Researchers just gave cellular traffic forecasting a context-aware upgrade, and it beats the old fixed-map approach.
A team publishing on arXiv describes SCRR-Net, a framework for predicting SMS volume, network traffic, and call activity across a city. Unlike prior systems that lock spatial relationships between neighborhoods into a static graph, SCRR-Net uses a context encoder to read the current urban scene and then dynamically routes which spatial connections and which prediction tasks share information. The system pairs that routing logic with a temporal Transformer encoder to track how activity shifts over time. In tests on real telecom datasets from Milan and Trento, it beat competing methods across all three signal types.
This matters because telecom operators lean on these forecasts to plan network capacity, and cities change shape by the hour: a stadium emptying out, a commute starting, an outage rerouting traffic. Static graphs assume yesterday's map of who talks to whom still applies today, which is a shaky assumption. A model that adjusts its own spatial logic on the fly, and can show its routing decisions rather than hide them, is a meaningfully different bet than the fixed multi-task schemes it's replacing.
Milan and Trento are two cities, not a global stress test, so whether this routing trick holds up in denser or more chaotic 5G networks remains an open question.