Multi-Domain Neural Network Recommender

Multi-domain recommendation is adopted to alleviate data sparsity problem that hurts performance by utilizing information from other domains in recommendation system recently. Considering the powerful ability of knowledge extraction of neural network, we design a novel multi-branch network to discover sharing-pattern features and domain-specific features for multi-domain recommendation. Sharing-pattern features are general preference of user but values are distinct in every domain in our model which is quite different from present other practices. As others use same features as sharing factors among domains while we use same transformation as sharing pattern. Besides that, we conduct a non-linear procedure with probability to form final user latent factor rather than direct adding or multiplying in some works. Experiments on real-world dataset outperforming baseline methods shows effectiveness. Observing results, another finding is sparser domains have more room for improvement as they have more absorption from others.