Jointly Tackling User and Item Cold-start with Sequential Contentbased Recommendations
Deep learning has been successfully used in the context of recommender systems. Sequential recommender systems are a class of algorithms which model user-item interactions and their temporal relationship in order to generate relevant personalized recommendations. Recurrent neural networks have become the state-of-the-art approach for sequential modeling, but current approaches in the context of recommendation systems are tightly coupled with the catalog size and item identifiers. This imposes a problem when new items are to be incorporated into the list of recommendable products, the entire model needs to be retrained. Feature-rich item metadata has been successfully used to improve recommendation quality with both sequential and non-sequential recommenders. However, to the best of our knowledge, no attempt has been made to tackle the problem of newly encountered user and item in a sequence aware model with personalized recommendations. This work presents a novel architecture for context-aware item prediction based on embeddings. The model combines item embeddings within a sequence to dynamically predict an item embedding for the next interaction. This allows to incorporate new items without model retraining. Moreover, the proposed architecture implicitly models the user preferences from user-item interactions and is able to provide item embedding predictions that are personalized to the context of a user and therefore produce personalized recommendations. The results are compared with GRU4Rec and TransRec in the next interaction prediction task using the Amazon reviews public dataset, and our experiments show comparable or better results than state-of-the-art personalized models, with the added benefit of being able to add items or users without model retraining.
Graduation Thesis language
Graduation Thesis type
Master - Computer Science
Tambet Matiisen, Carlos Bentes