Drug-drug interactions (DDIs) occur when multiple drugs are administered concurrently, which can alter their effectiveness, safety, or patient tolerance. Understanding and accurately predicting these interactions is crucial for personalized medicine, drug development, and reducing adverse effects in clinical settings. However, most current deep learning approaches, particularly graph-based methods, primarily focus on molecular structure information while often neglecting the complex inter-view interactions that can influence outcomes. To overcome this notable limitation, we introduce a new graph-based hierarchical approach that effectively aligns both intra-view and inter-view embeddings for enhanced prediction. Our method employs distribution matching in both the feature and output spaces, maximizing mutual information between these views to improve prediction accuracy and robustness. Additionally, we introduce an innovative loss function based on central matching distribution (CMD), which balances the information flow between the two views without relying on traditional unsupervised contrastive learning techniques. This approach not only improves computational efficiency by 50%, but also maintains high levels of prediction accuracy. Experiments on three widely recognized datasets show that our approach consistently surpasses multiple leading models in DDI prediction tasks, providing a reliable and scalable solution.