Learning the Spectrum via Collaborative Filtering in Cognitive Radio Networks

Abstract

Secondary users in cognitive radio networks need to learn the statistics of spectrum in order to achieve efficient communications. Due to the spatial correlation, the efficiency of learning is improved by letting secondary users collaborate and exchange information. Due to the similarity between the collaborative learning in cognitive radio networks and the recommendation systems of electronic commerce like Amazon, the technique of collaborative filtering is applied. Prediction oriented and reward oriented criteria are proposed to derive the procedure of collaborative filtering. For the former criterion, linear prediction is used for the parameter estimation, heuristic metric is derived for channel selection, and similarity based Boltzman distribution is used for collaborator selection. For the latter criterion, the technique of multi-armed bandit is applied to maximize the total reward of spectrum access. Numerical simulation shows that the proposed collaborative filtering scheme can significantly improve the performance of spectrum learning.