Leveraging Reconstructive Profiles of Users and Items for Tag-Aware Recommendation

It is an effective recommendation method by revealing user preferences and extracting latent semantic information of items through social tag information. Recent research shows impressive recommendation performance by using neural network-based methods to transform tag-based user or item profiles to abstract feature representations. However, in the process of training a neural network, these methods need an more effective measurement to balance the tag-based profiles and the abstract representations to further improve item recommendation. This paper proposes a method based on Generative Adversarial Networks to tackle this issue. In this method, abstract features of users and items are extracted from their tag-based profiles by a disentangling network. These abstract features are then used to calculate the probability of a user preferring an item, and are also used to reconstruct new user and item profiles by a generative network. Furthermore, the discriminative network is introduced to identify generated profiles for enforcing smoothness in the representation of users and items. Experiments on two real-world data-sets demonstrate the state-of-the-art performance of the proposed method.