Design Considerations for a Social Network-Based Recommendation System (SNRS) - Community-Built Databases: Research and Development - page 90

Table 4.3 Comparison of the performance of SNRS with and without distant friend inference.
MAE Coverage

With distant friend inference 0.727 0.807

Without distant friend inference 0.683 0.364

friends. This observation leads us to further study the role of distant friends in
SNRS. Specifically, we compared the performance of SNRS with and without
distant friend inference in a tenfold cross-validation. The experimental results are
shown in Table 4.3 . From these results, we can see that by considering the
influences from distant friends, the coverage of SNRS is increased from 0.364 to
0.807, which is equivalent to a 122% improvement. However, the improvement is
achieved at the cost of a slight reduction in the prediction accuracy. In our experi-
ments, the MAE increases from 0.683 to 0.727, which is only a 6.4% difference.
This is consistent with our intuition that the impact from distant friends is not as
direct as that from immediate friends, and certain errors will be inevitably intro-
duced when considering distant friends, but compensated for by the enormous gain
in the coverage.
Our experimental results revealed that social network information can be used to
improve the performance of recommender systems. In the next section, we shall
discuss how to remedy some issues in SNRS that are caused by heterogeneities in
social network information.

4.6 Semantic Filtering of Social Networks

Social networks contain rich semantics that are valuable to SNRS. However, this
information can also interfere with the predictions of SNRS if not carefully applied.
In this section, we discuss the issues of SNRS caused by the heterogeneities in social
relationships and items.
Friends exhibit similar behaviors when selecting items; however, the favorite
items that friends have in common depend on their social relationships. For example,
two friends who have common interests in music CDs may not necessarily agree
about their favorite restaurants. Therefore, to find the favorite restaurants, we should
not consider friends who share only a common preference in music. Instead, an
appropriate set of friends needs to be selected according to the target items. In fact,
we considered this issue when performing experiments on the Yelp dataset. Rather
than considering all friends listed in users' profiles, we keep only those friends who
also have an interest in food. For example, even though two real friends may have
reviewed many common hotels on Yelp, they are not necessarily considered as
friends in SNRS unless they both have reviewed restaurants. However, this solution
is still a gross approximation, because even within the domain of restaurants, friends
can be further grouped based on their opinions on different food categories, price
range, restaurant environment, etc.