Harnessing deep learning for mobile service traffic decomposition to support network slicing
Network slicing enables operators to isolate and customize mobile network resources on a per-service basis. Provisioning resources to slices requires real-time information about the traffic demands generated by individual services, which is challenging to acquire. This talk will first present an original deep learning approach to service-level demand estimation that relies only on total traffic aggregates. Experimental results with metropolitan-scale network measurements will reveal that the proposed solution infers per-service traffic demands with 99% accuracy. The talk will then introduce a dynamic convolution operator that resolves the grid-structural data requirements of existing spatiotemporal inference models recently proposed for mobile traffic analysis. Further, it will be shown that this operator can be easily plugged into traditional LSTM architectures to deliver accurate long-term traffic predictions.