Visual Causality Analysis of Event Sequence Data.

Area of Science:

• Data Science
• Causal Inference
• Visual Analytics

Background:

• Understanding complex systems requires identifying causal mechanisms, often hidden in event sequence data.
• Observational event data (e.g., health records, clickstreams) contains causal information but is challenging to analyze due to high dimensionality and complex excitation mechanisms.
• Existing automated causal discovery methods lack explainability and human knowledge integration.

Purpose of the Study:

• To develop a visual analytics method for recovering causalities from observational event sequence data.
• To enhance causal model explainability and incorporate human expertise into the analysis.
• To improve the accuracy and interpretability of causal discovery in complex event data.

Main Methods:

• Extended Granger causality analysis within the framework of Hawkes processes.
• Developed an interactive visual analytics system for causal exploration and refinement.
• Incorporated a user-feedback mechanism for iterative improvement of causal models.
• Utilized novel visualizations and interactions for bottom-up exploration, verification, and comparison.

Main Results:

• Demonstrated quantitative improvements in causal models through user-feedback integration.
• Showcased the system's utility in diverse application domains via qualitative case studies.
• The visual analytics approach facilitates more interpretable and accurate causal discovery.

Conclusions:

• The proposed visual analytics method effectively recovers causalities in event sequence data.
• Integrating human feedback significantly enhances causal model refinement and explainability.
• This approach offers a powerful tool for understanding complex systems through event sequence analysis.