Hashtag Recommendation based on Neural Topic Model

Hashtag recommendation for microblogs is an appealing research topic that can assist users in selecting hashtags for their microblogs. Previous hashtag recommendations have been proposed based on the topic model approach. In spite of their improvement, two issues are still unresolved. First, prior studies inferred the topic distribution of users, words, and hashtags by using a non-neural network framework, making them unable to deal with the complex topic distribution of users, words, and hashtags. Second, previous work required word and hashtag frequency to process the topic distribution inference only, ignoring not only word ordering and word context but also the relation between word and hashtag. It is one of the crucial factors that affect hashtag recommendation since some words in the target microblog can indicate the relevant hashtags. In this paper, we propose a novel hashtag recommendation model called Hashtag-NTM to infer the topic distribution of users, words, and hashtags by using the neural network framework, making the model able to deal with the complexity of distribution. Moreover, we explore the relationship between words and hashtags to model word and hashtag representation before processing topic distribution inference, making the recommended hashtags more related to words in the target microblog. Extensive experiments on the Twitter dataset demonstrate that Hashtag-NTM consistently outperforms baseline methods over hit rate metrics.