Personalized Fashion Advice - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 230

acomparable range of information levels. Preferences for jackets, shoes, and jeans
carry similar information compared to customers' ages and weights. We notice that
missing a buying history even the most sophisticated recommendation method will
struggle. This is mostly due to lack of sufficient information contained in the available
data.
8.7 Conclusion and Outlook
In this chapter, we presented the use case of a menswear online shop with the problem
of how to improve the rate of kept items among its customers. We defined this as
a constraint satisfaction problem. The user preferences are the constraints and our
goal is to build a recommender that learns those preferences based on item features
and recommends items which the user most likely will buy. We therefore introduced
a recommender in Sect. 8.5.2 where we presented a first version of a case-based-
reasoning approach using an overlay user model to make recommendations for a
single user. Results are comparable with other approaches, see Sect. 8.6 .Weshowed
that without any further analysis of the impact of different features, we already
reached good results. The next step here is to take a closer look at the different
features, which are currently all handled equally, and see if boosting some of them
improves the results. We also want to use more of the data, given in the profile, such
as shirt size, etc. as explicit constraints. Another direction would be using algorithms
like singular value decomposition, to deduce latent features, which could then be
used to further improve our CBR approach.
The question of solving the cold-start problem could not be answered satisfacto-
rily. To personalize, we need information about the user, more than currently available
in the user model. The purchase history is still the most important source of informa-
tion. As a cold-start user does not have such history, one possible solution could be
learning cluster of similar users and then use those combined purchase histories as
a start-up model for the cold-start user. The field of unsupervised machine learning
offers a variety of techniques which allow to cluster users as well as items. This
toolbox includes k -Means, hierarchical clustering, and component analysis amongst
others methods [ 15 ]. All these methods take users represented as feature vectors
as input. They seek to find patterns between these vectors. As a result, the sys-
tem can project new users onto clusters even without purchase history. Recently,
researchers have proposed representation learning [ 5 ]. Learning representations of
objects provides similar capabilities. The system may select the most similar repre-
sentations instead of particular clusters. Auto-Encoders represent a successful tech-
nique for representation learning. They consist of a multi-layered architecture of
neural nets. Each layer provides a more abstract representation of the target object.
This allows systems to learn hidden concepts. In the case of fashion recommen-
dation, we may detect taste patterns among subsets of customers. Additionally, we
plan to apply meta-learning techniques to further improve our approach's quality.
Bagging and Boosting represent two possible choices of meta-learning procedures.