移动阅读系统中的多源动态混合推荐

Multisource Dynamic Hybrid Recommendation in Mobile Reading System

针对超大型移动阅读系统的个性化推荐问题,提出了UFM模型以评价用户隐式反馈的兴趣偏好,并基于信息熵进行了评价指标赋权。提出了基于用户和属性的多源动态协同过滤算法UA-MDCF,在算法中提出属性偏好贡献度的概念,以反映不同属性对于每件物品的评分的不同贡献,创新地基于时变衰减函数综合了多种来源的属性偏好评分(用户显式、隐式反馈以及第三方来源的用户隐式反馈),混合了基于Jac UOD相似度的用户评分协同过滤算法和基于属性偏好的协同过滤算法,很好地解决了数据稀疏、冷启动、时效性、多种偏好信息引入等问题,降低了用户反馈依赖,提升了推荐准确率。并通过中国移动咪咕阅读平台的实际营销推荐应用案例证明了该算法的有效性。

As to the personalized recommendation problem in large-scale mobile reading system, UFM model is put for-ward to evaluate implicit interests and preferences of a user, evaluating indicators are weighted basing on information entro-py. An User and Attribution Based Multisource Dynamic Collaborative Filtering Algorithm, i.e. UA-MDCF, is put forward.In the algorithm, to reflect different contribution of each attribution of an item to the preference rate of the item, a new con-cept of attribution preference contribution rate is raised. Based on time attenuation function, multisource attribution prefer-ence scores are synthesized, including users＇ explicit feedback, implicit feedback and the third party sourced implicit feed-back. User scoring with Jac UOD similarity based collaborative filtering and attribution preference based collaborative filter-ing are mingled. As the result, problems of data sparsity, cold-start, timeliness, multisource preference information intro-duction, etc., are solved more effectively, while dependency on user feedback is decreased, and recommendation precisionrate is increased. Finally, through practical marketing and recommendation application in Migu Reading Platform of CMCC,effect of the algorithm are proven ideal.