图表示学习短视频智能推荐研究综述

随着短视频数量的爆发式增长, 精准的个性化短视频推荐成为学术界和工业界的迫切需求。然而,现有的推荐方法没有考虑实际的短视频具有数据多源异构多模态、用户行为复杂多样、用户兴趣动态变化等特点。短视频模态间的语义鸿沟、社交网络用户多行为挖掘、用户动态兴趣捕捉依然是短视频推荐领域面临的三个重要问题。针对当前推荐系统存在的问题,并充分考虑短视频推荐系统的实际需求,本文介绍了短视频推荐中基于图表示学习的短视频推荐方法;研究了短视频异构多模态特征表示,充分挖掘视频内容特征并进行高效融合;研究了短视频社交网络用户多行为表示,通过社交网络用户多种行为挖掘更细粒度的用户偏好;研究了用户的动态偏好表示方法,通过利用时序信息建模用户的动态兴趣,保证推荐结果的准确度并增加其多样性与个性化。本研究可在理论和实践上推进基于图特征学习的短视频推荐研究,也可作为短视频推荐系统的关键技术。

Survey of Graph Representation Learning for Micro-Video Intelligent Recommendation

With the explosive growth of micro?video, academic and industrial communities highly desire an effective and personalized micro?video recommendation. However, existing recommendation methods do not take into account the three characteristics of real micro?videos: (1) The data is characterized with multi?source, heterogeneousness, and multi?modality; (2) Social users have complex and multi behaviors; (3) The interests of micro?video users are dynamically changing. As a result, they suffer from three critical issues, i.e., the semantic gap between micro?video modalities, multi?behavior mining of social network users, and dynamic interest capture of users. To solve the key issues and fully consider the actual needs of the micro?video recommender system, an effective micro?video recommendation method was proposed and investigated based on graph representation learning. First, a heterogeneous multimodal feature representations for micro?video was proposed to fully mine video content features and perform an efficient fusion. Second, the multi?behavior mining of micro?video social network users was studied and the user preferences in a fine?grained manner were captured by mining users" various behaviors through social networks. Finally, the user"s dynamic preference representation method was studied. By using the time series information to model the user"s dynamic interest, the proposed method can ensure the accuracy of the recommendation results while increasing the diversity and personalization. Our research can substantially advance the study of micro?video recommendations based on graph feature learning both theoretically and practically. Moreover, it can also serve as the key technique for the micro?video recommender system.