<p>Visual recommender systems leverage images from sensor devices to enhance recommendations by integrating user feedback with visual features. Learning to rank algorithms are commonly used, with Visual Bayesian Personalized Ranking (VBPR) being popular. VBPR utilizes pairwise optimization to maximize the margin between positive and negative predictions. Adversarial Multimedia Ranking (AMR) enhances VBPR by robustifying it through adversarial pairwise ranking (APR). However, a challenge with APR is that not all unobserved feedback is negative; some unknowns could be positive, mislabeled as negative. This paper presents the Adversarial Graph Pairwise Training (AdvGPT) model, employing negative sampling for robust visual recommendation. Our contributions are threefold: (1) Integrating non-visual features in a factorization machine to capture users’ aesthetic preferences. (2) Introducing a graph APR algorithm that constructs item neighbor sets based on similarity in both visual space and graphs. This strategy assumes that items in the neighbor set of a purchased item (positive sample) are more likely to match the user’s aesthetics. We consider these potential positives as a third label type, between positive and negative. (3) Perturbations in connected items can cumulatively affect a target item in visual and graph spaces, making user-item graph structures vulnerable. To improve AdvGPT’s robustness and generalization, we learn user and item embeddings, exploring various interaction types. Employing the adversarial pairwise ranking method, we optimize AdvGPT to generate more robust embeddings for recommendations. Extensive experiments on real-world datasets validate the effectiveness of our approach.</p>