Invited Speakers

Title: Causal Discovery in Earth Science

Abstract: I'll overview recent methods to perform observational causal discovery in Earth sciences: (1) Granger causality explicitly in reproducing kernel spaces comes with statistical guarantees and allows us to identify spatial patterns of ENSO on global moisture, (2) an efficient EM method allows us to estimate transition matrices in state-space models for operational climate model intercomparison; and (3) learning Granger spatio-temporal causal representations can be made easy with regularized variational autoencoders. I'll showcase the CAUSEME platform that implements and runs a wide variety of causal discovery methods automatically, and introduce the chatPC which uses large language models for fast causal graph prototyping when unreliable or no data is available. Many opportunities open to understand complex socio-economic-environmental system relations from data, but many challenges still remain. 

Title: Causal inference with time-varying treatments in health research

Link: Many causal questions involve treatment strategies that are sustained over time. Estimating the causal effects of sustained treatment strategies requires longitudinal data on time-varying treatments and time-varying confounders. If treatment-confounder feedback exists, then valid estimation of the causal effects requires the use of Robins’ g-methods:  g-formula, g-estimation, or inverse probability weighting. The plug-in g-formula is the most flexible g-method but also the most computationally intensive. This talk reviews a g-formula estimator and demonstrates its versatility for causal inference from observational data using several applications to health research. While most of these applications have relied on fully parametric models, efforts are ongoing to characterize the settings under which neural networks (ie, deep learning) may be advantageous.

Title: Causal representation learning in temporal settings

Abstract: Causal inference reasons about the effect of unseen interventions or external manipulations on a system. Similar to classic approaches to AI, it typically assumes that the causal variables of interest are given from the outset. However, real-world data often comprises high-dimensional, low-level observations (e.g., pixels in a video) and is thus usually not structured into such meaningful causal units. Causal representation learning aims at addressing this gap by learning high-level causal variables along with their causal relations directly from raw, unstructured data, e.g. images or videos. 

In this talk I will focus on learning causal representations in temporal sequences, e.g. sequences of images. In particular I will present some of our recent work on causal representation learning in environments in which we can perform interventions or actions. I will start by presenting CITRIS (https://arxiv.org/abs/2202.03169), where we leveraged the knowledge of which variables are intervened in each timestep to learn a provably disentangled representation of the potentially multidimensional ground truth causal variables, as well as a dynamic bayesian network representing the causal relations between these variables. I will then show iCITRIS (https://arxiv.org/abs/2206.06169), an extension that allows for instantaneous effects between variables. Finally, I will focus on our most recent method, BISCUIT (https://arxiv.org/abs/2306.09643), which overcomes one of the biggest limitations of our previous methods: the need to know which variables are intervened. In BISCUIT we instead leverage actions with unknown effects on an environment. Assuming that each causal variable has exactly two distinct causal mechanisms, we prove that we can recover each ground truth variable from a sequence of images and actions up to permutation and element-wise transformations. This allows us to apply BISCUIT to realistic simulated environments for embodied AI, where we can learn a latent representation that allows us to identify and manipulate each causal variable, as well as a mapping between each high-level action and its effects on the latent causal variables.

Title: Causal Feature Selection in the Era of Big Data

Abstract: Causal feature selection is a fundamental problem in many fields such as healthcare, economics, environmental science, etc.  Unlike traditional feature selection methods that focus on correlation or predictive power, causal feature selection seeks to uncover the underlying causal relationships. The emergence of big data has profoundly enabled the integration of machine learning algorithms into causal feature selection, significantly improving both scalability and accuracy. In this talk, I will first delve into the foundational principles of causal feature selection. Subsequently, I will explain causal feature selection within the context of time series data, highlighting the unique challenges and advanced techniques employed to discern causal relationships in temporal sequences. Lastly, I will elucidate the recent advancements in machine learning in causal feature selection for spatio-temporal data, focusing on the integration of spatial and temporal dimensions to enhance the precision and robustness of causal inferences in complex datasets.

Title: Uncovering the CO2 Fertilization Effect: Overcoming Challenges in Causal Inference for Earth System Science

Abstract: Causal modeling is a promising approach for understanding complex relationships in earth system science. This presentation demonstrates the application of causal modeling to investigate the CO2 fertilization effect, a critical process in climate science. We first employ a double machine learning approach to estimate the causal effect of CO2 on plant growth, but encounter noisy results that undermine the reliability of the findings. We then discuss the practical challenges that may have contributed to these limitations. To overcome these issues, we propose an alternative approach that leverages wind patterns as a natural intervention to identify the causal effect of CO2 fertilization. Using synthetic and real-world data, we show that this method yields more robust results, providing new insights into the CO2 fertilization effect.

Title: Granger causality in extremes

Abstract: We present a novel mathematical framework for 'Granger causality in extremes,' which aims to identify causal links from extreme events in time series.  Traditional causality methods often overlook causal mechanisms during extreme and highly volatile periods; however, the most informative data points are typically the most extreme ones. We demonstrate equivalences between causality in extremes and other causal concepts, such as Granger causality, Sims causality, and structural causality. We introduce a new model-free inference method capable of handling non-linear and high-dimensional time series, outperforming current state-of-the-art methods in performance and speed, which is particularly effective under hidden confounders. This method proves useful in applications like stock market analysis and extreme weather events.

Title: Causal discovery with tiered background knowledge

Abstract: Causal discovery methods have well-known issues: The output in form of an estimated equivalence class (represented by a CPDAG) can be sensitive to statistical errors and is often not very informative. Including background knowledge, if correct, can only improve (and never harm) the result of causal discovery. This talk will focus on tiered background knowledge as would be available in longitudinal or cohort studies, but the results presented here are valid for any kind of data that has a tiered ordering of the variables.  Causal discovery algorithms that exploit tiered background knowledge output restricted equivalence classes represented by so-called tiered MPDAGs, which can be characterised as distinct from MPDAGs based on other types of background knowledge. This class of graphs inherits key properties of CPDAGs so that they retain the usual interpretation as well as computational efficiency. Moreover, we can determine exactly when tiered knowledge adds new information, and when it is redundant. Finally, we show that the estimated graphs from an extension of the PC algorithm using tiered background knowledge are not only more informative, but also more robust to statistical errors.