This paper provides a summary and review of embedding based recommender systems.
Word embedding frameworks like word2vec were originally developed for NLP tasks. However, they were quickly adopted in recommender systems to construct hybrid recommenders that incorporate side information in addition to user-item interaction to overcome common problems in recommender systems like cold start and popularity bias.
However, there are several proposed recommender systems that utilize embedding layers and each of them has its own strengths and weaknesses. A review and comparison between these different approaches is presented in this work. First, normal word embedding for NLP is introduced then different recommenders that utilize this method are presented and compared. Different evaluation metrics and standard datasets used for embedding based recommender systems are discussed afterwards and finally a unified comparison of all these datasets and evaluation metrics is presented in order to facilitate comparison between different embedding-based recommenders. Future work is then presented and discussed.