Abstract: 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.

Key words: Recommender system,Diversity,Top-N recommendation,Probabilistic selection