Harnessing Collaboration to Improve the Accuracy of Throughput Prediction in Cellular Networks

Throughput prediction in cellular networks has garnered considerable interest in recent years due to its demonstrated positive impact on quality of experience. Existing proposals operate by having each user device make its own predictions, in a standalone manner, on the basis of its local measurements. Our hypothesis is that pooling of device measurements in a collaborative way can yield more accurate predictions, by allowing a broader set of observations from within a cell to be combined. To this end, we identify shortcomings in existing datasets, and then present our collaborative approach, along with an extensive evaluation. When compared to operating standalone, the results show a reduction in prediction error of up to 66% for users that have been inactive, and up to 17% for active users.