GeMamba: Graph Convolution Network Enhanced-Mamba for Long-Term Cellular Traffic Prediction

With the rapid popularization of mobile networks, accurately predicting long-term cellular traffic is critical for optimizing network resource allocation. To address the limitations of Transformer architectures in capturing long-term dependencies and multivariate relationships this paper proposes a Graph Convolutional Network enhanced Mamba (GeMamba) that includes a Period GCN and a Variable Correlation GCN. These components aim to capture both the periodic characteristics and the homogeneous and heterogeneous relationships of cellular traffic data, respectively. Subsequently, this paper propose a bidirectional Mamba architecture that effectively addresses the unidirectional feature-extraction constraint inherent in standard Mamba blocks. By establishing complementary information series through forward and backward processing paths, this architecture enables comprehensive temporal modeling of long range dependencies in cellular traffic series. Specifically, the dual-directional processing mechanism facilitates spatiotemporal interaction learning between geographically distributed measurement areas, allowing the model to simultaneously capture both local periodic patterns and global contextual correlations. Evaluations on Milan and Trentino cellular traffic datasets demonstrate GeMamba's superiority over state-of-the-art models, achieving 4.1 and 1.5% MSE reductions respectively while improving the Period Capture Rate by 2.6 and 0.4%, respectively. The results validate the model's effectiveness in capturing both local periodic patterns and global contextual correlations.