Similarity-Based Context-Aware Recommendation

Context-aware recommender systems (CARS) take context into consideration when modeling user preferences. There are two general ways to integrate context with recommendation: contextual filtering and contextual modeling. Currently, the most effective context-aware recommendation algorithms are based on a contextual modeling approach that estimate deviations in ratings across different contexts. In this paper, we propose context similarity as an alternative contextual modeling approach and examine different ways to represent context similarity and incorporate it into recommendation. More specifically, we show how context similarity can be integrated into the sparse linear method and matrix factorization algorithms. Our experimental results demonstrate that learning context similarity is a more effective approach to contextaware recommendation than modeling contextual rating deviations.