Design Considerations for a Social Network-Based Recommendation System (SNRS) - Community-Built Databases: Research and Development - page 101

two friends trust each other. Therefore, the risk of shilling attacks can be further
reduced.
4.7.2 Misleading by Friends with Unreliable Knowledge
It is worth pointing out that malicious users are not the only cause of the trust
problems in recommender systems. Due to limited knowledge of target items, users
who are trustworthy may still provide inaccurate reviews that do not truly reflect the
truth of the items. Since SNRS relies on friends' opinions to make predictions, those
inaccurate reviews will produce misleading recommendations of SNRS. For exam-
ple, Alice has a taste similar to her friend Bob for Italian food, but Bob seldom goes
to Thai restaurants. In this case, even though Bob is trustworthy to Alice, his
opinion on Thai restaurants may not be so useful. To the best of our knowledge,
little research if any has been devoted to solving problems caused by users with
unreliable knowledge.
The key problem in this example is that the quantification of user correlations is
based on all of the common items that every pair of users has reviewed, and it does
not consider differences in item categories, like the difference between Thai food
and Italian food. Conceptually, SNRS can solve this problem by introducing item
clustering, for example, the clustering of items based on their contents or rating
similarities (as shown in Sect. 4.6 ). Therefore, SNRS can quantify two friends'
homophily effects based only on the items within the same cluster as a target item.
However, in practice, this solution may not work well because item clustering will
make the data sparser.
To solve this problem, we propose to relax item categories when quantifying
homophily effects. Instead of treating different categories as totally isolated, we
consider some of them as still related based upon domain knowledge such as item
taxonomies. For example, assuming we know from item taxonomies that Chinese
food and Thai food are all Asian food, thus Chinese food is more similar to Thai
food comparing with Italian food. Therefore, even though we cannot use Bob's
preference for Italian food, we can still leverage his preference for Chinese food, if
any, to guide the recommendation to Alice about Thai food.
In particular, we model item taxonomies into a type abstraction hierarchy (TAH)
[ 28 ]. A TAH is often used to facilitate approximate query answering. It has a tree
structure representing objects at different levels of abstraction. The leaf nodes in a
TAH are usually the most specific objects. As the level goes up, the nodes in the
TAH become more general. In Fig. 4.10 , we show a sample TAH generated from
food taxonomy. Let us refer to the leaf nodes in a TAH generated from item
taxonomy as item categories, such as Thai food and Chinese food. Thus, every
item in the system can be mapped into a corresponding leaf node according to its
category.
Let us assume that a target item belongs to category T ; C refers to each category
in item taxonomy; I C is the set of items of category C . We define W CT as the