GraFSTNet: Graph-based Frequency SpatioTemporal Network for Cellular Traffic Prediction

TL;DR

GraFSTNet is a novel graph-based neural network that jointly models spatio-temporal dependencies and periodic patterns in cellular traffic data, improving prediction accuracy without relying on predefined topologies.

Contribution

The paper introduces a combined spatio-temporal and time-frequency modeling framework with an attention-based spatial branch and an adaptive loss function for cellular traffic prediction.

Findings

Outperforms state-of-the-art methods on three datasets.

Effectively captures periodic traffic patterns.

Reduces prediction errors across different traffic intensities.

Abstract

With rapid expansion of cellular networks and the proliferation of mobile devices, cellular traffic data exhibits complex temporal dynamics and spatial correlations, posing challenges to accurate traffic prediction. Previous methods often focus predominantly on temporal modeling or depend on predefined spatial topologies, which limits their ability to jointly model spatio-temporal dependencies and effectively capture periodic patterns in cellular traffic. To address these issues, we propose a cellular traffic prediction framework that integrates spatio-temporal modeling with time-frequency analysis. First, we construct a spatial modeling branch to capture inter-cell dependencies through an attention mechanism, minimizing the reliance on predefined topological structures. Second, we build a time-frequency modeling branch to enhance the representation of periodic patterns. Furthermore, we introduce an adaptive-scale LogCosh loss function, which adjusts the error penalty based on traffic magnitude, preventing large errors from dominating the training process and helping the model maintain relatively stable prediction accuracy across different traffic intensities. Experiments on three open-sourced datasets demonstrate that the proposed method achieves prediction performance superior to state-of-the-art approaches.