Analyzing Personal Attribute Judgments - Modeling Users' Experiences with Interactive Systems - page 52

The average model, although it accounts for more attributes than each of the
diverse models, fails to predict semantically similar attributes. Thus, replicate at-
tributes (i.e. attributes pointing towards the same direction with overlapping confi-
dence ellipses) exist only for two attribute categories, namely “Fast access to in-
formation” and “Supports search” . The websites of the university of Mnchen and
Aachen are differentiated from the remaining ones as websites that provide fast ac-
cess to information, while the second attribute differentiates mainly the websites of
the universities of Mainz and Frankfurt as the ones that do not support searching.
These two attributes are present also in two of the three diverse models, model
1 and model 3. Model 1 further differentiates the websites of Aachen and Mnchen
as having a “graphical layout” , the website of the university of Aachen is mainly
differentiated from all others as a website that does not “refer to student life” .On
the contrary, model 2 provides a different insight. It reveals that the websites of
the Universities of Mannheim, Frankfurt and Mainz put “less emphasis on achieve-
ment” . The set of websites can also be split in two groups based on the amount of
information that they provide to the user.

3.4 Discussion

This chapter challenged traditional “averaging” practices in the analysis of repertory
grid data. It was illustrated that only 15% of the attributes were adequately predicted
by the average model and that the remaining attributes had minimal contribution to
the configuration space, as shown by the high correlation (r=0.99) of the two spaces,
i.e. one employing all attributes and one employing only the 15% of attributes be-
ing adequately predicted. The diverse models accounted for more than double the
information accounted by the average model (38 attributes out of all 118), how-
ever a substantial amount of attributed remained non-modeled. One might wonder
whether the information contained in the remaining attributes is truly non-modeled,
or whether these attributes just do not differentiate strongly between the stimuli and
thus do not contain any substantial information.
Figure 3.5a displays the variance in the attribute's ratings versus the accounted
variance by the three diverse models for two groups of attributes: the ones that are
adequately predicted by the three diverse models (38) and the remaining ones. One
may note that the hypothesis stated above does not necessarily hold; no substantial
differences may be found in the variance of the modeled and the non-modeled at-
tributes. It thus becomes evident that while the diverse models result in a substantial
increase in the information being modeled, still a substantial amount of attributes
containing high variance remain non-modeled.
One potential reason for this might be a limitation that we pointed earlier, the
use of heuristics in judging the goodness of fit of a model for a given attribute's
data. An alternative procedure proposed by (Martens, 2009a) employs an iterative
process, grounded on Singular Value Decomposition (Eckart and Young, 1936), that
aims at optimizing the overall accounted variance (R 2 ) for all attributes in the data.