Robust Factorial Causal Predictions with Observational and Interventional Data
Abstract
Causal inference requires many assumptions and poses serious statistical challenges. Machine learning methods for causal discovery aim at mitigating the former, but can still be sensitive to general assumptions about the stability of constraints that are exploited in order to infer causal structures, such as independence constraints. Statistical challenges amplify this issue. This project addresses the need for more robust algorithms for identifying causal structure and effects from a mixture of observational and interventional data.In particular, we will introduce families of algorithms that emphasizes the role of measuring as many confounders as possible outside the causal network of interest; enacting procedures to assess the degree in which the inferred causal structures vary as we allow for violations of the stability assumptions commonly used in machine learning; integrate the concept of causal discovery with predictive and bandit-based modeling of response to actions, allowing black-box models to leverage the information that we obtain from observing mediators between our actions and the outcome of interest.This project is of direct relevance to the Mathematical Data Science Program (Division 311). In particular, it is targeted to the focus areas of complex networks and of multi-modal, multi-scale integration. Within complex networks, we contribute to the goals of ~determining causal effects and influences.~ To be able to best aid decision making, it is well-understood that correlations alone do not tell the full story, hence advances in the area of inferring causal effects are essential to ONR as a whole. Moreover, we stress the need to make the most of multiple sources of evidence through the combination of observational and interventional data. This fits well the idea of principled information fusion, which is also core to the Mathematical Data Science Program.Publication targets include several conference submissions relating to specific items described in the full proposal. Conference venues will be top machine learning events such as the International Conference in Machine Learning (ICML), Neural Information Processing Systems (NIPS), Uncertainty in Artificial Intelligence (UAI) and Artificial Intelligence and Statistics (AISTATS). Two journal submissions are planned, focusing on the main results on robustness to confounding and sampling coverage, and integration of causal inference methods in bandit modeling. Journal submissions will include several details not covered in the original conference submissions.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 15, 2019
- Source ID
- N629091912096
Entities
People
- Ricardo S Silva
Organizations
- Office of Naval Research
- United States Navy
- University College London