Interactive Comprehensible Data Mining - Ambient Intelligence for Scientific Discovery - page 61

old of support and confidence can be selected interactively. The main advantages
of this method are that it is visually simple, and the rules it discovers are not
limited to rectangular regions.
Purple Insight's MineSet product (originally marketed by SGI), contains a
number of discovery visualization methods. The most relevant is a graphical table
illustrating the co-occurrences of “true” values for pairs of binary attributes.
Again this is limited to discoveries with two antecedents. MineSet also provides
decision tree and class probability visualizations [36, 37].
Han et al.[40] present a framework for interactive knowledge discovery, and
a number of examples of the AVis system. AVis uses scatter plots and simi-
lar visualizations for data preparation and cleaning, a choice of discretization
boundaries for numerical data, and visualization of associations between pairs
of attributes.
Graph Based Approaches. Klemettinen's system [35] visualizes interactions
between attributes in sets of rules using an attribute graph. Nodes represent
attributes, and links their co-occurrence. The number of rules visualizable is
greatly limited by the use of arc thickness to convey statistical information, and
also by clutter and crossing of arcs.
Hao et al.[39] describe the “Directed Association Visualization” (DAV) sys-
tem for 3D force based graph visualization of two item association rules. The
term “association rule” is a common one, however, when only two items are in-
volved, “association” might also be used. The visualization aims to identify pairs
of products that are bought together, one of their examples being that 85% of
purchasers of computer printers also buy paper.
The system visualizes graphs in which nodes represent items and directed
links represent associations. Colors of links represent confidence, and support
is reflected in the distance between items. In addition, clusters of items can be
detected and wrapped in a surface. This reduces visual complexity and is a form
of discovery in itself, indicating that items in the cluster a similar.
An example in [39] shows a graph with approximately 180 nodes and 1400
links, summarizing 250,000 transactions. This can be visualized in real time.
Attribute Utility. Bruzzese and Davino [38] present a method for the visu-
alization of multiple antecedent association rules, for pruning of rule sets and
identification of relevant attributes. Multiple rules can be plotted on parallel
co-ordinates using “Item Utility” for each of their attributes, based on the con-
tribution of the attribute to rule accuracy.
Hofmann et al [42] present a method for visualizing the support and confi-
dence of association rules with multiple conditions using mosaic plots and double
decker plots. These show bar chart like representations with an area proportional
to support, partially colored to reflect confidence.
Clustering. Discoveries Bruzzese and Davino [38] use Multiple Correspondence
Analysis to project rule antecedents and consequents into a 2D space, where clus-
tering is observed. Tsumoto and Hirano [43] calculate rule similarity measures