Database Reference
In-Depth Information
Item clustering can theoretically be applied to SNRS to select relevant friends
for inference. That is, by clustering similar items into different groups, homophily
effects among friends can be estimated based on the ratings of items within the same
group. Thus, it is possible for SNRS to identify friends who have a high correlation in
music CDs but a low correlation in restaurants. However, because the number of
items used to measure user similarity becomes less due to item clustering, the
estimated similarity values may not be as accurate as those without clustering.
A better way to select relevant friends is to utilize the semantics in social
relationships. Unfortunately, such semantics are not readily available in most
current OSNs. When a user indicates someone as a friend, it is not clear how and
why they became friends, and more importantly, we do not know in which aspects
they have homophily effects. Some OSNs ask how friends know each other, for
example, whether they were/are classmates or colleagues. Information like this
definitely helps us understand friend relationships. However, it is still too general to
have practical application in recommender systems. Instead, the semantics that we
really want to know from friend relationships should be more specific to the domain
of interest, in particular, the factors that influence users' buying decisions. For
example, in terms of dining, it would be ideal for SNRS to know whether two
individuals are friends because they have similar taste in food and/or similar
preference regarding the price of the meals. Although items have many character-
istics, the factors that matter in most users' buying decisions in choosing restaurants
may be limited to only a few common ones, such as food taste, nutrition value,
price, service, and environment. By carefully designed questionnaires or other
means of marketing analysis, such factors can be obtained. Thus, more semantics
regarding users' rating intentions and social relationships can be collected.
By providing users with the mechanism to rate items for each factor in their
buying decisions, for example, asking them to rate a restaurant based on food taste
and price, etc., recommender systems can improve the understanding of users'
rating intentions. Currently, most recommender systems ask users to input only
overall ratings which, however, consist of too many factors and are difficult to
understand. For example, when a user gives an overall rating of 4 to a restaurant, it
is not clear whether it is because of the food taste or the price of the meal. On the
other hand, if a user can provide ratings for those factors, the rationale behind the
overall rating can be well explained. Besides understanding users' rating intentions,
SNRS can also obtain the semantics in social relationships by asking users to rate
their friends on those factors. A user's high rating of a friend on a specific factor
means this user tends to agree with the friend's opinion, and together they have a
stronger homophily effect.
To predict a user's rating of a factor, SNRS needs to select those friends on
whom this user has a strong homophily effect regarding the same factor. The
selection of friends is thus dynamic according to the semantics in the factors of
user ratings. We call this process semantic filtering and denote SNRS with semantic
filtering as SNRS-SF. The framework of SNRS-SF is almost the same as that of
SNRS, except that immediate friend inference and distant friend inference are now
based on semantically filtered social networks.
Search WWH ::
Custom Search