Providing personalized recommendations of programs to the users is an important goal of the ViSTA-TV project. Since ViSTA-TV deals with content provided over the internet it has access to valuable user viewing behavior. This provides a world of information from which we can learn the user viewing preferences and subsequently recommend the most relevant shows for that user. This ensures the user a much richer experience of television watching as he does not have to squander time in going through the myriads of shows before finding the few he is interested in. The goal here is to provide a smart TV guide so to speak.
The domain however presents a few interesting challenges. Unlike providing recommendations for more perpetual items like movies, books or clothes the ephemeral nature of the TV programs makes it more demanding to use as an item by itself. Most programs like series, news, talk shows have a short lifetime and there are new shows coming on constantly. Another challenge is that subscribed users are continuously changing and increasing in numbers. So there is a significant cold start problem.
University of Zürich (UZH) and Technische Universität Dortmund (TUDo) have been tackling these very issues and learning recommender models under these constraints to provide informed recommendations. Both UZH and TUDo are building clustering based recommender models. UZH employs agglomerative clustering of users and programs based on the user program ratings. TUDo uses Tag Clustering to cluster users and programs based on the tags associated with the users and programs. The tags are the user and program features.
Both these approaches cluster the features of the users and programs and not the individual users and programs thereby circumventing the ephemeral nature of the specific users and programs themselves and dealing more with their inherent properties which are representative of these entities. From these clusters the individual users and programs can be identified and associated with each other to provide the recommendation.
ViSTA-TV successfully organized a workshop with UZH and TUDo as the main participants from 19 Aug 2013 to 23 Aug 2013 at UZH campus in Zürich with the aim of consolidating and collaborating the efforts of UZH and TUDo in designing and developing the recommender models and recommendation engine.
We will post the results of this cooperation in a future blog entry.