Supported matrix factorization using distributed representations for personalised recommendations on twitter.

Highlights • The work is unique as it is first known technique to exploit recurrent neural networks to support Prime Matrix Factorization. • Exploits the implicit content of microblogs to generate distributed representations. • Exploits the explicit factors like likes, favorites, followers and friends to model relationships between hashtags and users more accurately. • Uses implicit and explicit features to performed supported matrix factorization under regularization. • Experimentations show significant improvement over standard prime factorization methods. Abstract Microblogging is one of the most prevalent media for sharing news on the Internet. Microblogging platforms, such as Twitter have proven to be of great success in targeted marketing, alerting about natural disasters and promoting government policies among others; But most of this relevant information in microblogs is side-lined, owing to information overload, rendering any practical utility of the platform as ineffective. Hence, it is crucial to filter data and recommend only relevant information to the users. Interestingly, to pertain and appeal to a certain community, users make the use of hashtags (#), which in turn, helps in the efficient categorization and summarization of microblogs. In this paper, we exploit this advantage through a novel framework for a recommendation system, Distributed Representation based Supported Matrix Factorization (DRSMF) build on top of Probabilistic Matrix Factorization (PMF) and Recurrent Neural Networks (RNNs). The RNNs generate character-level distributed representations for each tweet to overcome the solecistic use of sentence structure in microblogs. The framework further performs a multi-modal analysis on the microblog posts to recommend similar users and hashtags , which assists in countering information-overload. Our framework outperforms standard PMF techniques by the use of constrained regularisation on latent factor representations. [ABSTRACT FROM AUTHOR]