Mining KPI correlations for non-parametric anomaly diagnosis in wireless networks

Abstract

The increase in mobile data traffic has imposed unprecedented pressure on wireless network management. KPI-based anomaly diagnosis can alleviate such pressure by automatically identifying the cause of abnormalities in the traffic and providing end-to-end monitoring and optimization. Previous approaches mainly focus on finding a subset of anomaly-inducing KPIs on the basis of supervised learning procedures. These studies have two possible limitations: (1) the inherent correlations between KPIs that are proven to be effective for the anomaly diagnosis, are still largely underexplored; (2) machine learning models heavily rely on human annotations, which are expensive and labor-intensive. Therefore, we propose random matrix theory-based KPI identification (RKI), a novel method that automatically mines rich interactions between KPIs for anomaly diagnosis without using any learnable parameters or human annotations. Specifically, RKI diagnoses the abnormal KPIs in two steps. First, we build a matrix for anomaly KPI detection to mine the spectrum of its covariances. Second, another new matrix is reconstructed to calculate the correlation difference. By doing so, the anomaly KPIs that have larger correlation difference scores can be efficiently identified in the wireless traffic without any trainable parameters. In extensive experiments on a public dataset, RKI yields a 6.5% higher true diagnostic rate and 11.36% lower false alarming rate than the statistical model, demonstrating its effectiveness. A 100 × larger scale synthetic dataset also demonstrates the capabilities of RKI to explore massive data traffic under real-word scenarios. Finally, we discuss RKI’s potential applications of our method in future 6G wireless networks.