面向互联网+智能电视平台的推荐系统研究

The platforms of the Internet+ Smart TVs are the TV systems embedded in the Internet ecotope, which can provide both the programs of CATV with more than hundred channels and the trillion videos from the Internet websites sharing the agreements with the makers of smart TV. Clearly it becomes very difficult for TV users to find the programs they want to watch. The recommendation systems on the platforms are asked to be able to provide the personalized program recommendation based on the preference of each ID user. Nevertheless, the known theoretical models of recommendation systems cannot be used directly to design the recommendation systems on the platforms for the following reasons. 1) There may be more than one different kinds of users corresponding to one TV ID; 2) The rating matrix of user-item is usually not available; 3) The preference on the programs of a user cannot be determined by the genres of programs that s/he has watched only, it also depends the preference of the user on attributes of the programs, e.g., the prize obtained, the director, the main actors/actresses and their current or historical popularity. 4) The recommendation of news and the channel programs is context-constraint. However, on the other hand, we discover that there are some characteristics of the users and programs in the platforms, such as, both the users and programs are of richer semantic description and it is possible to utilize the social networks of users and the rating matrices of the user-item and program popularity analysis in the sharing video websites, which can be used to enhance the performance of the program recommendation of the Internet + smart TV. Therefore this project tries to use the above mentioned characteristics to develop the creative recommendation models, algorithms and implementation techniques for the design of the recommendation systems for the Internet + Smart TV platforms. We believe the program recommendation systems developed based on our research on the Internet +Smart TV platforms can elevate the problems mentioned above. Furthermore our research will give the creative contribution in the aspects of new recommendation models, the implementing techniques, the system architecture and methods on online evaluation in the study of recommendation systems.

互联网+智能电视平台是融入互联网环境的电视系统，其节目包括了有线电视的上百频道和与TV厂家有共享协议的上百亿网络视频。但用户想找到喜欢的节目则变得十分困难。该平台上的推荐系统是根据每个ID用户的特点，为其进行个性化推荐。但目前已知的推荐系统理论模型均不能直接应用到这类平台上的推荐系统设计。原因是：1）同一个ID可能对应多个不同类型的用户；2）没有用户-节目评分矩阵；3）用户对节目的喜好不仅仅取决于过去看过的节目类别，而且还与获奖、导演、演员和流行度等节目属性有关；4）对新闻和直播节目的推荐是上下文受限。本申请课题将根据该平台上的用户和节目具有更丰富的语义描述和可以利用共享视频网站上用户之间、用户-节目评分等特点，给出能部分解决上述问题的推荐模型和实现技术。其研究成果可直接应用到互联网+智能电视的推荐系统设计，并且在推荐模型、系统框架、实现技术和在线推荐评估方法等方面给出创新成果。

本课题以智能电视的节目推荐系统为研究背景，在以下三个方面开展了研究：1）我们采用数据挖掘技术，基于智能电视用户点击日志去发现用户的兴趣偏好。考虑时间等上下文以及用户兴趣偏好设计个性化的智能电视节目推荐系统；我们深入地研究上下文感知推荐系统，这里上下文包括一天中不同的时间期间、工作日和周末、地点和用户使用应用系统时可能的陪伴者。2）由于用户日志实质上是由每天用户的点击序列组成的，而连续的序列可以认为是某种“会话”，我们采用深度学习技术开发出多个基于会话型的推荐系统模型。我们发现某用户点击日志中可能覆盖多个不同的领域，如娱乐和教育，我们研究了新颖的跨领域推荐系统模型。3）由于推荐系统的可解释性变得日益重要，我们研究了如何在影视节目、服装时尚和旅游线路推荐中产生可解释性方法。. 在上述三个方面研究中，我们分别给出了对用户兴趣偏好的分析模型、上下文感知推荐模型和解决多人共享同一账号的跨领域推荐模型。我们设计出一类能利用协同过滤技术和会话型推荐技术的新颖协同型会话推荐模型。除此之外，我们在影视节目和服装推荐公开数据集上，给出了多个基于用户评论和服装匹配常识的推荐解释产生方法。. 上述研究结果中有9篇论文发表在中国计算机学会列出的A类的国际学术期刊和会议、10篇B类期刊和会议。获得3项国家发明专利；培养了6名博士和15名硕士。其部分理论成果已经应用到青岛海信智能电视节目推荐系统，并在项目执行期间不断帮助企业进行系统升级和改进。