TopoFreer: Towards Topology-Free Cellular Traffic Prediction Through Self-Adaptive Local Spatio-Temporal Graph Convolution

The topology of cellular networks may vary over time or even be inaccessible due to privacy concerns. Although recent spatio-temporal neural networks employ learnable topology-independent adjacency matrices to model inter-cell dependencies, they capture spatial and temporal correlations among cellular traffic separately rather than simultaneously, thus resulting in limited prediction accuracies. To overcome these challenges, we propose TopoFreer, a topology-free cellular traffic prediction model based on self-adaptive local spatio-temporal graph convolution. The performance of TopoFreer is evaluated on two real-world datasets and the results demonstrate that it achieves lower prediction loss than baseline models and performs comparably to its topology-aware variants.