Personalized Fashion Advice - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 214

model-based collaborative filtering approaches, content-based filtering methods, or
hybrid strategies thereof to recommend items [ 1 ]. These approaches use historical
data to compute similarities between items or users which are then used to recom-
mendmatching items. The reliance on historical data can cause problems, particularly
for new users or items. The system lacks sufficient information about them and fails
to compute similarities. This is also known as the cold-start problem [ 34 , 35 ]. The
problem often occurs in domains with high sparsity. High levels of sparsity mani-
fest as users buy few items. Thus, systems struggle to create accurate user profiles.
Simultaneously, most items are seldom bought inducing similar issues.
Recommender systems face manifold challenges as they seek to suggest outfits.
These challenges arise as recommending outfits differs from recommending movies.
This is due to the fact that even customers with like-minded tastes cannot necessarily
wear identical clothes. Jurca et al. [ 27 ] presented a study where they examined human
foot sizes in large scale experiment with 10,000 foot scans. The results showed a
high dispersion of foot widths within the various length classes. As the shoe sizes
are computed using the length classes, all people with the same foot length should
have the same show size. Since people with the same foot length have different foot
widths, they cannot wear identical models. Narrower models will hurt people with
wider feet. Conversely, wider models fail to adequately support people with narrower
feet. This fact about shoes generalizes to clothes, as people with the same height do
not necessarily have the same body measures and thus cannot wear the same clothes.
Cacheda et al. [ 10 ] argue that pure CF approaches do not work in domains like cloth
recommendations due to the fact that they do not exploit knowledge about the item
itself. Therefore, different approaches to recommendations for clothes are exploited.
There exist RS that recommend similar products based on image retrieval.
Hasan et al. [ 23 ] present a method for recommending similar products based
on their images. The presented method removes noise, e.g. body parts, from
the pictures before computing similarities. The method showed slight improve-
ments over other approaches [ 21 , 24 , 29 ] using image-based methods. Other
approaches aim for recommending clothes of a user's wardrobe based on context—
mostly time of day, weather and purpose. Shen et al. [ 44 ] use a scenario-oriented
approach where users have a virtual wardrobe where clothes have attributes such
as brand or type. The clothes are also user tagged with information such as
“I am going to…” or “I want to look more….” This information is used to rec-
ommend outfits when the user asks for an outfit to wear at the beach. Selected
clothes are used as user feedback. Yu-Chu et al. [ 51 ] present a study where they used
a Bayesian network to recommend clothes from wardrobes for a specific situation.
The small evaluation suggests that the Bayesian network can learn user preferences
and recommend clothes.
What we learn from the related work is that recommending suitable clothes to a
user is hard and still not satisfyingly solved. To be able to compute good recommen-
dations for clothes for a user, different aspects have to be taken into account:

• The user preferences for clothes: For example preferred colors, brands, or material.

Also price preferences should be taken into account.