Hyperbolic Adversarial Learning for Personalized Item Recommendation

Personalized recommendation systems are indispensable intelligent components for social media and e-commerce. Traditional personalized item recommendation models are vulnerable to adversarial perturbations, resulting in poor robustness. Although adversarial learning-based recommendation models are able to improve the robustness, they inherently model the interaction relationships between users and items in Euclidean space, where it is difficult for them to capture the hierarchical relationships among entities. To address the above issues, we propose a hyperbolic adversarial learning based personalized item recommendation model, called HALRec. Specifically, HALRec models the interactions in hyperbolic space and utilizes hyperbolic distances to measure the similarities among entities. Moreover, instead of in Euclidean space, HALRec exploits the adversarial learning technique in hyperbolic space, i.e., HAL-Rec maximizes the hyperbolic adversarial perturbations loss while minimizing the hyperbolic based Bayesian personalized ranking loss. Hence, HALRec inherits the advantages of hyperbolic representation learning in capturing hierarchical relationships and adversarial learning in enhancing the robustness of the recommendation model. In addition, we utilize tangent space optimization to simplify the learning of model parameters. Experimental results on real-world datasets show that our proposed hyperbolic adversarial learning-based personalized item recommendation method outperforms the state-of-the-art personalized recommendation algorithms.