Model the Long-Term Post History for Hashtag Recommendation

The goal of this work is to provide a keyword-suggestion-like hashtag recommendation service, which recommends several hashtags when the user types in the hashtag symbol “#” while writing a post. Different from previously published hashtag recommendation systems, which only considered the textual information of the post itself or a few numbers of the latest posts, this work proposed to model the long-term post history for the recommendation. To achieve this purpose, we organized the historical posts of a user in the time order, obtaining a post sequence. Based on this sequence, we proposed a recurrent-neural-network-based framework, called the Parallel Long Short-term Memory (PLSTM), to perform the post history modeling. This was motivated by the success of the recurrent neural network in modeling the long-term dependency of dynamic sequences. The hashtag recommendation was performed based on both the current post content representation and the post history representation. We evaluated the proposed model on a real dataset crawled from Twitter. The experimental results demonstrated the effectiveness of our proposed model. Moreover, we quantitatively studied the informativeness of different parts of the post history and proved the feasibility of organizing the historical posts of a user in the time order.