Design Considerations for a Social Network-Based Recommendation System (SNRS) - Community-Built Databases: Research and Development - page 74

new items based on the user's past ratings of other items. The algorithms used in
recommender systems are usually two types - content-based filtering and collabo-
rative filtering . Let us define a target item as the item being considered for recom-
mendation, and a target user as the user who is receiving recommendations. In
content-based filtering, a target item is recommended to a target user if the item is
similar to the ones that the user liked in the past in terms of explicit content attributes
[ 1 , 2 ], while in collaborative filtering, a target item is recommended to a target user if
it is an item that has been liked in the past by people who are similar to this user.
Collaborative filtering finds users who are similar to a target user based on their
previous ratings of other items [ 3 - 5 ].
Despite all of the efforts above, recommender systems still face many challenges.
First, there are continuous demands for further improvements on the prediction
accuracy of recommender systems. Second, the algorithms for recommender sys-
tems suffer from many issues. For example, in order to measure item similarity,
content-based methods rely on explicit item descriptions. However, such descrip-
tions may be difficult to obtain for abstract items like ideas or opinions. On the other
hand, collaborative filtering has a data sparsity problem [ 6 ]. In contrast to the huge
number of items in recommender systems, each user normally rates only a few items.
Therefore, the user/item rating matrix is typically very sparse. It is difficult for
recommender systems to accurately measure user similarities from that limited
number of reviews. A related problem is the cold-start problem [ 6 ]. Even for a
system that is not particularly sparse, when a user initially joins, the system has no
reviews from this user. Therefore, the system cannot accurately interpret this user's
preference.
To tackle those problems, two approaches have been proposed [ 1 , 4 , 7 , 8 ]. The
first approach condenses the user/item rating matrix through dimensionality reduc-
tion techniques such as Singular Value Decomposition (SVD) [ 4 , 8 , 9 ]. By cluster-
ing users or items according to their latent structure, unrepresentative users or items
can be discarded, and thus the user/item matrix becomes denser. However, these
techniques do not significantly improve the performance of recommender systems
and sometimes even make the performance worse. The second approach “enriches”
the user/item rating matrix by: (1) using a default rating; (2) incorporating implicit
user ratings, for example, the time spent on reading articles [ 10 ]; (3) filling in with
half-baked rating predictions from content-based methods [ 7 ]; or (4) exploiting
transitive associations among users through their past transactions and feedback
[ 11 ]. These methods alleviate the data sparsity problem to some extent but still
cannot solve the cold-start issue. In this chapter, we plan to solve these problems
from a different perspective. Specifically, we propose a social network-based
recommender system (SNRS) [ 12 ] which predicts user interests by utilizing rich
semantic information in social networks, especially social relationships .
In a social network, two persons connected via a social relationship tend to have
similar attributes to each other. This is a fundamental property of social networks,
and it is also known as the homophily principle [ 13 ]. In product marketing, the
importance of social relationships has long been recognized [ 14 , 15 ]. Intuitively,
when we want to buy an unfamiliar product, we often consult with our friends who