ISMLL group wins this year's ECML/PKDD Discovery Challenge
The University of Hildesheim will not be easily forgotten after this year's ECML/PKDD Discovery Challenge, a yearly competition held together with the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD). Among the ten different research teams from seven different countries, two teams from the Information Systems and Machine Learning Lab (ISMLL) of the University of Hildesheim, Steffen Rendle as well as Leandro Balby Marinho and Christine Preisach, scored 1stand 2ndplace in challenge's task 2 respectively (see http://www.kde.cs.uni-kassel.de/ws/dc09/results ).
This year's discovery challenge consisted of three tasks in the area of social bookmarking, a popular practice in the Web where social bookmarking sites are distinguished by communities of practice, such as users sharing photos in Flickr, web pages and list of scientific publications in BibSonomy, and/or music in Last.fm. Common to all these sites is that users can attach tags - freely chosen keywords - to their uploaded resources (e.g., photo in Flickr, sound tracks in Last.fm, etc.), which help users to organize their resources by topic.
All tasks of the challenge targeted the support of the user during the tagging process by recommending tags, i.e., after a user has uploaded a resource to the system, the recommender algorithm must recommend a list of tags that are most likely to be used by the respective user to the respective resource. A complete dataset of the social bookmark and publication sharing system BibSonomy was provided for the challenge. A training dataset for all tasks is provided at the beginning of the competition, while the test dataset consisting of user/resource pairs with “unknown tags” is just released 48h before the deadline. For all tasks, given a test user/resource pair, the competitors should develop methods that try to predict a list of tags that are as close as possible to the real list of tags, which only the organizers know. The ISMLL teams took part on the task 2, which was especially intended for methods relying on the graph structure of the training data only, i.e., there were no new tags, resources or users in the test dataset.
The ISMLL group of the University of Hildesheim, headed by Prof. Dr. Dr. Lars Schmidt-Thieme, has been actively actuating on the areas of recommender systems and a broad range of machine learning topics, “know-how that proved to be one of the keys for achieving such successful results in the challenge”, according to Leandro Balby Marinho. The two teams responsible for this remarkable achievement are:
Steffen Rendle and Lars Schmidt-Thieme, with the approach entitled “Factor Models for Tag Recommendations in Bibsonomy”;
Leandro Balby-Marinho, Christine Preisach, and Lars Schmidt-Thieme, with the approach entitled “Relational Classification for Personalized Tag Recommendations”.
Another positive side effect resulting from the broad spectrum of machine learning topics covered in the ISMLL group, is that, despite the shared insights among the teams during the competition, both teams developed independent and distinct approaches, which end up presenting the research community with a suit of methods covering different and interesting aspects of the problem.
The University of Hildesheim will not be easily forgotten after this year's ECML/PKDD Discovery Challenge, a yearly competition held together with the European Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML/PKDD). Among the ten different research teams from seven different countries, two teams from the Information Systems and Machine Learning Lab (ISMLL) of the University of Hildesheim, Steffen Rendle as well as Leandro Balby Marinho and Christine Preisach, scored 1stand 2ndplace in challenge's task 2 respectively (see http://www.kde.cs.uni-kassel.de/ws/dc09/results ).
This year's discovery challenge consisted of three tasks in the area of social bookmarking, a popular practice in the Web where social bookmarking sites are distinguished by communities of practice, such as users sharing photos in Flickr, web pages and list of scientific publications in BibSonomy, and/or music in Last.fm. Common to all these sites is that users can attach tags - freely chosen keywords - to their uploaded resources (e.g., photo in Flickr, sound tracks in Last.fm, etc.), which help users to organize their resources by topic.
All tasks of the challenge targeted the support of the user during the tagging process by recommending tags, i.e., after a user has uploaded a resource to the system, the recommender algorithm must recommend a list of tags that are most likely to be used by the respective user to the respective resource. A complete dataset of the social bookmark and publication sharing system BibSonomy was provided for the challenge. A training dataset for all tasks is provided at the beginning of the competition, while the test dataset consisting of user/resource pairs with “unknown tags” is just released 48h before the deadline. For all tasks, given a test user/resource pair, the competitors should develop methods that try to predict a list of tags that are as close as possible to the real list of tags, which only the organizers know. The ISMLL teams took part on the task 2, which was especially intended for methods relying on the graph structure of the training data only, i.e., there were no new tags, resources or users in the test dataset.
The ISMLL group of the University of Hildesheim, headed by Prof. Dr. Dr. Lars Schmidt-Thieme, has been actively actuating on the areas of recommender systems and a broad range of machine learning topics, “know-how that proved to be one of the keys for achieving such successful results in the challenge”, according to Leandro Balby Marinho. The two teams responsible for this remarkable achievement are:
Steffen Rendle and Lars Schmidt-Thieme, with the approach entitled “Factor Models for Tag Recommendations in Bibsonomy”;
Leandro Balby-Marinho, Christine Preisach, and Lars Schmidt-Thieme, with the approach entitled “Relational Classification for Personalized Tag Recommendations”.
Another positive side effect resulting from the broad spectrum of machine learning topics covered in the ISMLL group, is that, despite the shared insights among the teams during the competition, both teams developed independent and distinct approaches, which end up presenting the research community with a suit of methods covering different and interesting aspects of the problem.