Learning Users’ similarity for hashtag recommendation on Twitter : A Graph Representation Learning based Approach

Learning Users’ similarity for hashtag recommendation on Twitter : A Graph Representation Learning based Approach by Irfan ul Haq Qureshi (2022)

Humans naturally seek to pursue and share information, thoughts, and beliefs. With the advancements in digital technologies, social media has now become a major part of today’s social system. Twitter is one of the major social media platforms that allows users to share their words through microblogs, called tweets. Users can also label their tweets with one or multiple hashtags representing the topic of discussion. Other users can browse hashtags of their interest to search for related tweets. That way, proper use of hashtags plays a vital role in the Twitter micro-blogging environment. Hashtags help in mitigating information overload problem by organizing similar tweets under relevant topics. Users are free to create hashtag terms themselves and may use informal or incorrect language. That creates a massive pool of different hashtags related to the same topic, making it challenging for users to use them effectively. A user may not find the appropriate hashtag of interest to browse related tweets or tag his tweet with, which creates the need for hashtag recommendation methods that recommend relevant hashtags to the users. Twitter hashtag recommendation has been an active area of research in the last decade, and various methods have been proposed. Most existing methods in the area are extensions of methods from the Information Retrieval field and standard recommendation systems, and predominantly suffer from data sparseness and dynamicity in tweets, which makes real-time hashtag recommendations a challenging task. In this thesis, we propose a deep learning-based solution based on social collaborative filtering paradigm combined with the state-of-the-art graph representation learning techniques. We propose that representations of users calculated using their hashtag usage history and social relations can be used to make practical hashtag recommendations. Experiments show promising results when compared to the baseline methods.