Causal Machine Learning – Summer 2019
Khoury College of Computer Sciences, Northeastern University
Prof. Robert Ness
This blog contains lecture notes and discussions from Causal Machine Learning course offered at Northeastern University in summer 2019.
How can we model and create decision making agents that develop an understanding of their environment? How can we make sure bandit algorithms are led astray by some unexpected influence? How can we measure the direct impact of a change to systems that serve ads, assign content to a feed, or direct a fleet of drivers impact bids/engagement/rides? How can we make sure our predictive algorithms are not amplifying biases in the training data in a way that leads to disastrous decision-making?
Causal inference answers these questions by enabling predictions about a system’s behavior after an intervention, something not possible even with the most advanced pattern recognition methods alone. For this reason, the skillful application of causal inference is essential for the success of in-production machine learning at scale.
This course introduces key ideas in causal inference and how to apply them to areas of machine learning where they have the most impact, such as predicting under intervention, A/B testing, bandit algorithms, reinforcement learning, agent modeling, and improving the performance of machine learning systems using past human-interaction data. The focus is on how to put these ideas into practice with real-world machine learning systems.
This course targets data scientists and ML engineers familiar with probabilistic generative modeling in machine learning. Learners familiar with how to use a tensor-based framework to build a Gaussian mixture model or a variational auto-encoder will find the material grafts directly onto this modeling intuition. This course focuses on causal probabilistic modeling and structural causal models because they fit nicely into that generative model framework and its toolsets. The course also covers elements of the Neyman–Rubin causal model that are commonplace in professional machine learning settings.
Learners who want to improve a multiarmed bandit or sequential decision process applied in a complex environment will also gain much from this course.
Learners who seek to see causal inference language from the statistics, econometrics, epidemiology, and cognitive science literature united under a set of case studies relating to the engineering of large scale machine learning systems, have also come to the right place.
See the syllabus to learn more.
Tutorial: Probabilistic modeling with Bayesian networks in bnlearn Lecture notes by Sara Taheri Installing bnlearn Open RStudio and in console type: install.packages(bnlearn) install.packages(Rgraphviz) If you experience problems installing Rgraphviz, try the following script: if (!requireNamespace("BiocManager", quietly = TRUE)) install.packages("BiocManager") BiocManager::install("Rgraphviz") ## Warning: package 'Rgraphviz' was built under R version 3.4.2...
Causal Machine Learning Pyro: Deep Universal Probabilistic Programming Prof. Robert Ness Khoury College of Computer Sciences – Northeastern University import torch import pyro pyro.set_rng_seed(101) Recap on Motivation: Our goal is to understand causal modeling within the context of generative machine learning. We just examined one generative machine learning framework called...
Causal Machine Learning - Introduction May 8th, 2019 Prof. Robert Ness Khoury College of Computer Sciences - Northeastern University 1. Introduction Discovering the direction of causality has been of interest for scientists for a long time. Researchers devote many years of their lives to discover and validate a single direction...