Link prediction is a critical research frontier in graph data analytics. In real-world scenarios, graph data often exhibit heterogeneity and temporal dynamism, adding layers of complexity to analytical models. Existing work typically considers either heterogeneity or temporal dynamism, with very few studies modeling both properties simultaneously. In this study, we propose a novel link prediction framework designed to integrate both heterogeneity and temporal variations within graph data, applicable across a wide range of domains. Our framework includes a Dual-Window Strategy and a Mix Information Graph Neural Network (MIGNN) model. The MIGNN synthesizes temporal node representations by aggregating heterogeneous temporal information, while the Dual-Window Strategy enhances the model’s ability to capture long-term distributional characteristics inherent in graph data. We also incorporate additional experiments using a more equitable data-splitting approach and perform comparisons with an expanded range of baseline methods.