Time: Wednesdays 11:35–12:50 and 13:10–14:25
Location: Upson Hall 202 (Ithaca) and Bloomberg Center 091 (New York)
Note: Physical and virtual lectures will alternate periodically between Ithaca and New York. The first two lectures will be physically in Ithaca and virtually in New York.
Registration options: Both grade and non-grade allowed. Auditors allowed.
Prerequisites: familiarity with basic statistics, probability, and calculus, or permission of instructor.
Reading material: The course will follow instructor’s lecture notes and some external reading materials such as book chapters and research papers, which will be distributed electronically.
Grade breakdown: participation and lecture scribing (25%), homework (25%), final project (50%).
Nathan Kallus
http://www.nathankallus.com
kallus@cornell.edu
Office hours by appointment
The course introduces students to fundamental principles in causality and machine learning for decision making. Some of the most impactful applications of machine learning, whether in online marketing and commerce, personalized medicine, or data-driven policymaking, are not just about prediction but are rather about taking the right action directed at the right target at the right time. Actions and decisions, unlike predictions, have consequences and so, in seeking to take the right action, one must seek to understand the causal effects of any action or action policy, whether through active experimentation or analysis of observational data. In this course, we will study the interaction of causality and machine learning for the purpose of making decisions. In the case of known causal effects, we will briefly review the theory of generalization as it applies to designing action policies and systems. We will then study causal inference and estimation of unknown causal effects using both classical methods and modern machine learning and optimization methods, considering a variety of settings including controlled experiments (A/B testing), regression discontinuity, instrumental variables, and general observational studies. We will then study the direct design of action policies and systems when causal effects are not known, looking closely both at the online (contextual bandit) and offline (off-policy learning) cases. Finally, we will study ancillary consequences of intelligent systems’ actions, such as algorithmic fairness. The course will culminate in a final project.
decision theory, learning, and generalization
foundations of causality
controlled experiments, optimal design, inference on effects
instrumental variables analysis
regression discontinuity designs
optimal matching and propensity scores for observational studies
covariate shift and domain adaptation
double robustness and semi-parametric efficiency
individual-level and heterogeneous causal effects
sequential decision making and contextual bandits
policy learning for personalization from observational data
algorithmic fairness of decision policies
https://docs.google.com/document/d/1QbsKXkNbdisH4E3P4TnsuERHJcSbUGQIue2tVa0v3iQ/
| Lecture | Topic |
| 1 | Introduction and course overview |
| 2 | What is causality and basic elements of statistical learning and decision theory |
| 3 | Generalization with known effects: Doob Martingale and McDiarmid's Inequality |
| 4 | Generalization with known effects: Rademacher averages, Massart, VC Dimension |
| 5 | Unknown effects: introduction to causal inference from controlled experiments |
| 6 | Inference on effect: parametric and permutation tests of hypotheses |
| 7 | Controlled-experimental design and covariate balance |
| 8 | Optimal covariate balance in experimental design |
| 9 | Unknown effects without experimental control: introduction to observational data |
| 10 | Noncompliance and instrumental variable analysis |
| 11 | Causal inference from regression discontinuity designs |
| 12 | Causal inference under unconfoundedness: matching and regression |
| 13 | Causal inference under unconfoundedness: propensity scores |
| 14 | Double robustness and semi-parametric efficiency |
| 15 | Generalized optimal matching |
| 16 | Generalized optimal matching: analysis |
| 17 | Individual-level and heterogeneous causal treatment effects |
| 18 | Sequential decision making: lower bounds |
| 19 | Sequential decision making: algorithms |
| 20 | Sequential decision making: contextual bandits |
| 21 | Causal policy learning from observational data |
| 22 | Causal policy learning from observational data: extensions |
| 23 | Balanced policy learning |
| 24 | Balanced policy learning: analysis |
| 25 | Fairness in supervised learning |
| 26 | Fairness in policy learning |
| 27 | Final project presentations |
| 28 | Final project presentations |