Spectrum Occupancy Prediction for Land Mobile Radio Bands Using a Recommender System

Abstract

Spectrum occupancy prediction is a key enabler of agile and proactive decision-making for dynamic spectrum management. In this paper, state-of-the- art statistical models and machine learning prediction methods are evaluated on real-world occupancy time series measured in the Land Mobile Radio bands. While there is no universally best method for forecasting spectrum usage data, significant accuracy improvements are shown to be achievable by selecting a suitable prediction method for different frequencies. Motivated by this observation, we treat the problem of automating the selection of a suitable prediction method from a candidate pool as a machine learning task. An approach is proposed that recommends the best performing method for new data instances by learning from prior predictions on spectrum data with similar characteristics. The merit of this approach is shown in terms of improving the prediction accuracy compared to baseline selections without the recommender.