Introduction

The Working Group (GT) aims to explore and advance causal inference, causal structure learning, and causal analysis. In a world increasingly dominated by complex predictive models that learn from correlations to classify, predict, or generate new examples, these methods often fall short when it comes to generalising, interpreting, and explaining. Understanding why a decision is made or why a phenomenon occurs has become an urgent necessity. This quest for causal understanding and transparency is fundamental to building trustworthy AI systems.

Context

Traditionally, Machine Learning excels at detecting correlations and making predictions. However, correlation does not imply causation. Recognising this distinction underpins the growing interest in causal inference, which aims to identify cause-and-effect relationships. Pioneers such as Judea Pearl, with his “do-calculus” framework and directed acyclic graphs (DAGs) [Pearl, 2000], laid the theoretical foundations of the field, enabling a formal distinction between interventions and passive observations. Other approaches, such as Rubin’s causal inference model based on potential outcomes [Rubin, 1974], have also structured the field, particularly in epidemiology and statistics.

Causal structure learning methods have become essential for discovering these causal graphs from observational data, where experimentation is impossible or too costly. From constraint-based methods relying on conditional independence tests, such as the PC algorithm [Spirtes et al., 2000], to score-based optimisation approaches like Bayesian networks, and methods based on non-Gaussian additive noise (e.g., LiNGAM [Shimizu et al., 2006]), the field is booming. The goal is to reconstruct the underlying causal relationships that generated the observed data.

At the same time, the need for explainability (eXplainable AI - XAI) has become central. While ML models, especially deep neural networks, achieve impressive performance, their “black box” nature limits their adoption in critical domains such as healthcare, finance, or justice, where transparency and accountability are paramount. XAI techniques aim to make these models more transparent, enabling users to understand the reasons behind their predictions. Post-hoc approaches such as LIME (Local Interpretable Model-agnostic Explanations) [Ribeiro et al., 2016] and SHAP (SHapley Additive exPlanations) [Lundberg & Lee, 2017] have become standards for attributing feature importance to a specific prediction. More recently, interest has shifted towards causal explainability, which seeks to explain a prediction not merely by its correlations, but by the true causal factors that influenced it [Karimi et al., 2020]. This approach promises more robust and transferable explanations.

Objectives

This GT aims to:

  • Bring together researchers from the GDR RADIA around the topic of causality
  • Federate expertise around these themes.
  • Share knowledge and best practices.
  • Identify open research challenges and innovation opportunities.
  • Contribute to the development of new methodologies and tools.
  • Promote the adoption of causal inference and explainability in ML applications, fostering more responsible and understandable AI.

Research Axes

  1. Causal Structure Learning: Exploration of advanced methods for discovering causal graphs from heterogeneous, dynamic, or high-dimensional data. This includes integrating expert knowledge to guide discovery [Claassen & Heskes, 2012], handling latent variables, and detecting causality in time series (e.g., Granger causality [Granger, 1969] and its modern extensions).
  2. Causal Impact Computation and Analysis: Development of techniques to quantify the effect of specific interventions, enabling prediction of the consequences of hypothetical changes in a system. The goal is to go beyond prediction to understand “what would happen if…” (counterfactuals) [Pearl, 2000], using techniques such as average treatment effect (ATE) estimation or methods based on matching or inverse probability weighting.
  3. Causal Explainability: Linking causal inference methods with explainability techniques. How can knowledge of causal relationships improve the interpretability of ML models? How can a prediction be explained in terms of underlying causes rather than mere correlations? This includes developing causal counterfactual explanations [Wachter et al., 2017] that identify the smallest causal changes necessary to significantly alter a prediction.
  4. Applications in Machine Learning: Concrete application of these concepts in various ML scenarios, such as recommendation personalisation (where causal interventions are key to suggesting items that cause satisfaction), policy optimisation (e.g., in personalised medicine or public policy), causal bias detection, and improving model robustness and generalisability in the face of distribution shift [Peters et al., 2016].

References

  • Claassen, T., & Heskes, T. (2012). Causal discovery in the presence of latent confounders. Journal of Machine Learning Research, 13(1).
  • Granger, C. W. J. (1969). Investigating causal relations by econometric models and cross-spectral methods. Econometrica, 37(3).
  • Karimi, A. H., Schölkopf, B., & Valera, I. (2020). Towards Causal XAI: Explaining with Real Causes. arXiv preprint arXiv:2006.00947.
  • Lundberg, S. M., & Lee, S. I. (2017). A Unified Approach to Interpreting Model Predictions. Advances in Neural Information Processing Systems (NeurIPS).
  • Pearl, J. (2000). Causality: Models, Reasoning and Inference. Cambridge University Press.
  • Peters, J., Janzing, D., & Schölkopf, B. (2016). Elements of Causal Inference: Foundations and Learning Algorithms. MIT Press.
  • Ribeiro, M. T., Singh, S., & Guestrin, C. (2016). “Why Should I Trust You?”: Explaining the Predictions of Any Classifier. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining.
  • Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66(5).
  • Shimizu, S., Hoyer, P. O., Hyvärinen, A., & Kerminen, A. J. (2006). A linear non-Gaussian acyclic model for causal discovery. Journal of Machine Learning Research, 7(Oct).
  • Spirtes, P., Glymour, C., & Scheines, R. (2000). Causation, Prediction, and Search (2nd ed.). MIT Press.
  • Wachter, S., Mittelstadt, B., & Russell, C. (2017). Counterfactual Explanations Without Opening the Black Box: Automated Decisions and the GDPR. Harvard Journal of Law & Technology.