Hybrid similarities for dynamic interaction recommendation

Similarity metric is the core of k-nearest neighbor collaborative filtering in recommender systems. However, traditional metrics measure the similarity among neighbors only in either direction or distance. In this paper, we propose a triangle similarity metric and two kinds of hybrid ones based on it for dynamic interaction recommendation. First, the triangle similarity metric combines both direction and distance. Second, two kinds of hybrid similarity metrics are designed to improve recommendation quality. The first hybrid one adds up the triangle, cosine and jaccard similarities, while the second one multiplies them. Third, we apply the hybrid similarity metrics to a dynamic user-recommender interaction system. Experimental results on the well-known MovieLens dataset indicate that the additive hybrid similarity outperforms traditional similarities on the Recall measure.