Abstract: There are two problems in the existing POI recommendation algorithm. First, most algorithms mainly utilize the historical check-in data of user, while ignoring the comments of users and label information. Thus, the cold-start problem cannot be solved effectively. Second, some algorithms only use the user’s check-in score when calculating the similarity. The high sparseness of the POI check-in matrix results in the inaccurate ness of the recommendation. In view of the above problems, this paper uses the LDA topic model to mine the user’s interest topic, and then integrates the check-in data for similarity measure to solve the cold-start problem. In recommendation period, a dynamic prediction method is proposed to dynamically fill the missing data and further alleviate the sparse data and improve the recommended quality. The experimental results on the real data set show that the proposed similarity fusion and dynamic prediction based recommend algorithm can effectively solve the problem of data sparseness and cold-start. The recommend performance is superior to traditional recommendation algorithms.

Key words: Latent Dirichlet Allocation（LDA） topic model, dynamic prediction, point of interest recommendation, similarity

摘要： 现有的兴趣点推荐算法大都存在两个问题：第一，算法中利用用户签到的历史数据，而忽略了用户的评论和标签等信息，不能很好地解决冷启动问题。第二，部分算法在计算相似度时仅使用用户的签到评分，而由于POI签到矩阵的高稀疏性，会导致推荐结果不准确。鉴于上述问题，提出了利用潜在的狄利克雷分配（Latent Dirichlet Allocation，LDA）主题模型挖掘用户的兴趣话题，融合签到数据进行相似度度量，很好地解决了冷启动问题。在推荐生成阶段提出了一种动态预测法，动态填补缺失的访问概率，进一步缓解数据稀疏，提高推荐质量。在真实数据集上的实验结果表明，基于相似度融合和动态预测的兴趣点推荐算法有效解决了数据稀疏性和冷启动问题，推荐性能优于传统的推荐算法。

关键词: 潜在的狄利克雷分配（LDA）主题模型, 动态预测, 兴趣点推荐, 相似度