 Abstract | Association mining, the discovery of interesting inferences from within a dataset,
is ultimately subjective as only the user can assess the practical usefulness of
an inference. To this effect, an association mining system harnesses the user's
perceptual capabilities and the computer's processing power to improve the quality of a set of inferences. Although current association mining systems tightly
involve the user within the pre-processing and presentation stages, the analysis stage of the association mining process remains relatively autonomous and
opaque. This lack of user involvement constrains domain space exploration and
subsequent inference derivation, potentially reducing inference quality, due to the
lack of user-computer synergy.
The theory of guided association mining and its realisation represents a timely
and logical step in the progression of association mining research. Early research
focused upon algorithmic efficiency, addressing issues such as I/O reduction and
scalability, however this seems to have reached a point of diminishing return.
The research focus has therefore shifted to improving result quality, or improving inference interest, rather than the speed at which the results are generated,
including areas of research such as measures of interestingness and semantic inclusion. However, these areas of research which attempt to incorporate domain
knowledge within analysis, fall short of providing user-computer synergy as the
specified constraints are statically included within an automated process. Given
this static constraint inclusion, the derivation of quality inferences often requires
an iterative analysis process, whereby a set of quality inferences is converged upon
through iterative constraint refinement.
This thesis argues that by maintaining the user-computer synergy during analysis, the quality of discovered inferences can be improved. This is achieved
by opening the opaque 'black box' analysis process and providing functionality through which the user can interact, and subsequently guide, domain space
exploration. Thus by enabling the user to dynamically focus exploration upon
concept areas of specific interest, the quality of the derived inferences will improve.
This thesis addresses the next step in providing analysis synergy by enabling
the user to dynamically refine constraints during analysis instead of between
analysis iterations. To this end a guided mining architecture is proposed that
merges the currently accepted knowledge discovery architecture with the model-view-controller architecture, enabling analysis synergy through the provision of
a transparent and interactive analysis environment. Furthermore this thesis also
makes novel contributions to the foundation fields of analysis and rule presentation, by way of an incremental closed-set association mining algorithm and an
association visualisation technique that accommodates hierarchical semantics. |
