Machine Learning in R, Class 4: Unsupervised learning (PCA, UMAP)

Class objectives:

• Explain why we might want to create a lower-dimension representation of a dataset.
• Explain the curse of dimensionality
• Explain the conceptual basis of PCA (principal component analysis) and what is meant by subsequent PCs
• Describe how PCA is used in practice and when to avoid using it.

Data set

• Leukemia
• Merge into a single csv but point students toward rmd that explains all the steps to get there.

Class outline

1. Review class 3

2. Briefly introduce supervised vs unsupervised machine learning - link out to concepts course

3. First step is exploring the data - JS uses a cocktail dataset. She does extensive cleaning up of terms - Goes through the process of building a piped mutate function the condesnces ingrediates (i.e. “fresh lemon juice|juice of a lemon” -> “lemon juice”), removes outliars. - Great example usage of pivot wider (seems important to demonstrate these in class bc they are newer functions)

4. Use recipes to implement principal componant analysis 1. Tell recipe what is going on with model (formula with no outcome) and what data to use 2. Update the role for cocktail name and category since we will use these as identifiers for roles but they are not a predictor or outcome 3. center and scale the numeric predictors (this is standard for pca) 4. use step_pca() for the actual analysis

pca_rec <- recipe(~., data = cocktails_df) %>%
  update_role(name, category, new_role = "id") %>%
  step_normalize(all_predictors()) %>%
  step_pca(all_predictors())
  
pca_prep <- prep(pca_rec)

pca_prep

Before using prep() the steps have been defined but not actually run or implemented. This last step is where everything gets evaluated.

5. Explore the results of the PCA - viz of first five PCs - discussion of how to read this plot - viz of PC1 by PC2 and PC1 by PC3 (or similar). Discussion of what these different visualizations are telling us. - Similarly discuss how coloring/labeling of points can help us make sense of PCA - Need to stress that unlike with supervised learning we aren’t given a quantitative outcome here. We’re given a data structure that we must make sense of in a more qualitative way.

6. Demonstrate the utility of tidymodels by applying to UMAP