|dc.description.abstract||The work in this dissertation was motivated by the desire to more fully understand
the results of a card sorting study of facial photographs. Twenty-five participants were
asked to sort 356 facial photographs (178 of Caucasian males and 178 of First Nations’
males) into an unconstrained number of piles according to their judgments of similarity.
Researchers have used card sorting to understand different concepts, but not facial
similarity. Therefore, the work presented in the dissertation is novel because it takes
an existing method and applies it to a new context and adapts analysis tools to this
Pairs of photos in the same pile were deemed to be judged “similar” and those in
different piles were deemed to be judged “dissimilar”. There are 63,190 pairs possible
from 356 photos. It is clear that participants did not, nor could not, make all these
pairwise comparisons directly.
The study was executed, analyzed, and described by other researchers at the University
of Regina, but there remain unresolved questions from the data that it produced.
Some participants made very few piles which others made very many: should the information
provided by each participant be treated equally? Is there enough information
in the sorting of the cards to uncover how participants have judged facial similarity,
with different participants using possibly different strategies? Perhaps it would be more
productive to work with a smaller number of photos, but how should these photos be
selected and what number of photos is neither “too many” nor “too few”?
If the majority of participants agree that a photo pair is similar or dissimilar, that
pair may not help to discern different strategies that participants may be using. Therefore,
it is possible to consider the data in terms of a three-way decision: if 16 or more
of the 25 participants judged a pair to be similar, the pair is labelled as Similar. If
9 or fewer of the 25 participants judged a pair to be similar, the pair is labelled as
Dissimilar. Finally, if between 10 and 15 participants judged the pair to be similar, the
pair is labelled as Undecided.
In order to explore the questions about information quality, the probability of each
pair of photos was calculated, modelled as the first two photos drawn from a deck
without replacement. If a participant placed the pair with many other photos, either
together in one pile or apart in two piles, the probability of that pair was large. Alternatively,
the probability of the pair was small if it was placed with few other photos.
The probability of the pair is hypothesized to be an indicator of confidence: a small
probability means high confidence and a large probability means low confidence. If the
Undecided group contains photo pairs about which some participants are very confident
of similarity and others are very confident of dissimilarity, these pairs may be most
useful to study further.
The eigenface method of evaluating facial similarity was used to provide a reference
for the judgements made in the card sorting study. The Euclidean distance between
each pair projected photos was used as a surrogate for the similarity judgements: pairs
with small intra-pair distances were deemed to be judged “similar” and pairs with large
intra-pair distances were deemed to be judged “dissimilar”.
A second study was conducted, based the analysis of the first, that asked forty-three
participants to rate the similarity of selected photo pairs on a scale from 0 (Similar) to
100 (Dissimilar), with the midpoint representing Undecided.
Analysis of the second, pairwise, study found agreement with the first, card-sorting,
study. The application of the 3-way decision framework also proved valuable.
In conclusion, the work has provided a means, previously unavailable, to understand
the importance of various photo pairs in participants’ judgment of similarity.||en_US