Primary Submission Category: Applications in Physical Sciences, Engineering, Environment and Miscellaneous Applications
Challenges and Opportunities in Causal Inference with Complex Treatments
Authors: Michael Valancius,
Presenting Author: Michael Valancius*
Most causal inference research has centered on estimating average effects of simple, well-defined interventions. While these methods have been highly successful, further innovation is needed to tackle the challenges posed when treatments are complex, high-dimensional objects such as images, audio, or text. The rapid proliferation of AI systems has only amplified this challenge: their outputs are rarely directly manipulable, defy straightforward characterization, and pose significant obstacles for causal discovery and intervention.
We argue that the next frontier for causal inference lies in developing methods that can reason about nuanced, individualized, and counterfactual questions in these complex settings. For example: Would a different dub of a specific movie have improved member satisfaction? Or, how can we define the relevant quality components of a video to understand their causal impact? In many real-world applications, particularly in R&D and creative support, actionable feedback and structured discovery are as important as traditional effect estimation, and often require “reverse” causal reasoning.
We illustrate these challenges and opportunities with examples from Netflix, where treatments like dubs and artwork are inherently complex, and understanding their causal impact is essential for supporting creative innovation. We discuss limitations of current methodologies, highlight open questions, and share early modeling directions.