Design Considerations for a Social Network-Based Recommendation System (SNRS) - Community-Built Databases: Research and Development - page 75

have already experienced the product, since they are those whom we can reach for
immediate advice. When friends recommend a product to us, we also tend to accept
the recommendation because we consider their inputs as trustworthy. Many mar-
keting strategies, such as Hotmail, that leveraged social relationships have achieved
great success [ 16 ]. Thus, social relationships play a key role when people make
decisions about products, and it is the basis for constructing SNRS.
The recent emergence of online social networks (OSNs) gives us an opportunity
to investigate the role of social relationships in recommender systems. With the
increasing popularity of Web 2.0, many OSNs such as Myspace.com and Facebook.
com have emerged. Members in those networks have their own personalized space
where they not only publish their biographies, hobbies, interests, blogs, etc., but
also list their friends . Here, friends are defined in a general sense: any two users who
are connected by an explicit social relationship are considered as friends. In reality,
they can be family members, buddies, classmates, and so on. In addition, we define
immediate friends as friends who are just one hop away from each other in a social
network graph, and distant friends as friends who are multiple hops away. OSNs
provide platforms where people can place themselves on exhibit and maintain
connections with friends. As OSNs continue to gain more popularity, the unprece-
dented amount of personal information and social relationships can promote social
science research which was once limited by the lack of data. In this chapter, we
design a new paradigm of recommender systems by utilizing such information in
social networks.
While the benefits of utilizing social network information in recommender
systems can be significant, how to materialize such an idea is especially challenging
considering the complexity of social networks. Many challenging questions can be
raised in this context. In particular, we investigate the following questions: (1) Does
homophily really exist when friends rate items? (2) How to effectively use different
types of social network information to make better predictions? (3) If predictions
rely on the opinions of immediate friends, what if a target user has no immediate
friend who has reviewed the same target item? (4) How does SNRS handle hetero-
geneities in social networks such as different types of friend relationships? (5) How
does SNRS handle situations where the reviews from immediate friends are not
trustworthy?
The remainder of the chapter is organized as follows. First, in Sect. 4.2 ,we
give a background of collaborative filtering algorithms. Then, in Sect. 4.3 ,we
introduce the dataset that we crawled from a real online social network, Yelp.com.
We will study this dataset to determine whether homophily exists when friends
rate items. In Sect. 4.4 , we present our SNRS system. Following that, we evaluate
the performance of SNRS on the Yelp dataset in Sect. 4.5 , focusing on its
prediction accuracy and coverage. In Sect. 4.6 , we propose to further improve
the prediction accuracy of SNRS by applying semantic filtering for social net-
works. We design a student experiment in a graduate class to validate its effec-
tiveness. In Sect. 4.7 , we propose extensions of SNRS to handle the trust issues
caused by users with unreliable domain knowledge. Finally, we review related
work in Sect. 4.8 .