关键词: 个性化推荐, 用户偏好, 推荐系统, 多样性, 划分拟阵约束, 子模函数

Abstract: Diversified recommendation methods aim to provide top-k recommendations that are both relevant and diverse. Most existing methods for diversity recommendation do not take into consideration diversity and accuracy at the same time, and these methods assume that each item is of equal significance. Inspired by this, this paper proposes a novel method for diversity recommendation based on user preference. For the personalized recommendation system, the approach models user’s aggregate category preference, inner category preference and relevance, which jointly merges diversity and relevance into a submodular function and enforces an non-uniform partition matroid constraints (i.e., different users have different preference degree for each item category even in the same category and each item is not equally important) on all categories. This paper proves that maximizing the proposed object fuction is NP-hard, and by local greedy select in the same category to solve submodular function can obtain a guaranteed approximation ratio of [1-1/e], and also reduces the complexity of the algorithm. Last but not least, a penalty factor is introduced to automatically adjust the hardness of item belonging to the same category included in recommendation list. Extensive experiments on different real data sets demonstrate that the proposed method can not only achieve a good tradeoff between diversity and relevance, but also have high efficiency.

Key words: personalized recommendation, user preference, recommender system, diversity, partition matroid constraints, submodular function