一种提高推荐多样性的概率选择模型

Improving Recommendation Diversity via Probabilistic Selection

传统的推荐算法多以优化推荐列表的精确度为目标,而忽略了推荐算法的另一个重要指标——多样性。提出了一种新的提高推荐列表多样性的方法。该方法将列表生成步骤转换为N次概率选择过程,每次概率选择通过两个步骤完成:类型选择与项目选择。在类型选择中,引入项目的类型信息,根据用户对不同项目类型的喜好计算概率矩阵,并依照该概率矩阵选择一个类型;在项目选择中,根据项目的预测评分、项目的历史流行度、项目的推荐流行度3个因素重新计算项目的最终得分,选择得分最高的项目推荐给用户。通过阈值TR来调节多样性与精确度之间的折中。最后,通过对比实验证明了该方法的有效性。

Typical recommendation algorithms focus on optimizing the accuracy of recommendation lists,however,diversity is also considered as a key property to measure the quality of recommendation lists from both user and system perspective.Many list diversification techniques improve diversity by re-ranking items.In this paper,a new probabilistic selection model for improving the diversity of recommendation lists was proposed.This model transfers the list generation process to N-times probabilistic selection process,and each selection includes two steps：genre selection and item selection.For the genre selection phase,genre information of items is included to compute user-genre probabilistic matrix,and a genre is chosen based on this matrix.For the item selection phase,three properties including estimated score of items,historical popularity of items,and recommending popularity of items are considered for item re-scoring.The item with the highest re-computed score will be selected into the recommendation list.The trade-off between diversity and accuracy can be controlled by changing threshold value TR.Experiments on two movie recommendation datasets show that our model can effectively improve recommendation lists diversity.At the same time,the comparative experiments show that our model outperforms re-ranking method in almost all experimental results,except the case of individual diversity for matrix factorization.