UAFA: Unsupervised Attribute-Friendship Attention Framework for User Representation

The problem of user representation has received considerable attention in recent years. A variety of social networks include not only network structures (friendships) but also information about users' attributes. Previous studies have explored the integration of the two information to encode users. However, these methods focus on how to fuse the target user's friendships as a whole with its attribute information to get its representation vector, without considering the inside information of friendships, that is the influence of intimacy difference between the target user and its each friend on its representation vector. In addition, most of the above methods are supervised, which can only be applied to limited social networks analysis tasks. In this paper, we investigate a novel unsupervised method for learning the user representation by considering the influence of intimacy difference. The proposed methods take both the users' attributes and their friendships into consideration with attribute-friendship attention network. Experimental results demonstrate that the user vectors generated by the proposed methods significantly outperform state-of-the-art user representation methods on two different scale real-world networks.