Abstract:
Personalized hashtag recommendations can provide relevant hashtags for a microblog. Despite performance improvement, three challenges remain unexplored. First, previous works construct user and hashtag representations based on relations from themselves. We argue that users and hashtags are influenced not only by their own relations (i.e., first-order relations) but also by the relations of a distant user/hashtag that is indirectly connected in multiple communities (i.e., high-order relations). Second, prior works perform personalization at the microblog level while ignoring the user aspects presented for each word in the microblog. Third, past studies capture correlations among hashtags in the same microblog by considering their sequence. We argue that hashtag correlations are sequenceless since they can reorder without changing their relevance to the microblog. To overcome these three challenges, we propose a personalized hashtag recommendation that consists of three parts. First, we employ graph neural networks to derive user and hashtag representation from high-order multiple relations in three communities: (1) user-hashtag interaction; (2) user-user social; and (3) hashtag-hashtag co-occurrence. Second, for word-level personalization, we extend the bidirectional attention to take both word and user representation as input. Finally, for sequenceless hashtag correlations, we feed the hashtag representation into the bidirectional attention and train using mask modeling. Experiments on the Twitter dataset show that our proposed method outperforms the state-of-the-art on precision, recall, and F1-score.
