Room: Caleta

Speakers:

• Jakob Runge, Imperial College, London
• Joris Mooij, University of Amsterdam
• Manuel Gomez Rodriguez, MPI for Software Systems, Saarbruecken
• David Lopez-Paz, Facebook AI, Paris
• Kun Zhang, Carnegie Mellon University, Pittsburgh
• Alex Peysakhovich, Facebook AI
• Ioannis Tsamardinos, University of Crete
• Antti Hyttinen, University of Helsinki
• Suchi Saria, Johns Hopkins University

Abstracts

Exact Constraint-based Causal Discovery

Abstract

Causal graphical model structure discovery is a challenging part of causal inference, graphical model research and machine learning. If absence of latent confounders can be assumed, the tool of choice are exact score-based methods, as they provably find the globally optimal equivalence class of structures. Constraint-based methods on the other hand are able to handle more general model spaces with cycles and latent confounders, but offer seriously poorer accuracy. We present research that bridges this gap between constraint-based and score-based methods to some extent. We use a combination of score-based and constraint-based ideas together with Boolean (maximum) satisfiability solving, to obtain an approach that retains exactness and the generality of the model space.

Causal discovery and prediction in nonstationary environments

Abstract

Time series data and data collected across different conditions often exhibit a distribution shift phenomenon, which poses challenges for both causal discovery and prediction -- the underlying causal model may change with the distribution shift, and traditional supervised learning assumes that the training set and test set were drawn from the same distribution, which is not the case when the data distribution changes. In this talk, we emphasize and exploit the coupling relationship between causal modeling and distribution shift: causal models imply how the data-generating processes or different modules of the joint distribution may change, and distribution shift actually exhibits such changes. Accordingly, we show how causal discovery could benefit from distribution shift, and how a causal perspective helps understand and solve the problem of domain adaptation or transfer learning.

Causal Inference from Single Cell, Mass Cytometry Data: an Integrative Approach

Abstract

Single Cell biological data, and particularly Mass Cytometry Data present significant opportunities not only for discovering novel biological data, but also for serving as a testbed of Causal Discovery algorithms. Such data seem ideal for Causal Discovery as they typically contain thousands of even millions of sample points, under several experimental conditions, and often measured at several time points. On the other hand, initial attempts at Causal Discovery demonstrate the challenges of learning causality in the microworld where confounding factors and feedback cycles are abundant. In this talk, we present approaches developed and applied on mass cytometry data that are based on Causal Probabilistic Graphical Models as well as novel approaches based on Ordinary Differential Equation Models. The results and the lesson's learnt will be discussed, as well as their influence in inventing new methods that can robustly discover causality in single cell biological data.

Joint Causal Inference

Abstract

The standard method to discover causal relations is by using experimentation. Over the last decades, alternative methods have been proposed: constraint-based causal discovery methods can sometimes infer causal relations from certain statistical patterns in purely observational data. We introduce Joint Causal Inference (JCI), a novel constraint-based approach to causal discovery from multiple data sets that elegantly unifies both approaches. Compared with existing constraint-based approaches for causal discovery from multiple data sets, JCI offers several advantages: it deals with several different types of interventions in a unified fashion, it can learn intervention targets, it systematically pools data across different datasets which improves the statistical power of independence tests, and most importantly, it improves on the accuracy and identifiability of the predicted causal relations.

Large-scale causal discovery from nonlinear time series datasets

Abstract

Detecting causal associations in observational time series datasets is a key challenge for novel insights into complex dynamical systems such as the Earth system. Dependencies in such a high-dimensional dynamical system may involve time-delays, nonlinearity, and strong autocorrelations, which present major challenges for causal discovery techniques. Here we are interested in time-lagged causal discovery using conditional independence testing and address two major problems: (1) Low power due to high-dimensionality and (2) detection biases. The latter refers to the problem that the detection power for individual links may depend not only on their causal strength, but also on autocorrelation and other dependencies. Here we introduce a method for large-scale, linear and nonlinear, time-delayed causal discovery. In extensive numerical experiments we find that our method yields more power than common methods and largely overcomes detection biases allowing to more accurately rank associations in large-scale analyses by their causal strength. We demonstrate the method on a large-scale climate time series dataset.

Uncovering the Dynamics of Crowdlearning and the Value of Knowledge

Abstract

Learning from the crowd has become increasingly popular in the Web and social media. There is a wide variety of crowdlearning sites in which, on the one hand, users learn from the knowledge that other users contribute to the site, and, on the other hand, knowledge is reviewed and curated by the same users using assessment measures such as upvotes or likes. In this talk, I will present a probabilistic modeling framework of crowdlearning, which uncovers the evolution of a user's expertise over time by leveraging other users' assessments of her contributions. The model allows for both off-site and on-site learning and captures forgetting of knowledge. I will then introduce a scalable estimation method to fit the model parameters from millions of recorded learning and contributing events and show the effectiveness of our model by tracing activity of ~25 thousand users in Stack Overflow over a 4.5 year period. The model reveals that answers with high knowledge value are rare, newbies and experts tend to acquire less knowledge than users in the middle range and prolific learners tend to be also proficient contributors that post answers

Learning causal effects from many randomized experiments using regularized instrumental variables

Abstract

Scientific and business practices are increasingly resulting in large collections of randomized experiments. Analyzed together, these collections can tell us things that individual experiments in the collection cannot. We study how to learn causal relationships between variables from such collections when the number experiments is large, many experiments have very small effects, and the analyst lacks metadata (e.g., descriptions of the interventions). Here we use experimental groups as instrumental variables (IV) and show that a standard method (two-stage least squares) is biased even when the number of experiments is infinite. We show how a sparsity-inducing $l_0$ regularization can --- in a reversal of the standard bias--variance tradeoff in regularization --- reduce bias and MSE of interventional predictions. We propose a cross-validation procedure (IVCV) to feasibly select the regularization parameter. Using a trick from Monte Carlo sampling, IVCV can be done using summary statistics instead of raw data, thus making it simple to use in many real-world applications.

Discovering causation in data

Abstract

The aim of machine learning is to discover dependencies between variables, and use these dependencies to provide predictions. This is in stark contrast to other sciences, which aim at discovering causal relations between variables, and using these causal relations to provide with explanations. Since machine learning ignores the differences between dependence and causation, current machine learning algorithms excel at prediction by exploiting indirect dependencies, are unable to explain their outcomes, need large amounts of data to solve the task at hand, and fail when facing distributional shifts --- such as adversarial perturbations. In this talk, we ask the question: Can machine learning evolve and address these problems by discovering and leveraging causation? We will conjecture that this is the case, by 1) developing a novel and theoretically sustained framework to uncover causal relations between pairs of variables, 2) exemplifying the preliminary use of these methods in computer vision and natural language understanding tasks, and 3) moving towards the multivariate case thanks to the use of Generative Adversarial Networks.