[Paper Review] Causal Deep Learning

Jeroen Berrevoets, Krzysztof Kacprzyk+2 more|arXiv (Cornell University)|Mar 3, 2023

Bayesian Modeling and Causal Inference4 citations

TL;DR

This paper introduces Causal Deep Learning (CDL), a framework integrating partial causal knowledge, parametric functional forms, and temporal dynamics to enhance real-world decision-making in healthcare, economics, and environmental science. By structuring models around testable assumptions across structural, parametric, and temporal dimensions, CDL enables robust, interpretable, and generalizable causal inference in complex, high-dimensional data settings.

ABSTRACT

Causality has the potential to truly transform the way we solve a large number of real-world problems. Yet, so far, its potential largely remains to be unlocked as causality often requires crucial assumptions which cannot be tested in practice. To address this challenge, we propose a new way of thinking about causality -- we call this causal deep learning. Our causal deep learning framework spans three dimensions: (1) a structural dimension, which incorporates partial yet testable causal knowledge rather than assuming either complete or no causal knowledge among the variables of interest; (2) a parametric dimension, which encompasses parametric forms that capture the type of relationships among the variables of interest; and (3) a temporal dimension, which captures exposure times or how the variables of interest interact (possibly causally) over time. Causal deep learning enables us to make progress on a variety of real-world problems by leveraging partial causal knowledge (including independencies among variables) and quantitatively characterising causal relationships among variables of interest (possibly over time). Our framework clearly identifies which assumptions are testable and which ones are not, such that the resulting solutions can be judiciously adopted in practice. Using our formulation we can combine or chain together causal representations to solve specific problems without losing track of which assumptions are required to build these solutions, pushing real-world impact in healthcare, economics and business, environmental sciences and education, through causal deep learning.

Motivation & Objective

• To address the gap in real-world applicability of causal inference by integrating deep learning with partial, testable causal knowledge.
• To develop a unified framework that explicitly distinguishes between testable and untestable assumptions in causal modeling.
• To enable practical deployment of causal models in domains with complex, high-dimensional, and temporally evolving data.
• To support decision-making in healthcare, economics, and environmental sciences by quantitatively characterizing causal relationships over time.
• To provide a structured taxonomy for classifying and guiding research in causal deep learning across different problem types and data assumptions.

Proposed method

• CDL structures models along three dimensions: structural (partial causal knowledge, including conditional independencies), parametric (functional forms of relationships between variables), and temporal (time-evolving interactions among variables).
• The framework uses distribution factorization to encode causal assumptions, enabling model construction based on known or assumed causal graphs.
• It incorporates deep neural networks to map input data to representations that respect the assumed causal structure, functional form, and temporal dynamics.
• The method supports model pipelines that chain causal representations while tracking required assumptions, ensuring traceability and validation.
• It enables evaluation using data matching input assumptions, including interventional or longitudinal data to validate causal structure.
• A classification map is proposed to categorize CDL methods based on input assumptions (e.g., non-parametric, noise models, fully known functions) and representation types, aiding method development and comparison.

Experimental results

Research questions

• RQ1 How can causal inference be made practically applicable in real-world domains with incomplete or untestable causal assumptions?
• RQ2 What is the role of deep learning in integrating partial causal knowledge, functional forms, and temporal dynamics to improve model interpretability and robustness?
• RQ3 How can causal deep learning models be systematically validated when key assumptions cannot be tested?
• RQ4 In what ways can CDL improve generalization, domain adaptation, and robustness in supervised learning beyond standard accuracy metrics?
• RQ5 How can researchers and practitioners systematically navigate, compare, and develop new methods within the causal deep learning landscape?

Key findings

• CDL provides a structured framework that explicitly separates testable from untestable assumptions, enabling more responsible and transparent deployment of causal models.
• By incorporating partial causal structure, parametric forms, and temporal dynamics, CDL enables more accurate and interpretable modeling of complex systems such as drug response over time in healthcare.
• The framework supports the construction of model pipelines that preserve traceability of required assumptions, enhancing auditability and trust in real-world applications.
• CDL enables improved performance in out-of-distribution generalization, domain adaptation, and robustness by learning causal representations that are invariant across environments.
• The proposed classification map identifies research gaps, such as the lack of methods assuming fully known parametric functions, and guides future method development.
• Empirical validation using interventional or longitudinal data is recommended to support causal structure assumptions, particularly in medical and economic applications.