RiskRoot:Multi-Dimensional Root Causes Location Model Based on Risk Score and Iterative Search

For internet companies, multi-dimensional KPIs (Key Performance Indicators) are crucial monitoring data and it is essential for operations and maintenance personnel to rapidly and precisely identify the attribute combination causing anomalies in these KPIs. In this paper, a model named RiskRoot is proposed to address the problem of small fluctuations of anomalies or the interaction of two changes that make anomalies inconspicuous and thus difficult to locate the root cause, as well as the impact of too large search space on computational efficiency. First, all elements are divided and weighted, risk score is defined to identify root causes and non-obvious abnormalities are located through cumulative weights. Then, the hierarchical data is searched iteratively and each iteration is filtered using element pruning to ensure operational efficiency. Finally, we validate the effectiveness of RiskRoot on multiple datasets, the experimental results show that RiskRoot outperforms the state-of-the-art baseline. Its F1-score is on average 3.54 % higher than the next best method. In particular, the F1-score improves by 10.1 % in more challenging localization scenarios.