Interpretable Unsupervised Anomaly Detection For RAN Cell Trace Analysis
Ashima Chawla, Paul Mazhuvanchary Jacob, Saman Feghhi, Devashish Rughwani, Sven van der Meer, Sheila Fallon · 2020
The high complexity of modern communication networks requires an increasing degree of automation for performance and fault management tasks. A key task is the classification and identification of anomalous operation modes (and faults). This is important to separate them from normal operation conditions. In addition, these diagnoses should be interpretable by domain experts to (a) gain acceptance by these experts and (b) support effective root cause analysis and localisation. In this paper, we investigate the analysis of multivariant network data in order to identify anomalous data instances. Root cause analysis benefits from this by filtering features whose values lead (to some extent) to such anomalies. We are using Deep Neural Networks (DNNs), a powerful tool for anomaly detection in the telecommunication domain. We demonstrate the effectiveness of autoencoders (an unsupervised technique) to detect multivariant anomalies and anomalous features. To overcome the black box nature of neural networks (and thus increase their acceptance by domain experts), we apply SHapley Additive exPlanations (SHAP), which are used to explain the output model of a neural network.