In EmilHvitfeldt/celery: A Common API to Clustering

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

library(tidyclust)
library(palmerpenguins)
library(tidymodels)

Fit

kmeans_spec <- k_means(k = 5) %>%
  set_engine("stats")

penguins_rec_1 <- recipe(~ ., data = penguins) %>%
  update_role(species, island, new_role = "demographic") %>%
  step_dummy(sex)


penguins_rec_2 <- recipe(species ~ ., data = penguins) %>%
  step_dummy(sex, island)

wflow_1 <- workflow() %>%
  add_model(kmeans_spec) %>%
  add_recipe(penguins_rec_1)


wflow_2 <- workflow() %>%
  add_model(kmeans_spec) %>%
  add_recipe(penguins_rec_2)

We need workflows!

# dropping NA first so rows match up later, this is clunky
pen_sub <- penguins %>% 
  drop_na() %>%
  select(-species, -island, -sex)

kmeans_fit <- kmeans_spec %>% fit( ~., pen_sub)

kmeans_fit %>%
  predict(new_data = pen_sub)

### try my new version
kmeans_fit %>%
  extract_cluster_assignment()

Needed flexclust install; probably not necessary, we could implement for kmeans with just dists
Missing values should probably return an NA prediction. Or for k-means, imputation isn't crazy...
We want a consistent return, and k-means is randomized. We should agree to a consistent default ordering throughout tidyclust - maybe by size (number of members) or something?

penguins %>%
  drop_na() %>%
  mutate(
    preds = predict(kmeans_fit, new_data = pen_sub)$.pred_cluster
  ) %>%
  count(preds, sex, species)

Diagnostics

Measure cluster enrichment with demographic variables (an official recipe designation?). Include Chi-Square or ANOVA tests???
Exploit confusion matrix from yardstick
Characterization of clusters, presumably by centers?
Automatic plot?!

talk to Emil

Ordering of clusters is really bothering me something with indices
Cluster density followup: is it kind of a model?

recipe( ~ demo1 + predictor1) %>%
  step_tidyclust(kmeans_fit) # doesn't quite make sense

Some of these followups feel like they need a recipe. Which vars are we using for within SS? Which vars for enrichments? Do we PCA first? etc.

get_SS(Cluster ~ v1 + v2)
recipe(Cluster ~ v1 + v2) %>%
  step_pca() %>%
  get_ss()

... or maybe fit enrichment() on a recipe and it automatically uses the variables with the "enrich" role?
how would PCA fit in on this?

How does fit() access the right variables?

... is this even worth it given we can attach cluster assignments? I say yes, because what if we are trying to "cross-validate".

penguins_2 <- penguins %>%
  drop_na() %>%
  mutate(
    preds = predict(kmeans_fit, new_data = pen_sub)$.pred_cluster
  ) 

debugonce(enrichment)
penguins_2 %>%
  enrichment(preds, species)