Radio link failure prediction in 5G networks using graph neural networks

5G networks have vitally contributed to meeting quality of service (QoS) requirements for IoT and traditional communication networks. Radio Access Networks in the 5G Infrastructure comprise radio base stations that communicate over wireless radio links. Wireless radio links have been found to be prone to weather changes. This can lead to link failure and interrupt communication. In this paper, we used Graph Attention Networks, a type of Graph Neural Networks on historical weather data and radio link site characteristics to predict radio link failure on a real telecom dataset. We compare our model with commonly used machine learning models such as support vector machines (SVM), logistic regression (LR) and Long Short-term Memory (LSTM). Our model achieves an F1-score of 0.717 performing significantly better than the other models.