This paper presents a novel link prediction model named ACSS (Attention-based Link Prediction with Contextualized Self-Supervision), which integrates attention mechanisms to address challenges faced by traditional link prediction algorithms. Conventional methods are limited by sparse links, noisy node attributes, and dynamic variations. The ACSS model combines contextualized self-supervised learning with attention mechanisms to better capture critical relationships between nodes and improve link prediction performance. Our model employs end-to-end training, optimizing it by simultaneously learning link prediction and self-supervised learning tasks. Experimental show that the ACSS model surpasses traditional and inductive link prediction tasks across multiple real-world benchmark networks. Furthermore, the ACSS model exhibits strong robustness and scalability when handling node attribute noise and large-scale networks.