Rolling forecasting forward by boosting heterogeneous kernels

Abstract

The problem discussed in this paper stems from a project of cellular network traffic prediction, the primary step of network planning striving to serve the continuously soaring network traffic best with limited resource. The traffic prediction emphasizes two aspects: (1) how to exploit the potential value of physical and electronic properties for tens of thousands of wireless stations, which may partly determine the allocation of traffic load in some intricate way; (2) the lack of sufficient and high-quality historical records, for the appropriate training of long-term predictions, further aggravated by frequent reconfigurations in daily operation. To solve this problem, we define a general framework to accommodate several variants of multi-step forecasting, via decomposing the problem into a series of single-step vector-output regression tasks. They can further be augmented by miscellaneous attributive information, in the form of boosted multiple kernels. Experiments on multiple telecom datasets show that the solution outperforms conventional time series methods on accuracy, especially for long horizons. Those attributes describing the macroscopic factors, such as the network type, topology, locations, are significantly helpful for longer horizons, whereas the immediate values in the near future are mainly determined by their recent records.