Invited Speakers

Title: Causal discovery, representation learning, and time series

Abstract: There is a deep connection between causal discovery and representation learning: Both can actually be formalized as estimation of latent-variable models. On the other hand, it has been recently shown that using temporal structure enables identifiability of latent-variable models even in very general nonlinear cases. In this talk, I explore the interplay of these concepts. In the end, I show how by judiciously using temporal structure, it is even possible to perform causal representation learning, i.e. to learn a causal model for latent variables.

Title: Commonalities and differences in the causal analysis of different types of time-structured data

Abstract: Real-world data is typically time-structured: For instance electronic health records are databases of patient information over time, cohort studies collect data in waves, panel-data are collect by regular questionnaires with occasional changes to the questions, and of course plenty of data come in form of time-series or event-histories. The causal analysis of such data must take into account that hypothetical intervention will occur at certain points in time, and may or may not be sustained or modified over time; similarly, confounding becomes time-dependent; less well-understood are the many guises of time-related selection-bias threatening the causal validity.  I will give an overview of these issues, how they can be related to graphical representation of dynamic causal structures, and how they affect all time-structured data. However, due to different modelling requirements and traditions, the solutions are often quite different. I conclude with an example of causal analysis of continuous-time event data which we formalise as stochastic interventions that change the intensities of certain types of events.

Title: Disentangling mechanisms and behaviors in complex systems

Abstract: Systems composed of many units, whose behavior goes beyond the sum of the individual behaviors of the singles, are ubiquitous. Examples relevant to what we do are the brain, the body as a whole, and the social systems we live in. When it comes to analyzing collective behavior we are often stuck with pairwise dependencies (often correlations). Another source of confusion is that the statistical dependencies we use are usually designed to address effects (or behaviors) rather than causal structures (or mechanisms). In this talk I will describe some conceptual frameworks to navigate these levels of knowledge and some approaches rooted in information theory to mine groups of variables sharing common information about the dynamics of complex systems, and provide some examples.

Title: Learning causal graphs of multivariate stochastic processes from tests of local independence or Granger causality

Abstract: Many algorithms use tests of conditional independence to learn about the causal structure in a set of random variables, e.g., the PC and FCI algorithms. Analogously, one may use tests of local independence to learn causal graphs of continuous-time multivariate stochastic processes. In discrete-time stochastic processes, one may use tests of Granger causality similarly. Both are asymmetric notions of independence that describe how a system of stochastic processes evolves over time. Let A, B, and C be three subsets of coordinate processes of the stochastic system. Intuitively speaking, B is locally independent of A given C (A is Granger noncausal for B given C) if at every point in time knowing the past of both A and C is not more informative about the present of B than knowing the past of C only.

In this talk, we will describe a graphical framework that allows causal learning in partially observed stochastic processes based on tests of local independence/Granger causality using so-called directed mixed graphs. Several directed mixed graphs may describe the same set of local independences/Granger noncausalities and therefore it is important to characterize such Markov equivalence classes. It turns out that directed mixed graphs satisfy a maximality property which allows us to construct a simple graphical representation of an entire equivalence class. This is convenient as the equivalence class can be learned from data and its graphical representation concisely describes which underlying structures could have generated the observed local independence/Granger causality structure.

In this graphical framework, several computational problems are hard. For instance, deciding Markov equivalence of two directed mixed graphs is coNP-complete. For this reason, we introduce a class of so-called weak equivalences of directed mixed graphs. These equivalences generalize Markov equivalence in that weakly equivalent graphs only need to agree on a certain subset of local independences/Granger noncausalities. We also extend the maximality result to the weak equivalence classes and this enables tractable learning of equivalence classes, even in large networks. Weak equivalence classes are less expressive than Markov equivalence classes, however, the difference is not necessarily substantial in sparse graphs. Finally, the maximality results enable straightforward edge-specific learning such that data can be seen as providing separate evidence for or against the inclusion of each edge in the output graph.

Title: Causal discovery for time series from multiple datasets with latent contexts

Abstract: Causal discovery from time series data is a typical problem setting across the sciences. Often, multiple datasets of the same system variables are available, for instance, time series of river discharge from different catchments. The local catchment systems then share certain causal parents, such as time-dependent large-scale weather over all catchments, but differ in other catchment-specific drivers, such as the altitude of the catchment. These drivers can be called temporal and spatial contexts, respectively, and are often partially unobserved. Pooling the datasets and considering the joint causal graph among system, context, and certain auxiliary variables enables us to overcome such latent confounding of system variables. In this talk, I give intuition about how temporal and spatial latent contexts can be modelled and present a non-parametric time series causal discovery method, J(oint)-PCMCI+, that efficiently learns such joint causal time series graphs when both observed and latent contexts are present in the time series setting.

Title: Enhancing Reinforcement Learning through Causal Representations and Graph Structures

Abstract: Deep reinforcement learning has witnessed significant interest in recent years, but it still faces crucial challenges such as data scarcity, interpretability concerns, and limitations in generalization. In this talk, I will show how these challenges can be mitigated through a causal perspective. Specifically, by incorporating causal representations and graph structures into the RL framework, we demonstrate how these challenges can be effectively mitigated, paving the way for more efficient and interpretable reinforcement learning algorithms.

Title: Causal AI in the industry - applications and open questions

Abstract: Causal AI has experienced a massive increase in adoption in the industry, with applications in supply chain, manufacturing, healthcare, customer retention, finance, amongst others. In this talk we'll provide an overview of some of the questions that are being asked by practitioners, and how that is creating interesting threads of research given the complexity of some of these topics.