Online tracking technology is a critical tool for user-centric platform practitioners to link users across multiple web pages and make detailed user profiles for the improvement of recommender systems like targeted advertising. Recently, due to the dynamic address allocation and security upgrade, mitigations indirectly make prior tracking techniques unreliable. To overcome the problem, traffic-based tracking techniques are proposed to link users’ dynamic addresses through similarity learning of user behaviors in their traffic interaction. However, prior work either provides poor similarity learning ability or is impractical when applied to a large scale. In this paper, we propose GALG, a graph-based artificial intelligence approach to link addresses for user tracking on TLS encrypted traffic. GALG uses the framework of graph autoencoder and adversarial training to learn the user embedding with semantics and distributions. Employing a new theory – link generation, GALG could link all the addresses of target users based on the knowledge of address-service links. When evaluated on real-world user datasets, GALG outperforms existing approaches in both performance and practicality.