Personalized TV and content recommender : collaborative filtering in recommender systems

The purpose of this thesis work done at Ericsson Research in Kista was to investigate the possibilities of improving the recommendations in a recommender system for TV content. In the first phase of the work investigations of current techniques were carried out. Once an understanding of those techniques was achieved the focus shifted to improving the way to measure similarity between users or items, which is commonly used in many algorithms. In the second phase a new double weighted correlation scheme was developed in order to solve some of the flaws with the existing ones. The hypothesis was that the double weighted correlation would measure similarities between users or items in a recommender system more accurate than the existing ones, but also that it could be used to compute user-clusters that could be used as pre-computed neighborhoods. Finally the third phase consisted mainly of implementing the algorithms and testing them on different datasets. The double weighted correlation's ability to measure similarity between users was tested with fabricated data. These tests where made in Matlab and the double weighted correlation showed good results in all cases, while other existing correlation schemes had more fluctuating results. In order to see how it works in a recommender system it was also tested with a nearest neighbors algorithm and compared with the most commonly used correlation, namely Pearson. These tests were implemented in Java and run on both the Movielens 100k dataset as well as a dense 100k subset of the Jester dataset. The evaluation was made using ROC- curves. The double weighted correlation enhanced the performance compared to when using Pearson. However, it performed worse when adding default ratings to the algorithm. The double weighted correlation's ability to cluster users was tested using combinations of k-means algorithms on both the Movielens 100k dataset as well as a dense 100k subset of the Jester dataset. Contrary to the hypothesis it could not find any clusters or structures in any of the two datasets.Validerat; 20101217 (root