A Causal Inference Framework for Combinatorial Interventions

Explore a comprehensive framework for causal inference with combinatorial interventions in this 54-minute talk by Anish Agarwal from Valence Labs. Delve into the challenges of estimating unit-specific potential outcomes for multiple interventions, addressing issues of scalability and confounding in observational data. Learn about a novel latent factor model that imposes structure across units and intervention combinations, enabling identification of causal parameters despite unobserved confounding. Discover the Synthetic Combinations estimation procedure and its advantages in terms of sample complexity and consistency. Examine the application of this framework to real-world scenarios such as movie recommendations and ranking interventions. Follow along as the speaker covers topics including potential outcomes as Boolean functions, representation methods, combinatorial structure, sparsity exploitation, model details, assumptions, and a summary of key findings.

- Discussant Slide
- Introduction
- Potential Outcomes as Boolean Functions
- How to Represent Boolean Functions
- Combinatorial Structure
- Exploiting Sparsity
- Model
- Synthetic combinations
- Assumptions
- Summary