Discovery of root causes of system failures by means of analysis of repair records

Repair records constitute an invaluable source of information for early detection of systematic failures, despite issues such as inherent noise and missing data. In this paper, we present methodology and algorithms for mining repair records to discover root causes of system failure. We employ both domain-driven and data-driven clustering approaches to reduce data noise and to consider system failures at different level of granularity. We use probabilistic graphical models for identifying potential causal relations among clusters of repair records. The models are acquired by learning from data. Our methodology and algorithms are captured in a comprehensive software environment, which assists analysts in performing an interactive discovery of root causes of failures based on repair records.