This paper addresses temporal graph learning, which plays a crucial role in dynamic networks where nodes and edges evolve over time and new nodes are continuously added to the system. Specifically, we focus on two key challenges: effectively representing new nodes and alleviating noisy or redundant graph information. To achieve this, we propose a multi-objective framework, GTGIB, which integrates Graph Structure Learning (GSL) and Temporal Graph Information Bottleneck (TGIB). We design a novel two-stage GSL-based structure enhancer to enrich and optimize node neighborhoods, and demonstrate its effectiveness and efficiency through theoretical proofs and experiments. TGIB improves the optimized graph by regulating both edges and features through a tractable TGIB objective function derived via variational approximation, enabling stable and efficient optimization. We evaluate the link prediction performance of the GTGIB-based model on four real-world network datasets. The GTGIB-based model outperforms existing methods in the inductive setting across all datasets, and demonstrates significant and consistent performance improvements in the transitive setting.
