Unsupervised learning

Grouping or categorizing observational units (objects) without any pre-assigned labels or scores (no outcome information!)

Some examples:

• Latent Dirichlet Allocation: Topic Modeling of TSL Articles

• Network & Clustering: Characters in ‘Love Actually’

Distance measures (for clustering)

• Euclidean Distance

$$d_E(\mathbf{\text{x}},\mathbf{\text{y}})=\sqrt{\sum _{i=1}^p(x_i-y_i{)}^2}$$

• Pearson Correlation Distance

$$d_P(\mathbf{\text{x}},\mathbf{\text{y}})=1-r_P(\mathbf{\text{x}},\mathbf{\text{y}})$$ $$\text{ or }{d}_P(\mathbf{\text{x}},\mathbf{\text{y}})=1-|r_P(\mathbf{\text{x}},\mathbf{\text{y}})|$$ $$\text{ or }{d}_P(\mathbf{\text{x}},\mathbf{\text{y}})=1-(r_P(\mathbf{\text{x}},\mathbf{\text{y}}){)}^2$$

Correlation distance isn’t a distance metric!

x1 <- c(1,2,3)
x2 <- c(1, 4, 10)
x3 <- c(9, 2, 2)

# d(1,2)
1 - cor(x1, x2)

[1] 0.01801949

# d(1,3)
1 - cor(x1, x3)

[1] 1.866025

# d(2,3)
1 - cor(x2, x3)

[1] 1.755929

# d(1,3) > d(1,2) + d(2,3)
1 - cor(x1, x2) + 1 - cor(x2, x3)

[1] 1.773948

Correlation distance isn’t a distance metric!

Using absolute distance doesn’t fix things.

# d(1,2)
1 - abs(cor(x1, x2))

[1] 0.01801949

# d(1,3)
1 - abs(cor(x1, x3))

[1] 0.1339746

# d(2,3)
1 - abs(cor(x2, x3))

[1] 0.2440711

# d(2,3) > d(1,2) + d(1,3)
1 - abs(cor(x1, x2)) + 1 - abs(cor(x1, x3))

[1] 0.1519941

More distance measures

• Cosine Distance

$$d_C(\mathbf{\text{x}},\mathbf{\text{y}})=\frac{\mathbf{\text{x}}\cdot \mathbf{\text{y}}}{||\mathbf{\text{x}}||||\mathbf{\text{y}}||}$$ $$=\frac{\sum _{i=1}^p{x}_i{y}_i}{\sqrt{\sum _{i=1}^p{x}_i^2\sum _{i=1}^p{y}_i^2}}$$ $$=1-r_P(\mathbf{\text{x}},\mathbf{\text{y}})\text{if }\overline{\mathbf{\text{x}}}=\overline{\mathbf{\text{y}}}=0$$

More distance measures

• Hamming Distance

$$\begin{array}{r}{d}_H(\mathbf{\text{x}},\mathbf{\text{y}})=\sum _{i=1}^{p}I(x_i\ne y_i)\end{array}$$

The Hamming distance across the two DNA strands is 7.

dist function in R

The function dist in R calculates the distances given above.

String distances

Comparison of string distance metrics from https://www.kdnuggets.com/2019/01/comparison-text-distance-metrics.html.

Hierarchical Clustering

is a set of nested clusters that are organized as a tree. Note that objects that belong to a child cluster also belong to the parent cluster.

Definitions

Agglomerative methods start with each object (e.g., gene, penguin, etc.) in its own group. Groups are merged until all objects are together in one group.

Divisive methods start with all objects in one group and break up the groups sequentially until all objects are individuals.

Single Linkage algorithm defines the distance between groups as that of the closest pair of individuals.

Complete Linkage algorithm defines the distance between groups as that of the farthest pair of individuals.

Average Linkage algorithm defines the distance between groups as the average of the distances between all pairs of individuals across the groups.

Toy example

of Single Linkage Agglomerative Hierarchical Clustering

see class notes to walk through the process.

Should we use hierarchical clustering?

strengths

• Provides a clustering for a range of values of $k$.
• Can be used with any distance metric.

shortcomings

• Forces a hierarchical structure (almost always agglomerative)
• Linkage choice can change the outcome.

Hierarchical clustering in R

penguins <- penguins |>
  select(bill_length_mm, bill_depth_mm,
         flipper_length_mm, body_mass_g,
         island, species) |>
  drop_na()

hc_rec <- recipe(~ ., data = penguins) |>
  update_role(island, new_role = "ID") |> 
  update_role(species, new_role = "ID") |>
  step_rm(all_nominal()) |>
  step_normalize(all_predictors())

hc_model <- hier_clust(
  num_clusters = 3,
  linkage_method = "average") |> 
  set_engine("stats")

hc_model

Hierarchical Clustering Specification (partition)

Main Arguments:
  num_clusters = 3
  linkage_method = average

Computational engine: stats 

hc_work <- workflow() |> 
  add_recipe(hc_rec) |> 
  add_model(hc_model)

hc_work

══ Workflow ════════════════════════════════════════════════════════════════════
Preprocessor: Recipe
Model: hier_clust()

── Preprocessor ────────────────────────────────────────────────────────────────
2 Recipe Steps

• step_rm()
• step_normalize()

── Model ───────────────────────────────────────────────────────────────────────
Hierarchical Clustering Specification (partition)

Main Arguments:
  num_clusters = 3
  linkage_method = average

Computational engine: stats 

hc_fit <- hc_work |>
  fit(data = penguins)

hc_fit |>
  summary()

        Length Class      Mode   
pre     3      stage_pre  list   
fit     2      stage_fit  list   
post    1      stage_post list   
trained 1      -none-     logical

hc_summary <- hc_fit |> extract_fit_summary()

hc_summary |> str()

List of 7
 $ cluster_names         : Factor w/ 3 levels "Cluster_1","Cluster_2",..: 1 2 3
 $ centroids             : tibble [3 × 4] (S3: tbl_df/tbl/data.frame)
  ..$ bill_length_mm   : num [1:3] -0.369 0.603 2.253
  ..$ bill_depth_mm    : num [1:3] 0.617 -1.123 -0.368
  ..$ flipper_length_mm: num [1:3] -0.65 1.13 2.05
  ..$ body_mass_g      : num [1:3] -0.612 1.06 1.977
 $ n_members             : int [1:3] 219 119 4
 $ sse_within_total_total: num [1:3] 283.19 111.49 2.01
 $ sse_total             : num 656
 $ orig_labels           : NULL
 $ cluster_assignments   : Factor w/ 3 levels "Cluster_1","Cluster_2",..: 1 1 1 1 1 1 1 1 1 1 ...

hc_fit |> extract_centroids()

# A tibble: 3 × 5
  .cluster  bill_length_mm bill_depth_mm flipper_length_mm body_mass_g
  <fct>              <dbl>         <dbl>             <dbl>       <dbl>
1 Cluster_1         -0.369         0.617            -0.650      -0.612
2 Cluster_2          0.603        -1.12              1.13        1.06 
3 Cluster_3          2.25         -0.368             2.05        1.98 

Prediction

hc_preds <- hc_fit |> predict(penguins)

hc_preds

# A tibble: 342 × 1
   .pred_cluster
   <fct>        
 1 Cluster_1    
 2 Cluster_1    
 3 Cluster_1    
 4 Cluster_1    
 5 Cluster_1    
 6 Cluster_1    
 7 Cluster_1    
 8 Cluster_1    
 9 Cluster_1    
10 Cluster_1    
# ℹ 332 more rows

It’s important to note that there is no guarantee that predict() on the training data will produce the same results as extract_cluster_assignments(). The process by which clusters are created during the agglomerations results in a particular partition; but if a training observation is treated as new data, it is predicted in the same manner as truly new information.

Prediction vs. label

bind_cols(hc_preds,
  extract_cluster_assignment(hc_fit) ) |> 
  select(.pred_cluster, .cluster) |> 
  table()

             .cluster
.pred_cluster Cluster_1 Cluster_2 Cluster_3
    Cluster_1       218         0         0
    Cluster_2         0        99         0
    Cluster_3         1        20         4

The basic dendrogram

library(ggdendro)
penguin_hclust <- hc_fit$fit$fit$fit
penguin_dend <- penguin_hclust |> as.dendrogram()
ggdendrogram(penguin_dend)

Zooming in on a cluster

ggdendrogram(penguin_dend[[1]])

ggdendrogram(penguin_dend[[2]])

ggdendrogram(penguin_dend[[1]][[1]])

Colored dendrogram

library(dendextend)
penguin_dend |>
  set("branches_k_color", k = 3) |>  # color of the branches
  set("labels_colors", k = 3) |>     # color of the labels
  set("branches_lwd", 0.3) |>        # width of the branches
  set("labels_cex", 0.3) |>          # size of the label 
  as.ggdend() |> 
  ggplot(horiz = TRUE)

We don’t usually have a way to check (i.e., “unsupervised”)

penguins |> 
  drop_na(bill_length_mm, bill_depth_mm, flipper_length_mm, body_mass_g) |>
  select(island) |> 
  cbind(cluster = cutree(penguin_hclust, k = 3) ) |> 
  table()

           cluster
island        1   2   3
  Biscoe     44 119   4
  Dream     124   0   0
  Torgersen  51   0   0

penguins |> 
  drop_na(bill_length_mm, bill_depth_mm, flipper_length_mm, body_mass_g) |>
  select(species) |> 
  cbind(cluster = cutree(penguin_hclust, k = 3) ) |> 
  table()

           cluster
species       1   2   3
  Adelie    151   0   0
  Chinstrap  68   0   0
  Gentoo      0 119   4

$k$-means Clustering

$k$-means clustering is an unsupervised partitioning algorithm designed to find a partition of the observations such that the following objective function is minimized (find the smallest within cluster sum of squares):

$$\text{arg}\underset{C_1,\dots ,C_k}{min}\{\sum _{k=1}^K\sum _{i\in C_k}\sum _{j=1}^p(x_{ij}-{\bar{x}}_{kj}{)}^2\}$$

A fun applet!!

https://www.naftaliharris.com/blog/visualizing-k-means-clustering/

Should we use $k$-means?

strengths

• No hierarchical structure / points can move from one cluster to another.
• Can run for a range of values of $k$.

shortcomings

• $k$ has to be predefined to run the algorithm.
• $k$-means is based on Euclidean distance (only).

$k$-means clustering in R

library(ClusterR)
penguins <- penguins |>
  select(bill_length_mm, bill_depth_mm,
         flipper_length_mm, body_mass_g,
         island, species) |>
  drop_na()

kmean_rec <- recipe(~ ., data = penguins) |>
  update_role(island, new_role = "ID") |> 
  update_role(species, new_role = "ID") |>
  step_rm(all_nominal()) |>
  step_normalize(all_predictors())

kmean_model <- k_means(
  num_clusters = 3) |> 
  set_engine("ClusterR", initializer = "random") # need the ClusterR package

kmean_model

K Means Cluster Specification (partition)

Main Arguments:
  num_clusters = 3

Engine-Specific Arguments:
  initializer = random

Computational engine: ClusterR 

kmean_work <- workflow() |> 
  add_recipe(kmean_rec) |> 
  add_model(kmean_model)

kmean_work

══ Workflow ════════════════════════════════════════════════════════════════════
Preprocessor: Recipe
Model: k_means()

── Preprocessor ────────────────────────────────────────────────────────────────
2 Recipe Steps

• step_rm()
• step_normalize()

── Model ───────────────────────────────────────────────────────────────────────
K Means Cluster Specification (partition)

Main Arguments:
  num_clusters = 3

Engine-Specific Arguments:
  initializer = random

Computational engine: ClusterR 

kmean_fit <- kmean_work |>
  fit(data = penguins)

kmean_fit |>
  summary()

        Length Class      Mode   
pre     3      stage_pre  list   
fit     2      stage_fit  list   
post    1      stage_post list   
trained 1      -none-     logical

kmean_summary <- kmean_fit |> extract_fit_summary()

kmean_summary |> str()

List of 7
 $ cluster_names         : Factor w/ 3 levels "Cluster_1","Cluster_2",..: 1 2 3
 $ centroids             : tibble [3 × 4] (S3: tbl_df/tbl/data.frame)
  ..$ bill_length_mm   : num [1:3] -1.047 0.66 0.656
  ..$ bill_depth_mm    : num [1:3] 0.486 0.816 -1.098
  ..$ flipper_length_mm: num [1:3] -0.89 -0.286 1.157
  ..$ body_mass_g      : num [1:3] -0.769 -0.374 1.09
 $ n_members             : int [1:3] 132 87 123
 $ sse_within_total_total: num [1:3] 122 113 143
 $ sse_total             : num 1364
 $ orig_labels           : int [1:342] 1 1 1 1 1 1 1 1 2 1 ...
 $ cluster_assignments   : Factor w/ 3 levels "Cluster_1","Cluster_2",..: 1 1 1 1 1 1 1 1 2 1 ...

kmean_fit |> extract_centroids()

# A tibble: 3 × 5
  .cluster  bill_length_mm bill_depth_mm flipper_length_mm body_mass_g
  <fct>              <dbl>         <dbl>             <dbl>       <dbl>
1 Cluster_1         -1.05          0.486            -0.890      -0.769
2 Cluster_2          0.660         0.816            -0.286      -0.374
3 Cluster_3          0.656        -1.10              1.16        1.09 

Partitioning Around Medoids (PAM)

Find the observations (data values!) $m_k$ that solve:

$$\text{arg}\underset{C_1,\dots ,C_k}{min}\{\sum _{k=1}^K\sum _{i\in C_k}d(x_i,m_k)\}$$

Important: $m_k$ is a data point. It is not an average of points or any other summary statistic.

Should we use PAM?

strengths

• No hierarchical structure / points can move from one cluster to another.
• Can run for a range of values of $k$.
• It can use any distance measure.

shortcomings

• $k$ has to be predefined to run the algorithm.

Evaluating clustering (which $k$?)

• Silhouette Width (use $k$ with smallest silhouette width) – good for PAM (because uses any distance)

• Elbow plot (use $k$ at elbow on plot of $k$ vs. within cluster sum of squares) – good for $k$-means (because based on Euclidean distance)

Silhouette width

Consider observation $i\in$ cluster $C_1$. Let

$$d(i,C_k)=\text{average dissimilarity of }i\text{ to all objects in cluster }{C}_k$$ $$a(i)=\text{average dissimilarity of }i\text{ to all objects in }{C}_1.$$ $$b(i)=\underset{C_k\ne C_1}{min}d(i,C_k)=\text{distance to the next closest neighbor cluster}$$ $$s(i)=\frac{b(i)-a(i)}{max\{a(i),b(i)\}}$$ $${\text{average}}_{i\in C_1}s(i)=\text{average silhouette width for cluster }{C}_1$$

Note that if $a(i)<b(i)$ then $i$ is well classified with a maximum $s(i)=1$. If $a(i)>b(i)$ then $i$ is not well classified with a minimum of $s(i)=-1$.

We are looking for a large silhouette width.