Database Reference
In-Depth Information
Fig. 4.6 Form for reviewing a group member
reviewed every article in this experiment, the student/article rating matrix is
completely filled in. Thus, the data sparsity of the dataset is 0. Compared with the
extremely sparse data in the Yelp dataset, the fully observed students' ratings in this
experiment allow us to measure the performance of SNRS-SF under the sparseness
test in a full range.
4.6.2 Experiment Setup
We implement the following methods for performance comparison.
Collaborative filtering ( CF ). When predicting fine-grained student ratings, we
select similar users based on their fine-grained ratings on all articles.
Collaborative filtering with item clustering based on item category ( CF-C ). In
this method, 21 articles are clustered into four groups according to their categories.
To predict a student's rating of an article, we measure his/her Pearson coefficient
with other students based on their ratings of the articles in the corresponding group.
Collaborative filtering with item clustering by running K-means on students'
ratings ( CF-K ). In this method we use K -means to cluster 21 articles based on their
rating similarities. Since there are four types of ratings (three fine-grained ratings
and an overall rating), we have four sets of clusters. In each set, the articles are
clustered into three groups. Similar to CF-C, to predict each student's rating of an
article, we measure Pearson coefficient of student pairs based on their ratings of the
articles in the same cluster.
SNRS . In this method, we consider student V as student U' s friend if U rates V
with a value 3 on at least one of the three factors. The social network of these
students is shown in Fig. 4.7a . Each node in the figure represents a student. If student
Search WWH ::




Custom Search