NIPS 2016

Timezone: »

Workshop

"What If?" Inference and Learning of Hypothetical and Counterfactual Interventions in Complex Systems

Ricardo Silva · John Shawe-Taylor · Adith Swaminathan · Thorsten Joachims

Fri Dec 09 11:00 PM -- 09:30 AM (PST) @ Room 133 + 134

in Saturday Workshops »

Event URL: https://sites.google.com/site/whatif2016nips/ »

One of the promises of Big Data is its potential to answer “what if?” questions in digital, natural and social systems. Whether we speak of genetic interactions in a cell, passengers commuting in railways and roads, recommender systems matching users to ads, or understanding contagion in social networks, such systems are composed of many interacting components that suggest that learning to control them or understanding the effect of shocks to a system is not an easy task. What if some railways are closed, what will passengers do? What if we incentivize a member of a social network to propagate an idea, how influential can they be? What if some genes in a cell are knocked-out, which phenotypes can we expect?

Such questions need to be addressed via a combination of experimental and observational data, and require a careful approach to modelling heterogeneous datasets and structural assumptions concerning the causal relations among the components of the system. The workshop is aimed at bringing together research expertise from a variety of communities in machine learning, statistics, engineering, and the social, medical and natural sciences. It is an opportunity for methods for causal inference, reinforcement learning and game theory to be cross-fertilized with more traditional research in statistics and the real-world constraints found in practical applications. Ultimately, this can lead to new research platforms to aid the assessment of policies, shocks and experimental design methods in the discovery of breakthroughs in a variety of domains.

Fri 11:20 p.m. - 11:30 p.m.	Welcome (Opening remarks)	🔗
Fri 11:30 p.m. - 12:15 a.m.	The Data-Fusion Problem: Causal Inference and Reinforcement Learning (Invited Talk)	Elias Bareinboim 🔗
Sat 12:15 a.m. - 1:00 a.m.	A Contextual Research Program (Invited Talk) A theory of contextual interventions has developed and matured to the point where contextual bandits can be routinely deployed to solve appropriate problems. A more general theory of contextual interventions in complex settings appears desirable and is under development leading to developments in two new areas: 1) Sequential decision making around deviations from existing solutions 2) Global exploration strategies for arbitrary contexts.	John Langford 🔗
Sat 1:00 a.m. - 2:00 a.m.	Poster Session I (Poster Session) The first of two sessions. Each session will include all posters.	🔗
Sat 2:00 a.m. - 2:30 a.m.	Optimal and Adaptive Off-policy Evaluation in Contextual Bandits (Contributed Talk) We consider the problem of off-policy evaluation—estimating the value of a target policy using data collected by another policy—under the contextual bandit model. We establish a minimax lower bound on the mean squared error (MSE), and show that it is matched up to constant factors by the inverse propensity scoring (IPS) estimator. Since in the multi-armed bandit problem the IPS is suboptimal, our result highlights the difficulty of the contextual setting with non-degenerate context distributions. We further consider improvements on this minimax MSE bound, given access to a reward model. We show that the existing doubly robust approach, which utilizes such a reward model, may continue to suffer from high variance even when the reward model is perfect. We propose a new estimator called SWITCH which more effectively uses the reward model and achieves a superior bias-variance tradeoff compared with prior work. We prove an upper bound on its MSE and demonstrate its benefits empirically on a diverse collection of datasets, often seeing orders of magnitude improvements over a number of baselines.	Yu-Xiang Wang 🔗
Sat 2:30 a.m. - 3:00 a.m.	Joint Causal Inference on Observational and Experimental Datasets (Contributed Talk) We introduce joint causal inference, a powerful formulation of causal discovery over multiple datasets in which we jointly learn both the causal structure and targets of interventions from independence test results. While offering many advantages, joint causal inference induces faithfulness violations due to deterministic relations, so we extend a recently proposed constraint-based method to deal with this type of violations. A preliminary evaluation shows the benefits of joint causal inference.	Sara Magliacane 🔗
Sat 4:45 a.m. - 5:30 a.m.	Extracting Templates from Media Event Sequences (Invited Talk) In this preliminary research we'll present early results on extracting repeatable probabilistic templates from global media-event sequences. Such patterns could hint on some weak forms of causality in the global social dynamics. As a basis, we are using the evolving graph of interlinked events generated by the "Event Registry" system (eventregistry.org), where each event is represented as an object composed from three main components: social, topical and temporal. In the analysis we will show early results on the structure of the problem and the spectrum of extracted templates from simple to hard ones.	Marko Grobelnik 🔗
Sat 5:30 a.m. - 6:00 a.m.	Estimating What-if Outcomes for Targeting Interventions in a Clinical Setting (Contributed Talk) Individuals have heterogeneous outcomes from interventions. In a clinical setting, estimating how patients will respond to different treatments is critical for targeted care. Clinicians constantly ask themselves, given a patient’s history, what would happen to the patient’s clinical trajectory if they were given one treatment versus another. However, in practice it is often unknown how the patient’s signals will change in response to treatment until that treatment is actually administered. Even then, it is impossible to observe the counterfactual from real data, i.e., what would have happened to the patient if the doctor had made a different choice. In order to solve this causal question, we use the g-formula with proper assumptions to estimate physiologic trajectories and treatment responses from observed data. To demonstrate this we model blood pressure and heart rate for patients in the intensive care unit (ICU) and estimate their responses to six types of treatments that are used in their management. These two signals are among the most commonly used vital signs in the ICU and are critical for identifying life-threatening conditions like septic and hemorrhagic shock. To model the signal with treatment response from observed data, we use two different Bayesian non-parametric (BNP) methods to build the estimator. BNP are known to have an extremely flexible functional form, which helps to overcome the model mis-specification problem and makes the estimator more robust.	Suchi Saria 🔗
Sat 6:00 a.m. - 7:00 a.m.	Poster Session II (Poster Session)	🔗
Sat 7:00 a.m. - 7:30 a.m.	Long-term Causal Effects in Policy Analysis (Contributed Talk) In most causal problems we want to evaluate the long-term effects of policy changes but only have access to short-term experimental data. For example, for the long-term effects of minimum wage increase we may only have access to one-year worth of employment data. In this technical note we argue that such conceptual gap between what is to be estimated and what is in the data has not been adequately addressed. To make our criticism constructive we describe our approach in studying multiagent systems and the long-term effects of interventions in such systems. Central to our approach is behavioral game theory, where a behavioral model of how agents act conditional on their latent behaviors is combined with a temporal model of how behaviors evolve.	Panagiotis Toulis 🔗
Sat 7:30 a.m. - 8:15 a.m.	Causal Inference for Recommendation Systems (Invited Talk) We develop a causal inference approach to recommender systems. Observational recommendation data contains two sources of information: which items each user decided to look at and which of those items each user liked. We assume these two types of information come from different models---the exposure data comes from a model by which users discover items to consider; the click data comes from a model by which users decide which items they like. Traditionally, recommender systems use the click data alone (or ratings data) to infer the user preferences. But this inference is biased by the exposure data, i.e., that users do not consider each item independently at random. We use causal inference to correct for this bias. On real-world data, we demonstrate that causal inference for recommender systems leads to improved generalization to new data. (Joint work with Dawen Liang and Laurent Charlin)	David Blei 🔗
Sat 8:15 a.m. - 9:00 a.m.	Panel & Closing (Panel Discussion and Closing Remarks)	🔗

Author Information

Ricardo Silva (University College London)

John Shawe-Taylor (University College London)

Adith Swaminathan (Cornell University)

Thorsten Joachims (Cornell)

More from the Same Authors