融合实体和交互上下文信息的社会化推荐方法研究

Recommendation systems are significant tools in handling the problem of information overload, as well as providing personalized information. With the extensive applications of social media and portable devices, great amount of the contextual information is collected. However, only few works can effectively integrate the contextual information of entities and interactions, commonly existing on social media. In order to achieve the objective of this project, developing the social recommendation algorithms based on multi-source contextual information, we should solve three sub-problems as follows. Firstly, when it comes to the integration of multi-view contextual information of entities, existing works fail in utilizing the correlations between multi-view information to decrease the effect of noise, redundancy and dimensions’ divergence. Secondly, existing work lacks efficient strategies of integrating multi-class contextual information of interaction, as well as the extraction and utilization of its operational semantic. Thirdly, the existing work has not used entity and interactive information and has difficulty in mining the social relations and domain relations from social networks. To solve those problems, this project focuses on revealing the relationships of multi-view context of the entities, extracting operational semantic of multi-view interactive context. Then we take the contextual information into consideration in mining the social network. In this case, we are able to build a social recommendation framework and provide the theoretical and technical basis for contextual social recommendation algorithms.

推荐系统是应对信息过载并为用户提供信息推荐的有效工具。随着社交媒体和移动设备的广泛应用，系统中收集了大量上下文信息，而现有社会化推荐算法又难以有效融合这种上下文信息。为实现常见的实体和交互上下文融合的社会化推荐算法这个整体目标，亟需解决三个问题：第一，现有工作在融合多视角实体上下文时，没能有效利用多视角间的关联特性，来消除噪声、冗余和维度差异；第二，现有工作缺乏有效的多类别交互上下文融合的策略，亦没有提取和利用其操作性语义的方法；第三，现有工作未能有效结合实体和交互信息，来挖掘社会化网络中的社交关系和领域关系。针对以上问题，本项目着眼于把握多视角实体上下文的关联关系，提炼多视角交互上下文的操作性语义，结合实体和交互信息深度挖掘社会化网络，构建融合实体和交互上下文的社会化推荐算法框架，为融合上下文信息的社会化推荐算法应用提供理论依据和技术基础。

本项目对实体特征建模和情境上下文感知建模两个部分开展了研究。在实体特征建模中，我们主要针对多视角、不完备特征和文本特征进行建模。在情境感知研究中，我们针对一般化的情境进行探索，特别是时间序列信息。项目在情境建模方法上获得了突破，发现情境操作性是具有普遍性的现象。另一方面，我们将循环神经网络方法有效的引入用户行为建模中来，并针对用户行为的特点，实现情境信息、时序信息、长距短距的依赖关系深入分析，开展了深入系统的研究。相关成果获得企业的关注，获得企业的横向研究计划的支持。在本项目的支持下，课题成员在国际学术会议和期刊上共发表论文18篇，其中包括顶级国际会议AAAI，ICDM，SIGIR，CIKM, IJCAI, WWW；多篇国际核心期刊TKDE,THMS,TIST,TNNLS和PR。在执行期间协助培养3名博士生，1名硕士生。