags algorithm. To implement the fuzzy association prediction algorithm, the “growing
window” and the proximity computation pruning were introduced to reduce both I/O and
CPU costs in computing the fuzzy semantic proximity between time-series.
For new techniques and algorithms, theoretical analysis and experimental results
on synthetic data sets and real-world datasets were presented and discussed in the thesis.