Design Considerations for a Social Network-Based Recommendation System (SNRS) - Community-Built Databases: Research and Development - page 92

Since overall rating is not determined by a single factor, relevant friends for
predicting a user's overall rating cannot be selected in the same way as we did for
predicting fine-grained rating. To do so, we consider friends as those who are
selected as relevant friends for two or more of the most important factors. For
example, if a system considers that price and taste are the most important decision
factors in terms of dining, then a user's relevant friends for predicting overall
ratings are those users who are considered relevant friends for predicting the
user's ratings on price and taste. In the following section, we shall use this
approach.
4.6.1 Semantic Filtering Experiments
Since the Yelp dataset does not have fine-grained user ratings, we cannot use the
Yelp dataset for semantic filtering experiments. Therefore, we designed an experi-
ment for a graduate student class and collected a social network and fine-grained
user ratings from students.
The goal of this experiment is to predict students' ratings for reading online
articles. It was conducted in a graduate student class, “Intelligent Information
Systems”, with 22 students. We first selected 21 articles which focus mainly on
four topics: local news, US news, technologies, and culture. These articles all
contain strong opinions expressed by the authors. The article information and the
corresponding categories of these 21 articles are listed in Table 4.4 .
Before asking these students to review the online articles, we first collected their
demographic information, including gender, age, student type, employment, and
religion. We then asked the students to answer a set of survey questions related to
the articles as shown in Table 4.5 . These survey responses will provide prior
information about the students.
We then asked the students to review, as shown in Fig. 4.5 , every article and give
ratings (from 1 to 5, with 5 being the best) on the following four factors: (1)
Interestingness: Is the article interesting? (2) Agreement: How much do you
agree with the author? (3) Writing: Is the article well written? and (4) Overall:
Overall evaluation. The reason we include the first three ratings is because they
usually play the most important roles when we give an overall score to an article.
Since most students did not know each other before the experiment, it would be
difficult to form a social network from their original relationships. We therefore
divided the students into groups and let them get to know each other through
discussions of the articles. Specifically, we divided the students into three groups
twice. The first grouping was based on students' ethnicities, and the second
grouping was based on students' responses to the survey questions. The goal of
these groupings was to organize the students in such a way that the students in a
group were more likely to be friends after the group discussions. Each group then
had a meeting to discuss the articles. During the discussions, every student needed
to explain the reasons why s/he liked or disliked each article. Thus, the other