The challenge's data and the videolectures.net's current RS

Questions, answers, discussions related to VideoLectures.Net Recommender System Challenge

The challenge's data and the videolectures.net's current RS

Postby nhanitvn » Sun Apr 24, 2011 5:51 am

I have a question: Was the usage data in file pairs.csv collected before or after the videolectures.net's recommender system was deployed? That means whether the site's recommender system has had any effect on the gathered usage data or not? It is because if that situation happens, you are asking the contestants to guess the ranking algorithm currently used at the site, not "to improve the website’s current recommender system", right?
nhanitvn
 
Posts: 3
Joined: Sat Jun 26, 2010 5:09 am

Re: The challenge's data and the videolectures.net's current RS

Postby ninoaf2 » Wed Apr 27, 2011 3:33 pm

Dear contestant,

VL.Net recommender system was running when the pairs.csv was collected,
but still we are not asking the contestants to guess the ranking
algorithm currently used at site. We will try to explain why.

Current VL.Net recommender system recommends 10 "most relevant" items.
But the list of recommended items is constructed from the whole item set and
not only from "new lectures". Due to the fact that "new lectures" have
been introduced to the system more recently, there exists significantly
less probability for them to appear in the top 10 items (the cold start
problem of the current recommender). Furthermore, we are asking recommended
list of size 30 for task 1. So our final answer is that the current recommender
did in fact influence, however not significantly nor crucially, the
viewing statistics of new lectures for task 1.

p.s. Sorry for delay, we had a collective Easter holiday

Best regards,
Organizers
ninoaf2
 
Posts: 25
Joined: Tue Mar 08, 2011 2:53 pm

Re: The challenge's data and the videolectures.net's current RS

Postby ninoaf2 » Mon May 02, 2011 11:06 am

Dear contestants,

Some additional comments regarding this question:

The given data (reflecting „cumulative“ users' behavior - pairs.csv) is actually a combination of following the recommendations, searching, and browsing the categories of interest. We do not follow how the user chooses his next video, since the user had to actually watch the lecture for a certain predefined amount of time for it to be entered into our database. We believe that this implies the user's interest in the lecture and that the viewing sequence thus indeed portray the user's interests.

So, it's not about modeling the current recommender system but rather about modeling the users' viewing sequences. In fact, the current recommender system proves ineffective in finding solutions for the challenge tasks.

Best regards,
Organizers
ninoaf2
 
Posts: 25
Joined: Tue Mar 08, 2011 2:53 pm


Return to VideoLectures.Net Recommender System Challenge

Who is online

Users browsing this forum: No registered users and 1 guest

cron