Context-aware recommender systems

Abstract

Recommender systems try to predict users' preferences for certain items, given a set of historical data. Multiple different techniques are available that make these systems accurate and one of them that delivers promising results is matrix factorization. This thesis explores how these systems work and presents a method to incorporate contextual data into a factorization technique to get predictions based on context. Specifically, a music recommender based on Candecomp/Parafac tensor factorization is proposed that uses implicit feedback collected from music listeners. The results are empirically tested and compared with other non-contextual recommender techniques. The prediction quality of the matrix factorization technique is unfortunately not improved by our proposed tensor factorization recommender on the used Last.fm dataset. However, an adjusted dataset with artificially made contextual data does get better results, but this may not reflect a real-world situation.