Hierarchical Diffusion-Based Ad Recommendation with Variational Graph Attention and Adversarial Refinement

Sequential advertisement recommendation is an important task in intelligent advertising systems. It needs models that can learn user–ad interaction sequences from different types of data. Many existing methods have trouble keeping a good balance between generation quality, robustness, and efficiency. This becomes more difficult when there is little user data or feedback. To solve this, we present GDAR (Generative Diffusion-based Advertisement Recommendation). GDAR is a single framework that includes a hierarchical diffusion process with trainable noise steps for fast sequence generation. It also includes a variational graph attention network to learn dynamic co-occurrence and time relations using uncertainty-aware embeddings. In addition, it has an adversarial module that uses contrastive learning to improve diversity and meaning. The model is trained with several types of loss, including diffusion, variational, adversarial, and contrastive loss. Data-level augmentation is used to help the model generalize better. This framework is designed to improve sequential ad recommendation in real-world systems.