Personal Attribute Judgments - Modeling Users' Experiences with Interactive Systems - page 30

Fig. 2.5 User segmentation map based on correlations between dissimilarity scores
81 attributes for designers and 95 for users were obtained in the study (6 to 11 per
participant). Two rounds of analysis were performed: an exploratory semantic clas-
sification of attributes, followed by a confirmatory analysis of the classification. For
the first round, sixteen semantically distinct attribute categories were first formed
out of the data. To minimize the researcher's bias, the naming of the attribute cate-
gories was restricted to choosing one of the attribute names that reflect this semantic
value. Subsequently, the first author and two additional experimenters independently
assigned every attribute to one of the sixteen categories (Table 2.3). Interrater agree-
ment (Fleiss et al., 2003) of the initial classification was satisfactory (K=0.72). All
sixteen categories were then classified into three overall classes: Effectiveness , Effi-
ciency ,and Emotional Appreciation (interrater agreement, K= 0.80).
During the confirmatory analysis of the classification, statistical consistency
across attributes within the same category was being sought. Attribute scores were
submitted to a cluster algorithm where Euclidean distances between attributes
were calculated and visualized in two or three dimensions by means of Multi-
Dimensional Scaling. Outlier attributes, i.e. ones that did not belong in the same
or a neighbor cluster to the one that were characterized primarily by attributes of
a given semantic category, were identified. The prospect of transferring the outlier
attribute to one if its statistically-neighbor categories was explored. If there was no
argument for a transfer to another category, the attribute was deleted.