ICWSM Session 3: Ranking

CourseRank: A Closed-Community Social System through the Magnifying Glass

This paper discusses a social-media course selection site for Stanford University.  It combines official university course information, grade distributions, and course reviews with user generated comments, reviews, etc.  Has a course planning/recommendations, course clouds to find courses related to certain topics.

85% of Stanford undergrads use the site, way more than open community sites.

Using Tranactional Information to Predict Link Strength in Online Social Networks

Analyzed the Purdue Facebook network.  Generated different friend graphs for: Friends, Wall Posting, Pictures.  The Wall/Picture graphs have a much lower InDegree/OutDegree than the Friends network.  This may indicate that the wall postings may be a better indication of who your ‘real’ friends are.  I thought it was interesting that people had, on average 21 people writing on your Wall, but you only write on 7 people’s Wall.

Used the ‘Top Friends’ application as ‘truth’ of who your top friends are.  This paper compares 3 types of supervised learning algorithms, and four types of features to predict link strength through four separate experiments.

  • Experiment 1: Found 12 of 15 top features are network-tranactional type features, with wall information used best.
  • Experiment 2: Network transactional features had highest accuracy
  • Experiment 3: Compared link type.  Wall features had best accuracy.  Picture information quite bad
  • Experiment 4: Bagged decision trees had the best accuracy.  97% of performances comes from network transactional features

Network transactional features take into account transactions between person A to person B, moderated by # transactions A makes to everyone else.

RevRank: A Fully Unsupervised Algorithm for Selecting the Most Helpful Book Reviews

We use product reviews to make purchasing decisions.  Many reviews (on Amazon) are repetative, limited contribution, poorly written, unnoticed (and, as we learned this morning, confusing or plagerized).  Amazon has User Voting, which has some problems (imbalance vote bias, early bird bias, Winner Circle bias).

This work locates helpful reviews based on dominant concepts.  Term Dominance is similar to TF-IDF.

Examined 12,000 reviews of 5 books.  Compared algorithm to a human user vote and random sample.

RevRank did a good job of finding ‘helpful’ reviews, better than the other two conditions.