Structural Causal Bottleneck Models

Simon Bing, Jonas Wahl, Jakob Runge

Keywords: causal graphical models, dimension reduction, causal representation learning, causal abstraction, identifiability

TL;DR: We introduce a class of models that assume the effects between high-dimensional variables can be described by low-dimensional summary statistics, show that we can learn these models from data and that they provide an advantage for transfer learning.

Abstract: We introduce structural causal bottleneck models (SCBMs), a novel class of structural causal models. At the core of SCBMs lies the assumption that causal effects between high-dimensional variables only depend on low-dimensional summary statistics, or *bottlenecks*, of the causes. SCBMs provide a flexible framework for task-specific dimension reduction while being estimable via standard, simple learning algorithms in practice. In addition to an analysis of identifiability in SCBMs, we provide experimental results evidencing that we can estimate bottlenecks in practice. We also demonstrate the benefit of bottlenecks for effect estimation in low-sample transfer learning settings.

Submission Number: 7