Machine Learning for Causal Inference, Counterfactual Prediction, and Autonomous Action (CausalML)

July 15, 2018
Stockholm, Sweden

Many of the most impactful applications of machine learning are not just about prediction, but are about putting learning systems in control of selecting the right action at the right time. Examples of such systems range from search engines that act by displaying a ranking, to recommender systems, ad placement systems, medical decision support systems, conversational systems, automated trading platforms, computer games, and cyber-physical systems like self-driving cars. This focus on acting requires some causal understanding of the world, since actions are interventions that change the distribution of data unlike in standard prediction problems. This gives rise to challenging counterfactual and causal prediction problems. However, causality is only a means to an end - namely being able to take the right actions; one typically does not have the burden of providing strong proofs of causal discovery.

Another interesting property is that these systems are both producers and users of data. In particular, the logs of the selected actions and their outcomes (e.g., derived from clicks, ratings, reaction times, or revenue) can provide valuable training data for learning the next generation of the system, giving rise to some of the biggest datasets we have collected. What makes machine learning in these settings challenging is that these system logs do not fit the standard supervised learning setting, since the system in operation biases the log data through the actions it selects and outcomes remain unknown for the actions not taken. Instead, learning methods have to reason about how changes to the system will affect future outcomes.

Recent successes in establishing the theoretical foundations and designing practical algorithms that include counterfactual reasoning and estimation have given rise to an emerging research area, which this workshop aims to fully develop. In particular, this workshop will bring together researchers working on the following topics not only from the ICML community, but naturally including IJCAI-ECAI around the topic of causal inference and AAMAS around autonomous systems:

  • Predicting counterfactual outcomes
  • Estimation of (conditional) average treatment effects
  • Contextual bandit algorithms and on-policy learning
  • Batch/offline learning from bandit feedback
  • Off-policy evaluation and learning
  • Interactive experimental control vs. counterfactual estimation from logged experiments
  • Online A/B-testing vs. offline A/B-testing
  • De-biasing observational data and feedback cycles
  • Fairness of actions and causal aspects of fairness
  • Applications in online systems (e.g. search, recommendation, ad placement)
  • Applications in physical systems (e.g. cars, smart homes)
  • Applications in medicine (e.g. personalized treatment, clinical trials)

Confirmed Invited Speakers:

  • Alekh Agarwal

  • Victor Chernozhukov

  • Alexandra Chouldechova

  • Mohammad Ghavamzadeh

  • Jonas Peters

  • Don Rubin

  • Suchi Saria

  • Csaba Szepesvari

  • Stefan Wager


Thorsten Joachims (Cornell)

Nathan Kallus (Cornell)

Adith Swaminathan (Microsoft Research)

Clement Calauzenes (Criteo)

Philip Thomas (UMass Amherst)