Abstract: Collaborative Filtering has been widely used in modern recommendation systems,and it assumes that similar users prefer similar items.A key ingredient of CF-based recommendation model is the score function,which measures the preference of users on items.However,there are some shortages in the most popular score functions.The inner product score function fails to capture the user-user similarity and item-item similarity effectively,and Euclidean distance measurement function reduces the expressiveness of the model because of its geometrical restriction.This paper proposes a novel hybrid score function by mixing the inner product-based similarity and the Euclidean distance metric,and further theoretically analyze its properties,thus proving that the new score function can avoid the aforementioned shortages effectively.In addition,the new hybrid score function is a general technique and can help to improve the quality of recommendation for existing models (e.g.,SVD++,MF,NGCF,CML).Extensive empirical studies over 6 datasets demonstrate the superior performance of the proposed hybrid score function.

Key words: Collaborative filtering, Recommendation system, Score function