Abstract: Personal service is a hot topic. But how to construct an integrated user model remains a challenge for us. This paper makes use of the topic model LDA to infer the user model. In order to improve precision, CTR is put into use for restrict of feature vector. With a few manual factors, the machine generates a training topic model. Based on this model, 100 users’ micro-log messages regarded as test data will be applied for evaluating the quality of recommendation. The results show that the recommendation of celebrity performs better than the recommendation of news. Generally speaking, personal service is satisfying.

Key words: Latent Dirichlet Allocation（LDA）, topic model, Collaborative Topic Regression（CTR）, user model, recommendation

摘要： 个性化服务一直是研究的热点，但是如何构建完整的用户模型是一个颇有挑战性的问题。将基于主体模型LDA对用户模型进行预测，在用户和推荐项目的特征向量上采用CTR进行约束，使结果更为准确。在只需要少量人为因素下，由机器来训练最初的主题模型，在训练模型的基础上，通过选取100名用户的微博作为测试，用等级打分制来对推荐的项目进行打分，最终的结果显示，在新闻推荐上，微观满意度达到82.5%；而在名人推荐上，微观满意度达到了84.3%，综合以上，推荐服务的满意度还是令人满意的。

关键词: 隐形狄雷克雷分布（LDA）, 主题模型, 基于主题模型的协同过滤（CTR）, 用户模型, 推荐