中文摘要：

该文提出一种基于关注关系和多用户行为的图推荐算法AttentionRank＋,目的是为网络系统用户提供感兴趣的物品推荐.算法思路如下：首先根据用户对物品的多种反馈建立＂用户-物品＂反馈图,根据用户间的关注行为建立用户兴趣图;分别从每个用户节点出发,在反馈图上完成一轮Random Walk,得到每个用户节点与反馈图上各节点间的相似度;将用户节点与物品节点的相似度信息在兴趣图上进行扩散,计算通过关注关系扩散后用户节点与物品节点间新的相似度;重复上述Random Walk和信息扩散的过程,直到反馈图上用户节点与各节点间的相似度收敛到稳定值;最后根据用户节点与物品节点间的相似度信息,计算每个用户的物品推荐列表.该文采用包含关注、收藏、上传等用户行为的YouKu数据集对推荐算法进行评价,实验结果表明AttentionRank＋能够在用户行为稀疏的情况下,为用户提供高质量的视频推荐.

英文摘要：

In this paper, we present a graph-based recommendation, AttentionRank＋, which considers not only the multi-behaviors but also the attention relationship between different users by extending the basic Random Walk algorithm. First, we build a weighted user-item graph based on the multi-behaviors, and then we build the attention graph based on the follower-followee relations between users that have similar interests. After that, AttentionRank＋ conducts a Random Walk on the weighted user-item graph and calculates the similarities between the target user node and any other node. If a user （follower） pays attention to other users （followees）, we assume that the similarities between the follower and other nodes can be affected by his/her followees. So the similarity information of a user node can be spread along the interest edges to the followers on the attention graph. Each follower updates his/her own similarity information, and the new similarities are taken as the initial values for the next Random Walk. Repeat the above process until the similarity information of each user node converge to stable values. Finally, AttentionRank＋ creates a recommendation list of items for the target user according to the similarities between the user node and the other nodes in descending order. We evaluate the performance of AttentionRank＋ by using a YouKu dataset containing users follow, upload and collection records. The extensive experimental results show that AttentionRank＋ is capable of providing users with personalized and high quality video recommendation even when user behaviors are sparse.