Monday 25 August 2025
9:15 - 10:30 Mats Stensrud
11:00 - 12:15 Urmi Ninad
Lunch
13:45 - 15:00 Aurore Lomet
15:30 - 16:45 Hugh Dance
17:30 - 20:00 poster session and aperitive
Tuesday 26 August 2025
9:15 - 10:30 Julius von Kügelgen
11:00 - 12:15 Ruta Binkyte
Lunch
13:45 - 15:00 Emmanuel Flachaire
15:30 - 16:45 Joshua Loftus cancelled
15:00 Closing
Mats Stensrud
Julius von Kügelgen
Urmi Ninad
Ruta Binkyte
Aurore Lomet
Emmanuel Flachaire
Hugh Dance
NEW: Slides are available here!
Mats Stensrud
On algorithm-assisted human decision making
Systems for algorithmic decision-making are on the rise. Self-driving cars have been a classical example, but such systems are also used to individualize decision rules in many other domains. In particular, the current focus on precision medicine reflects the interest in individualized decision rules, adapted to a patient's characteristics. In this talk, I will introduce new theory and methods for finding optimal decision rules. In particular, I will discuss an apparent paradox in the optimal regimes literature: in plausible decision settings, there is no formal guarantee that conventional optimal regimes, learned algorithmically from data, will outperform human decision-makers, like medical doctors. Then I will introduce superoptimal decision rules, which resolve this ostensible shortcoming. I further discuss how the superoptimal rules can be identified and estimated in different contexts, using both experimental data and (possibly confounded) observational data. The results will be illustrated by examples from medicine and economics.
Urmi Ninad
Causal discovery on vector-valued variables and consistency-guided aggregation
In this talk, we address the problem of causal discovery (CD) when the causal variables are vector-valued. We begin by exploring various data-generating models and examining the strengths and limitations of commonly used CD approaches in this setting. Given the high dimensionality of vector-valued variables, practitioners often resort to aggregations—such as averaging—for the sake of efficiency and robustness. However, in the absence of interventional data, it is difficult to assess whether such aggregated variables provide consistent abstractions that faithfully map low-level to high-level structural causal models (SCMs) and recent work has highlighted the stringent conditions required to test for such consistency.
Here we focus on the problem of consistency of aggregation for the specific task of constraint-based CD. We present the argument that the consistency of causal abstractions must be separated from the task-dependent consistency of aggregation maps. As an actionable conclusion of our findings, we propose a wrapper “Adag” to optimize a chosen aggregation consistency score for aggregate-CD, to make the output of CD over aggregate variables more reliable.
Aurore Lomet
Causal Discovery for Time Series with PC-HSIC: Kernel Methods and Computational Improvements
In situations where dependencies between variables are nonlinear or where distributional assumptions are difficult to justify, kernel-based methods provide a flexible alternative for causal discovery in multivariate time series. A kernel-based variant of the PC algorithm (PC-HSIC) integrates the Hilbert-Schmidt Independence Criterion (HSIC) for conditional independence testing, while accounting for autocorrelation and limiting distributional constraints. Preliminary experiments suggest that this approach leads to consistent causal graph reconstruction. However, the large number of independence tests required significantly increases the computational cost. To reduce this overhead, computational optimizations are introduced to accelerate HSIC-based independence testing. In parallel, an approach based on Shift-HSIC is introduced to address the performance degradation of standard HSIC under autocorrelation. This method, designed for dependent data, maintains control over type-I error. To improve scalability, approximations are introduced to reduce the computational cost of independence testing. These methods address independence testing in causal discovery for time series with consideration of computational efficiency. Future work may involve their application in causal inference pipelines and evaluation in large-scale settings.
Hugh Dance
Counterfactual Cocycles: Noise-Agnostic and Coherent Transport-based Couplings
Estimating joint distributions (a.k.a., couplings) over counterfactual outcomes are central to personalized decision-making and treatment risk assessment. Two emergent frameworks with identifiability guarantees are: (i) bijective structural causal models (SCMs), which are flexible but brittle to mis-specified latent noise; and (ii) optimal-transport (OT) methods, which avoid latent noise assumptions but can produce incoherent counterfactual transports which fail to identify higher-order couplings. In this work, we bridge the gap with \emph{counterfactual cocycles}: a framework for counterfactual transports that use algebraic structure to provide coherence and identifiability guarantees. Every counterfactual cocycle corresponds to an equivalence class of SCMs, however the cocycle is invariant to the latent noise distribution, enabling us to various sidestep mis-specification problems. We characterize the structure of all counterfactual cocycles; propose flexible and identifiable model parameterizations; introduce a novel cocycle estimator that avoids any distributional assumptions; and derive mis-specification robustness properties of the resulting counterfactual inference method. We demonstrate state-of-the-art performance and noise-robustness of counterfactual cocycles across synthetic benchmarks and a 401(k) eligibility study.
Julius von Kügelgen
Causal Representation Learning: Overview and Applications to Singe-Cell Biology
Many scientific questions are fundamentally causal in nature. Yet, existing causal inference methods cannot easily handle complex, high-dimensional data. Causal representation learning (CRL) seeks to fill this gap by embedding causal models in the latent space of a machine learning model. In this talk, I will provide an overview of CRL across a variety of settings. I will then present ongoing work on leveraging CRL methods for problems in bioinformatics, specifically for predicting the effects of unseen drug or gene perturbations from omics measurements. CRL requires rich experimental data and single-cell biology offers unique opportunities for gaining new scientific insights by leveraging such methods.
Ruta Binkyte
A Causal Perspective on Balancing Competing Goals in Trustworthy Machine Learning
As machine learning systems are deployed in increasingly sensitive and high-stakes environments, ensuring their trustworthiness is no longer optional—it’s imperative. However, core pillars of trustworthy ML—fairness, privacy, robustness, accuracy, and explainability—often come into conflict when addressed in isolation. This talk explores how causal reasoning provides a powerful framework to understand and navigate these trade-offs, with a special focus on the challenge of achieving fairness without compromising other goals.
Building on insights from the paper “Causality Is Key to Understand and Balance Multiple Goals in Trustworthy ML and Foundation Models”, we delve into practical examples where causal approaches have enabled more equitable outcomes while preserving performance and interpretability. We’ll discuss how causal graphs, counterfactual reasoning, and intervention-based analysis can expose hidden biases and offer actionable pathways toward fairer algorithms. The talk also addresses the limitations and open questions in operationalizing causality at scale, especially in the context of large foundation models.
Emmanuel Flachaire
Decomposing inequalitiies using Machine Learning and Overcoming Common Support Issues
The Kitagawa-Oaxaca-Blinder decomposition splits the difference in means between two groups into an explained part, due to observable factors, and an unexplained part. In this paper, we reformulate this framework using potential outcomes, highlighting the critical role of the reference outcome. To address limitations like common support and model misspecification, we extend Neumark’s (1988) weighted reference approach with a doubly robust estimator. Leveraging Neyman orthogonality and double machine learning, our method avoids trimming and extrapolation. This improves flexibility and robustness, as illustrated through two empirical applications. Nonetheless, we also highlight that Neumark’s reference outcome is more sensitive to the inclusion of irrelevant explanatory variables. Joint work with Bertille Picard.
Marianne Clausel, Lucas de Lara, Emilie Devijver, Luca Ganassali