In this master's thesis we prepare a recommender system for recommending TV programmes. The data documenting users' preferences was acquired implicitly. The recommender system will be based on collaborative filtering, because very little additional information about TV programmes is required. The main issue with recommending TV programmes is that they rarely repeat so we cannot use the types of recommenders commonly used in other domains. We deal with this problem by describing each TV programme with attributes and then recommending sets of attributes. These will be later converted to actual TV programmes using a programme guide. A recommender system for TV programmes must be able to include new data in real time and without retraining. We tested a few different methods for recommending TV programmes: history, BRISMF (Biased Regularized Incremental Simultaneous Matrix Factorization), kNN (k Nearest Neighbours) and ECOCLE (Evolutionary Co-clustering with Ensembles). Because BRISMF is not incremental, we propose building BRISMF models for every single day and then joining recommendations from these daily models. When compared to other methods, which are already incremental by design, our proposed method gives good results.
|
