Human readable network troubleshooting based on anomaly detection and feature scoring

Network troubleshooting is still a heavily human-intensive process. To reduce the time spent by human operators in diagnosis process, we present system based on (i) unsupervised learning methods for detecting anomalies domain, (ii) an attention mechanism to rank features feature space and finally (iii) expert knowledge module able seamlessly incorporate previously collected domain-knowledge. In this paper, thoroughly evaluate performance of full its individual building blocks: particularly, consider 10 anomaly detection algorithms as well mechanisms, that comprehensively represent current state art respective fields. Leveraging unique collection expert-labeled datasets worth several months real router telemetry data, perform thorough evaluation contrasting practical results constrained stream-mode settings, with achievable ideal oracle academic settings. Our experimental shows proposed effective achieving high levels agreement expert, even simple statistical approach extract useful information from gained past cases, significantly improving performance.