AutoARTS: Taxonomy, Insights and Tools for Root Cause Labelling of Incidents in...
summary
Empirical study of over 2,000 incidents across 450+ Microsoft Azure services yields a 4-level hierarchical Arts taxonomy with 346 distinct contributing factors and the finding that 75% of incidents have multiple causes (average four), contradicting single-root-cause practice. An automated pipeline AutoARTS uses a graph convolutional hierarchy (higherGM), domain-tuned text encoders, and a fine-tuned Pegasus summarizer to extract context; hierarchical classification hits a weighted F1 of 89.63 and a user study rates generated context 4.6/5 while revealing missed factors.
AI-written summary · about
chapters (9)
- 0:00AutoARTS project
- 3:29Manual root cause labeling challenges
- 7:23Incident analysis insights
- 10:10AutoARTS taxonomy benefits
- 11:13AutoARTS automated labeling tool
- 13:40Root cause classification challenges
- 15:23Context extraction and evaluation
- 17:32Tool evaluation and feedback
- 18:36Summary and open-source call
related talks
32mGenerative AI for Observability: Automating Root Cause Analysis in Modern IT Systems
20mAutomated Incident Management
19mIncident Archaeology: Extracting Value from Paperwork and Narratives
38mHolistic root cause analysis of software breakages through structural causal modeling
30mThere is No Root Cause...
24mAutomated Root Cause Analysis
19mPrincipled Identification of "Root Causes" Using Techniques...
35mAgentic Incident Response