![](bg2.png)
selling prices) is revealed. In addition anonymisation
algorithms (data modification and privacy
preserving) are used.
The virtual interior designer (VID) is
AsIsKnown’s customer consulting and ordering
system. Products are displayed according to
characteristics and preferences of the customer as
different style worlds in showrooms, where the user
can try out different product combinations.
Consumer behaviour is logged and stored in a data
warehouse which also stores (aggregated) ordering
data from producers. The association miner accesses
the data warehouse and enables market basket
analysis on these data, to analyse frequent
combinations of products and features. The analysis
is performed by expert users of the trusted third
party. The trends are formulated as association rules
into the rule editor based on a commonsense
ontology. The Smart Profiler (SP) identifies
preference changes of customer groups in these
rules, e.g. which age-group prefers which design
style or which colour is favored by which group of
customers and generates new style worlds, to
advance customer related product proposals in the
VID. Trend rules are also provided to producers as
trend reports. See (Becks, 2007) for the overall
design of AsIsKnown’s Trend Analyser.
2.1 Detecting Association Rules
For association mining, we use the flexible
explorative visual data mining tool InfoZoom
(Spenke, 2001), developed at Fraunhofer FIT. It uses
special information visualisation and interaction
techniques to support the user in analysing and
gaining a deep understanding of the data. The
interaction possibilities offered are based on the
Information Seeking Mantra (overview, zoom and
filter, details on demand) (Shneiderman, 1996).
InfoZoom displays data with attributes as rows
and their values as columns. In the AsIsKnown data
warehouse each entry represents a transaction of a
customer visit in the VID. Selection of values
restricts the table to this value (zoom in). Clicking
on the arrow outline right of an attribute sorts the
table by that attribute. The user can see all values of
the attributes at a glance by using the so called
“compressed mode” (Figure 1). This view causes
that adjacent cells, having identical values, namely
the attribute values presented in columns, are
merged. The width of each cell indicates the number
of objects with this specific value. Cells with
numeric values, too small to be labelled with the
related value, are represented through a horizontal
line. The level of that line reflects the height of the
value.
Interesting dependencies between different
attributes can be identified by performing
consecutive sorts in the “compressed mode”.
Suppose we would like to verify if the attribute
age=20+ and the attribute style preferences do have
an interesting correlation. Figure 1 (top) shows the
customer attributes and their values for all
transactions of all customers. The attributes have
been sorted by style preference and age. Each style
preference is preferred by some customers of every
age. However, customers of age older than 20 seem
to prefer mainly the style pop. Zoom in (double
click) this customer group, as it is shown in Figure 1
(bottom) confirms this observation. Measurements
such as support, confidence, lift, etc., typically
calculated in association mining, are done based on
performing counts of different item groups. These
counts are calculated by defining new, “derived”
attributes in InfoZoom. These attributes store
functions to perform calculations on already existing
attributes.
The following steps are performed to determine the
measurements for the rule: Count of basic
population. Zoom in the body, i.e. double click at the
cell age = 20+. Count of the items for which holds
age = 20+. Zoom in style preference = pop, the
zoom of the body is still fixed, meaning that we can
now count the number of items age= 20+
∪
style
preference = pop, now. Zoom out the body, hence
double click on age= 20+. Count of items for which
the head (style preference = pop) holds.
3 EVALUATION
In this section we report on the evaluation of the
Association Miner. The evaluations were based on
questionnaires and “thinking aloud” protocol. The
aim was to evaluate the procedure of rule detection
and measures and to evaluate the quality of the
results and their usefulness for the producers. The
evaluation was performed in two steps.
According to the relevant questions of the producers
(Which type of customer buys what? Which kind of
products are bought in conjunction?), we developed
a questionnaire with different scenarios. In a second
evaluation step the calculated rules were evaluated
through producers, concerning their usefulness. This
step was performed in comparison to the fully
automatic association mining tool WEKA (Witten,
2005).
EXPLORATIVE ASSOCIATION MINING - Cross-sector Knowledge for the European Home Textile Industry
501