Graph-based learning methods effectively analyze citation networks by uncovering structural relationships between scientific publications. Multi-layer Graph Convolutional Networks (MGCNs) have demonstrated competitive performance on traditional citation benchmarks such as Cora, CiteSeer, and PubMed. However, such models produce purely structure-based recommendations that lack personalization and temporal sensitivity. Recently, DyBooster has shown promising results in post-hoc boosting modules, re-ranking predictions based on user interaction history and temporal self-correction. This paper combines MGCN with a DyBooster-inspired re-ranking method for citation recommendations. We first execute an MGCN pipeline using link prediction on three benchmark datasets to generate candidate recommendations. We then introduce a post-processing step that computes temporal indicators (Indication Tendency and Addition Tendency) based on user reading history and adjusts the recommendation list by balancing familiar and novel research categories. Experimental results show that MGCN exhibits strong structural learning capabilities with test Area Under the Curve (AUC) rates of 95.15% on Cora, 94.77% on CiteSeer, and 95.66% on PubMed. The impact of the temporal re-ranking using DyBooster is quite pronounced on the distribution of categories in top K recommendations. This is because it facilitates exploration in a controlled manner while maintaining structural relevance.