Learning temporal event distributions via confidence-constraint diffusion for enhanced knowledge graph forecasting

Temporal Knowledge Graph Embedding (TKGE) facilitates the modeling of evolving relationships within dynamic systems. While conventional deterministic models struggle to account for the stochasticity inherent in future events, emerging generative paradigms frequently lack the structural grounding required to maintain topological fidelity. To bridge this gap, we introduce the Temporal Knowledge Graph Embedding through Constraint Diffusion Model (TCD), a framework that reformulates event forecasting as a constrained temporal fact generation task. TCD captures time-dependent event distributions while enforcing structural validity through an integrated architectural design. The Structural Element Embedding Module (SEEM) encodes local connectivity patterns into continuous representations, effectively aligning discrete graph structures with the diffusion process. Complementing this, the Confidence-Constraint Event Denoiser (CCED) utilizes a minimum confidence principle within a residual learning network to guide the reverse denoising trajectory. This mechanism prunes structurally inconsistent predictions, ensuring that generated facts remain anchored to the underlying graph logic. Empirical evaluations across benchmark datasets demonstrate that TCD significantly improves predictive accuracy, offering a robust approach to structurally-aware generative reasoning.