Data-Driven Semi-Supervised Anomaly Detection Using Real-World Call Data Record

5G and beyond networks are expected to provide ubiquitous, ultra-reliable low latency connectivity to cellular users. Maintaining this stringent B5 performance requirement will be a challenging task for cellular service providers. A key factor that may affect network performance will be anomalies such as sleeping cells or congestion due to high traffic volumes. In the worst cases, these anomalies may cause a partial or complete networkoutage.Traditionaloutagemanagementtechnique s,such as drive-testing, may prove unsuitable in the B5G era as they are time consuming and costly. These outdated mechanisms are also unable to provide real-time data analysis. Hence future networks will rely on data-driven self-organizing networks (SON) with selfhealing capabilities to detect anomalies. Machine learning will be an essential component of such systems. In this paper we have proposed a semi-supervised learning algorithm to detect anomaly using real-world Spatio-temporal call data records (CDRs). We will demonstrate that our proposed algorithm can detect anomalies with high accuracy. The CDR is collected for the entire city of Milan, Italy in the form of spatial grids. We will demonstrate that once trained using the single-cell grid record, our model can accurately predict anomalies for the neighboring grids as well.