Explore a 56-minute conference talk on the complexities of causal aggregation and paradoxical confounding in data analysis. Delve into the challenges of defining causality in aggregated variables, where different micro-level realizations of macro-interventions can lead to varying outcomes. Examine how this ambiguity can transform unconfounded causal relations into confounded ones and vice versa. Discover the concept of natural macro interventions and their role in aggregating cause-effect relations. Learn through practical examples, including a linear Gaussian model, to better understand these intricate concepts. Gain insights from speaker Yuchen Zhu of Valence Labs on the importance of considering micro-level states when analyzing macro causal relationships in data science and machine learning applications.
Overview
Syllabus
- Introduction
- Challenges of Causal Abstraction
- Macro Intervention and Micro-realism
- Linear Gaussian Example
- Summary
Taught by
Valence Labs
Related Courses
-
Identifiability of Causal Models and Applications to Perturb-Seq Data
-
A Causal Inference Framework for Combinatorial Interventions
-
Causal Abstractions Using Generalized Functions
-
Hidden Variables in Linear Non-Gaussian Causal Models
-
Weakly Supervised Causal Representation Learning
-
Causal Representation Learning via Soft Interventions
