Disentangled Item Representation for Recommender Systems

Item representations in recommendation systems are expected to reveal the properties of items. Collaborative recommender methods usually represent an item as one single latent vector. Nowadays e-commercial platforms provide various kinds attribute information for items (e.g., category, price, and style clothing). Utilizing this better is popular recent years. Some studies use given side information, which concatenated with vector augment representations. However, mixed fail fully exploit rich or explanation systems. To end, we propose a fine-grained Disentangled Representation (DIR) article, where represented several separated vectors instead In way, at level, can recommendation. We introduce learning strategy, LearnDIR, allocate corresponding show how DIR be applied two typical models, Matrix Factorization (MF) Recurrent Neural Network (RNN). Experimental results on real-world datasets that models developed under framework effective efficient. Even using fewer parameters, proposed model outperform state-of-the-art methods, especially cold-start situation. addition, make visualizations our proposition users applications.