Reproduce “An Influence-based Approach for Root Cause Alarm Discovery in Telecom Networks” on OpenResearch

About this paper

Keli Zhang, Marcus Kalander, Min Zhou, Xi Zhang, Junjian Ye

The authors rank telecom alarms to help localize faults. They learn alarm relationships from historical data, then rank incident alarms by influence using a causal graph and IRIE. On 6,000 labeled transactions, CPBE reaches 61.8% top-1 accuracy versus 57.6% for IT. This test used an HPCI graph augmented with 44 known causal edges, limiting the result to that graph.

Key insights

• A hybrid approach integrating Hawkes processes for temporal event influence and Conditional Independence tests for structural refinement effectively learns more accurate and sparser causal graphs than methods relying on either technique alone.
• Network embedding applied to causal propagation contexts provides a robust way to learn global influence weights for causal graph edges, capturing semantic similarities between alarm types based on their roles in propagating faults.
• Framing root cause identification as an influence maximization problem on a globally weighted causal graph allows for effective prioritization and ranking of potential root causes from an 'alarm flood' in real-time.

Summary by alphaXiv.