[talks] A Suleimenov MSE talk

[Melissa M. Lawson]

Arman Suleimenov will present his MSE talk on Thursday May 3 at 10AM 
in Room 402.  The members of his committee are Andrea LaPaugh, advisor, 
and Adam Finkelstein, reader.  Everyone is invited to attend his 
talk.  His abstract follows below.
----------------

Title
Twitter News: harnessing Twitter to build an article recommendation system

Abstract
With more than 140 million active users and 340 million tweets a day
(as of March 2012), Twitter presents a great source of recommendation
knowledge for articles shared on the platform. Collaborative filtering
methods (based on matrix factorization or neighborhood-based
algorithms) suffer from extremely sparse user-article matrix as well
as cold start problem. Content-based filtering requires us to fetch
the text/title of an article given the url which (given the fact we
don't limit ourselves to a small subset of well-formatted news sites)
becomes a challenge in itself. In this work, we analyze 836 Twitter
users from the technology and entrepreneurship space with 78,508 links
shared by them. We explore and evaluate the following (old and novel)
techniques for an article recommendation engine: bag-of-words Naive
Bayes, vector-to-vector similarity where the user vector is
constructed from the text of the tweets produced, topic-modeling based
approach where we learn the topic distribution for each article and
thanks to that reduce dimensions of the user-article matrix, model
where apart from relevance and novelty we take into account connection
clarity and transition smoothness between articles, content-boosted
collaborative filtering (with probabilistic matrix factorization)
where pseudo user-ratings are created as well as the hybrid model of
some of the best-performing techniques above.