• bert
• bert4rec
• deep learning
• embeddings
• gat
• gcn
• graph neural networks
• machine learning
• recommendation systems
• sage
• sequential recommendations
• transformers

roduct recommendation is of paramount importance for improving customer experience in online retail. In this work, we propose to tackle the problem by integrating BERT4Rec with a deep graph network that allows learning an item representation fusing multimodal information, namely visual, textual, and transaction history. We validate our approach on an industrial scale dataset and we demonstrate increased recommendation performance with respect to BERT4Rec using only a single source of information for the items. The thesis has been developed in collaboration with H&M.