Toward Interpretable and Actionable Data Analysis with Explanations and Causality

We live in a world dominated by data, where users from different fields routinely collect, study, and make decisions supported by data. To aid these users, the current trend in data analysis is to design tools that allow large-scale analytics, sophisticated predictive models, and beautiful visualizations. At this exciting time when both data and analytics tools are widely accessible to users, treating analyses as magical black boxes can painfully mislead users and make troubleshooting frustratingly time-consuming. For instance, although the perils of interpreting correlations inferred by predictive models as causation are well-documented, making such a distinction can be tricky for many users who do not have formal training in computer science or statistics. In this paper, we give an overview of our research toward bridging this gap along two main thrusts of explanations and causality. Explanations support a primary goal of data analysis – empowering users to be able to interpret the results in data analysis and troubleshoot the process. Causality complements explanations by supporting prescriptive or actionable analytics with counterfactuals and interventions, thereby helping sound decision making. In these thrusts, we explore the symbiotic relationship between core database techniques and complementary techniques from machine learning and statistics via interdisciplinary collaborations, and employ them to applications in domains like computer science education, law, and health.

Duke Scholars

Author Sudeepa Roy Computer Science