Temporal graph attention and contrastive learning model for link prediction in dynamic networks

Abstract

Existing discrete time-slicing methods suffer from three critical limitations: coarse temporal processing granularity, exclusively model connection establishment events while neglecting the impact of disconnection events, and vulnerability to sample imbalance and transient connection noise. These limitations severely constrain adaptability in dynamic networks characterized by connection instability and interaction volatility, where both connection establishment and disconnection events govern topological evolution. To address these challenges, this paper proposes a novel dynamic link prediction framework integrating Contrastive Learning with enhanced Temporal Graph Attention Network (CLTGAT). Our model employs continuous timestamp encoding to explicitly incorporate connection moments while fusing disconnection moments via weighted mechanisms. This dual-event modelling enables joint influence of connection/disconnection timestamps on link states. Crucially, we design a connection-duration-based Top-K contrastive sampling strategy to simultaneously mitigate transient connection noise and sample imbalance, selecting stable neighbours as positive samples while controlling distribution to alleviate prediction bias. Evaluations on three highly volatile real-world dynamic network datasets, viz., Infocom05, Hyccups and Infocom06, demonstrate CLTGAT’s superiority. Compared with seven other methods, our approach achieves higher link prediction accuracy with reduced training time, exhibiting enhanced adaptability to rapidly evolving network scenarios.