A Twitter Hashtag Recommendation Model that Accommodates for Temporal Clustering Effects

Abstract

Hashtags in social medial platforms such as Twitter are important for accessing related messages as well as for tracking and detecting events. Motivated by the observation that the average hashtag experiences a life cycle of increasing and decreasing popularity, the authors propose the Topic-over-Time Mixed Membership Model (TOT-MMM), a hashtag recommendation approach that captures the temporal clustering effect of latent topics in tweets. Their experiments on 1 million tweets suggest that TOT-MMM outperforms other hashtag recommendation approaches on tweet similarity and latent Dirichlet allocation. Combining TOT-MMM with the similarity-based approach yielded additional performance improvements. The authors' simulation studies on the British Petroleum oil disaster, which happened in April 2010, suggest that the combined approach successfully identifies a higher volume of additional event-related tweets and generates signals that lead to the lowest signal-detection delay at a reasonable false alarm rate of 1.34 percent.