Machine Learning 4 Data Science
From prediction to action
Supervised machine learning methods excel at predicting an outcome. But being able to predict an outcome does not mean we know how to change it, or that we should.
Less interpretable methods
Neural networks and ensemble methods like bagging, random forests, and boosting can greatly increase predictive accuracy at the cost of ease of interpretation.
More nonlinear methods
We continue our exploration of non-linear supervised machine learning approaches including tree based methods, GAMs, and neural networks and graphs structured learning.
Nonlinear methods
Non-linearity may result in models that trade interpretability for increased predictive accuracy. These notes discuss the challenges of non-linearity and introduce nearest neighbors and kernel methods.
Regularization and validation
When optimizing an ML model there are a variety of strategies to improve generalization from the training data. We can add a complexity penalty to the loss function, and we can evaluate the loss function on validation data.
Optimization and overfitting
Optimization is about finding the best model. With greater model complexity it becomes increasingly important to avoid overfitting: finding a model that is best for one specific dataset but does not generalize well to others.
Classification
Categorical or qualitative outcome variables are ubiquitous. We review some supervised learning methods for classification, and see how these may be applied to observational causal inference.
Multiple regression and causality
Multiple linear regression does not, by default, tell us anything about causality. But with the right data and careful interpretation we might be able to learn some causal relationships.
Linear regression
Reviewing linear regression and framing it as a prototypical example and source of intuition for other machine learning methods.
Introduction and foundations
A brief introduction to the course, preview of things to come, and some foundational background material.