Spectrum Hole Detection For Cognitive Radio Using Machine Learning

Abstract

Detection of spectrum hole is one of the main objectives of spectrum sensing in a cognitive radio network. Spectrum hole detection for cognitive radio in machine learning, is a binary classification problem. In this work, classification  algorithms such as feed-forward neural network, , K-Nearest Neighbors, Decision Tree, Random Forest and Gradient Boosting  are used to classify the extracted features as spectrum hole detected  or not detected class. The algorithms are trained on a dataset of primary users which is a combination of several analog and digital modulations. The performance metrics are evaluated based on the confusion matrix obtained for each of the algorithms. Receiver operating characteristics obtained show that both the Random Forest and Gradient Boosting algorithms have resulted in higher classification accuracy, spectrum hole detection probability, F1 score, area under the curve and reduced misclassification probability when compared to other algorithms used in this work.