VL.Net newsletter (2011-07-04 16:55:50)
Dear VL.Net challenge contestants, It is now few days before the closing of the competition! There are couple of announcements that we want to make and one slight change in the schedule of the challenge. Challenge: We plan to close the challenge 12 hours earlier, at 11.59 AM on Friday, 8th of July, rather than at 11.59PM (midnight), as scheduled originally. We hope that this will not trouble anyone of you; the change is made purely for technical reasons. We plan also to release results on the day after – i.e. 9th of July. Workshop (5th of the September, Athens ECML/PKDD 2011): We have prepared a website for the workshop related to the challenge that should be held on. From the beginning of the next week all the changes and news related to the workshop (submission of papers, key dates, etc) will be announced at http://lis.irb.hr/challenge . Please, note that we expect you to submit a paper describing your solution in order to receive the award. The submission of a paper is also a prerequisite to be announced as the winner of the challenge and to present the results at the ECML/ PKDD 2011 workshop. Due to the tight schedule and closeness of the workshop/conference, we will try to finish reviewing process earlier, and leave more time for you to organize your travel to Athens! So, please follow the changes on the workshop site. Good luck in the time left for competing ! Organizers of VL.Net
Categories - contains cycles (2011-06-17 08:55:12)
In readme.txt file, categories was described as a DAG (direct acyclic graph), but this is not true. It contains cycles ! We have updated the description in the readme file.
Snapshot time-additional information (2011-06-03 10:07:08)
Snapshot time or the moment when the lecture (co)viewing frequencies were taken (for all lectures!) was (July 2010). 1.7.2009 was taken in order to have reasonable training/test split for the task 1. Note however, that training/test split is both "vertical" (all test lectures are published after the 1.7.2009. !) and "horizontal" (approximately half of the lectures published after 1.7.2009. are in the training set!). By this split we wanted to provide samples in the training set that can be used to learn the temporal impact on lecture (co)viewing frequencies.
Readme.txt file was updated to new version (2011-05-30 15:31:15)
One of the participants has found that two of the challenge datasets are not described accurately, with respect to their actual content. Actually events, as well as lectures, have been related to categories and authors. Therefore the files categories_lectures.csv and authors_lectures.csv contain not only category-lecture and author-lecture relations, but also category-event and author-event relations respectively. As the lectures and events have different IDs, there are no errors, so the files stay the same (we did not want to change their names either, not to confuse participants). We did however change/correct the description of these two files in README file. For more details, visit our forum under the topic "Events".
Simple pairs solution (2011-05-19 10:50:17)
For each triplet (id1,id2,id3) we can generate simple ranked list of lectures by the following procedure: 1.) Identify all links to other lectures that are related to the particular triplet (data in pairs.csv). 2.) Sum all the co-viewing frequencies related to these links 3.) Rank the lectures according to the summed frequencies - and this is the simple solution (for one triplet in the test set!) This simple CVS solution achieves leaderboard score of 0.12684 on Track 2.
Maximum number of submissions (2011-05-17 16:28:14)
Maximum number of submissions for Track 1 and Track 2 has been increased from 30 to 60.
Baseline random solution (2011-04-29 10:13:20)
Baseline random solution recommendation lists and pooled viewing sequences for track 1 and track 2 were created by random sampling set of lectures_test and set of lectures_train. Examples of submission files for baseline random solution can be found under the site for track 1 and track 2.