Causal analysis
Interpretable fault diagnosis based on smartphone event tracking data
Problem: The causes of smartphone quality problems are complex. Currently, data analysts mainly analyze the high-dimensional event tracking data manually to find the causes of failures, which is very time-consuming and laborious.
Approach: This project proposes a causal discovery algorithm based on multi-causal graph fusion and an efficient causal effect estimation approximation method to support the detection of reliable causal relationships. A visual analysis system is also developed to support users to interactively explore and analyze causal graphs, as well as to validate detection results.
Evaluation: Quantitative experiments on simulated data demonstrated the accuracy of our causal discovery algorithm at over 70%, and case studies demonstrate that the system helps analysts discover the cause of abnormal smartphone heating.
