Design Considerations for a Social Network-Based Recommendation System (SNRS) - Community-Built Databases: Research and Development - page 100

This is because item clustering makes the number of items in each cluster smaller;
thus, the coverage has a greater impact due to the sparseness. On the other hand, we
notice that the coverage of SNRS and SNRS-SF are rather insensitive to the
sparseness. It starts to drop after the sparseness exceeds 0.8 and with a slower
rate than CF, CF-C and CF-K. For example, even when the sparseness is 0.9, the
coverage of SNRS and SNRS-SF is still 0.789 and 0.645, as shown in Fig. 4.9a . The
coverage of SNRS-SF is slightly lower than that of SNRS because of fewer friends
for each student.
4.7 Trust in SNRS
SNRS implicitly assumes that all users in the social network are trustworthy.
However, in most recommender systems, this assumption is not necessarily valid.
In this section, we shall discuss two trust issues and propose how SNRS can be
extended to handle them.
4.7.1 Shilling Attacks from Malicious Users
Intrigued by incentives, malicious users in recommender systems can purposely
provide false reviews to promote their own products or attack similar products of
competitors. For example, in a user-based collaborative filtering system, a mali-
cious user can simply fake a set of reviews with the exact same ratings as those of a
target user. Then this malicious user will be considered as the most similar user of
the target user. If malicious users want to promote their own products, they can
simply give the products high ratings, and these products will have a high chance of
being recommended to the target user. This problem is known as shilling attacks.
The main reason that shilling attacks can become threats is that recommender
systems rely too much on user rating similarity but overlook another important
aspect, i.e., trust among users. Some studies have introduced explicit trust defined
by users [ 27 ] and implicit trust inferred from user ratings [ 26 ] and have shown some
improvements. However, unlike these approaches, SNRS is in essence built on
trust, and thus it is able to handle shilling attack problems. Instead of using rating
similarity, SNRS makes predictions by exploiting homophily among friends. Since
users know their friends themselves, it is less likely for them to add malicious users
as friends. If a user suspects that some friends may be potential malicious users, the
user can remove those friends from the friend lists. Thus, in SNRS, the fact that two
users are friends indicates the trust between them. In addition, with the capability of
rating friends on each factor of a user's buying decisions (as discussed in Sect. 4.6 ),
SNRS not only knows who are friends, but also on which aspect of buying decisions