ML4SocialScience

Class notes for University of Pennsylvania course on machine learning for social science (CRIM6012/SOCI6012)

---

Project maintained by gregridgeway Hosted on GitHub Pages — Theme by mattgraham

ML4SocialScience

These notes are best viewed at the ML4SocialScience github.io site.

These are the class notes for my course on machine learning for social science (CRIM6012/SOCI6012) that I have taught at the University of Pennsylvania since 2024. The course aims to

• build foundational skills essential for machine learning (calculus, linear algebra, probability)
• cover a range of machine learning methods, primarily supervised learning methods
• show applications on a range of social science problems
  • predicting recidivism risk
  • predicting high school dropout (NELS88)
  • exploring media censorship (Varieties of Democracy, V-Dem)
  • studying links between arrest and opiod use (NSDUH)
  • measure racial disparities in pedestrian stop outcomes
  • text analysis of officer-involved shooting incident reports
  • building a small language model based only on Crime and Punishment

Table of contents

1. Probability review
2. Naïve Bayes classifier
3. Prediction, bias, variance, and noise
   • k-nearest neighbor regression and classification
   • Example: Predict dropout risk from the NELS88 data
   • Spam example
4. Differential calculus review
5. Classification and regression trees
6. Linear algebra
   • Basic matrix operations, including matrix derivatives
   • Ordinary least squares and ridge regression
   • Multivariate Taylor series, Newton-Raphson, logistic regression, iteratively reweighted least squares (IRLS)
7. Singular value decomposition
   • Image compression
   • Image classification with emojis
8. Boosting and L1 regularization
   • Lasso
   • Forward stagewise selection
   • Gradient boosting
9. Propensity score estimation
   • Simpson’s paradox and confounders
   • Neyman-Rubin causal model
   • Propensity score weighting
     • using machine learning to estimate propensity scores
     • fastDR package
10. Neural networks
    • Backpropagation “by hand”
    • neuralnet package
    • Tensorflow and Keras
    • Convolutional layers
    • MNIST postal digits dataset
11. Text analysis
    • Working with text2vec
    • DTM and TFIDF
    • SVD for text
12. Long short-term memory (LSTM) neural networks
    • LSTM models
    • a small language model