NORA Summer Course 2026 - Causality and Machine Learning

Causal reasoning is a key feature of scientific inquiry and human intelligence. Current artificial intelligence (AI) and machine learning (ML) models work only on a probabilistic level evaluating associations between variables; causal reasoning would allow us to develop systems able to reason and to control systems in a more grounded, reliable and interpretable way. Causality is thus crucial for the development of more trustworthy AIs, as proven by its adoption in many sensitive sub-fields of ML, including explainable AI (XAI), large language models (LLMs) benchmarking, or optimization of decision-making by autonomous agents.

The course we propose aims at introducing students to the rich theory of causality developed in the past decades, as well as to the methodologies devised to perform valid and sound causal inferences. It will start presenting the graphical and probabilistic formalism used to represent causal systems; it will discuss grounded ways to estimate causal effects relying on observational data and graphical structures; it will delve into the problem of discovering graphical structures from data; and, finally, it will review modern intersections between causality and machine learning. These topics will provide a solid understanding of the concepts related to causality, allowing the students to apply such ideas to their own domain and research.

Contents:

• Causal modelling (probabilistic graphical models, structural causal modelling, Pearl’s causal hierarchy, potential outcomes)
• Causal inference (identifiability, do-calculus, identification methods)
• Causal discovery (independence-based methods, score-based methods, functional assumption methods)
• Machine learning and causality (integration of causal inference and causal discovery in ML, ML methods for causal inference and causal discovery)

Learning outcome: A student completing the course is expected to be able to:

• Explain the importance and relevance of causal reasoning and how it differs from probabilistic reasoning.
• Be aware of the pitfalls of correlational learning, and be knowledgeable about the requirements, the strenghts and the limitations of causal reasoning.
• Distinguish between observational, interventional and counterfactual reasoning, be aware of the connections between these forms of reasoning, and express their own queries in the correct form.
• Decide whether a causal query is identifiable, and, if so, use proper causal inference methods for identification.
• Perform graph discovery on data, assess the necessary conditions to run discovery algorithms and make sense of the output.