Improving Hashtag Recommendation for Instagram Images by Considering Hashtag Relativity and Sentiment.

Abstract: Extracting knowledge from user-generated content (UGC) in social media platforms is a very hot research topic in the area of machine learning, nonetheless, the main challenge resides in the fact that UGC carries inference, abstraction and subjectivity alongside objectivity. With the aim of recognising the importance of subjectivity as an influential aspect for providing humanoid results from a machine learning algorithm, this study proposes a novel approach to improve Instagram hashtag recommendation by considering sentiment that can be expressed for images. Two main points are studied in this thesis; evaluating the relativity of Instagram image to hashtag for both objective and subjective features of an image and the effect of sentiment on said relativity. This work examines three machine learning methods for hashtag recommendation: AWS service, developed algorithms with and without sentiment considerations. The models are tested on a collected dataset of de-identified Instagram posts in location London gathered from public profiles. The results show that considering sentiment significantly improves Instagram hashtag recommendation.