Clustering

Everything Data
CompSci 216 Spring 2018
 2

Announcements (Thu. Mar 2)

•  Homework #7 will be posted today.

•  Projects teams and number assignments

are posted. Please let me know of
changes.
–  Once you submit your proposal on Tuesday
no more changes will be entertained.
 3

Announcements (Thu. Mar 2)

•  Project presentations on Tuesday to
instructors
–  3 minutes per team
–  Introduce your team members
–  Describe problem, dataset and how you will
quantify success
–  You may use 1-2 slides (PDF format)
 4

Geo-tags of tweets
80

60

40

20

-20

-40

-60
-200 -150 -100 -50 0 50 100 150 200
 5

Trending topics
•  How would you compute trending topics?
–  Most frequent hashtags
–  Frequent keywords or phrases (which are not
stopwords)
–  …

•  But interesting trends in one region may

not represent interesting trends in another.
 6

Idea: Cluster tweets by geography

80 12
"geo_data_head10000_kmeans_10" using 1:2:3

60
10

40
8

20
6
0

4
-20

2
-40

-60 0
-200 -150 -100 -50 0 50 100 150 200
 7

Trending topics by geography

•  We can now compute trending topics
within each cluster (region).
 8

Example: Market Segmentation

http://www.esriro.ro/library/fliers/
pdfs/tapestry_segmentation.pdf#page=2
 9

Example: Phylogenetic Trees

 10

Other Examples
•  Image segmentation
•  Document clustering
•  De-duplication …
 11

Outline
•  K-means Clustering

•  Distance Metrics

•  Using distance metrics for clustering

–  K-medoids
–  Hierarchical Clustering
 12

How did we create 10 clusters?

80 12
"geo_data_head10000_kmeans_10" using 1:2:3

60
10

40
8

20
6
0

4
-20

2
-40

-60 0
-200 -150 -100 -50 0 50 100 150 200
 13

Can compare apples vs oranges …

•  … if they are in the same feature space.

•  X = {x1, x2, …, xn} is a dataset

•  Each xi is assumed to be a point in some

d-dimensional space
–  xi = [xi1 , xi2 , …, xid ]
–  Each dimension represents a feature
 14

K-means
•  Partition a set of points X = {x1, x2, …, xn}
into k partitions C = {C1, C2, …, Ck} that
minimizes
! !
!!
!""(!) = !!" !! − !! !
!!
!!! !!!

aij is 1 if
xj is assigned to cluster Ci

Assignment Function
 15

K-means
•  Partition a set of points X = {x1, x2, …, xn}
into k partitions C = {C1, C2, …, Ck} that
minimizes
! !
!!
!""(!) = !!" !! − !! !
!!
!!! !!!

!! = !!! , !!! , … , !!! !

!! µi is the mean of
!!" = ! ! points in cluster Ci.
!∈!! |!! |
Cluster Representative
 16

K-means
•  Partition a set of points X = {x1, x2, …, xn}
into k partitions C = {C1, C2, …, Ck} that
minimizes
! !
!!
!""(!) = !!" !! − !! !
!!
!!! !!!

Square of the straight line distance

between xj and its center µi.
 17

Chicken-and-Egg problem
•  How do we minimize RSS(C) ?

–  If we know the cluster representatives (or the

means), then it is easy to find the assignment
function (which minimizes RSS(C))
•  Assign point to the closest cluster representative
–  If we know the assignment function,
computing the cluster representatives is easy
•  Compute the mean of the points in the cluster
 18

K-means Algorithm
•  Idea: Alternate these two steps.
–  Pick some initialization for cluster
representatives µ0.

–  E-step:
Assign points to the closest representative in µi.
–  M-step:
Recompute the representatives µi+1 as means of
the current clusters.
 19

K-means: Initialization
80 12
"geo_data_head10000_run_0" using 1:2:3

60
+ +
10

40 + + + +
+ 8

20 + + 6
0

+ 4
-20

2
-40

-60 0
-200 -150 -100 -50 0 50 100 150 200
 20

K-means: Iteration 1
80 12
"geo_data_head10000_run_1" using 1:2:3

60
10
+ +
40
+ + +
+ 8

20 + +
+
6
0

-20 + 4

2
-40

-60 0
-200 -150 -100 -50 0 50 100 150 200
 21

K-means: Iteration 2
80 12
"geo_data_head10000_run_2" using 1:2:3

60
10
+ +
40 + + +
+ 8

20 +
6
0 + +
4
-20 +
2
-40

-60 0
-200 -150 -100 -50 0 50 100 150 200
 22

K-means: Iteration 10
80 12
"geo_data_head10000_kmeans_10" using 1:2:3

60
+ +
10

40
+ + +
+ 8

20

+ +
6
0
+
4
-20
+
2
-40

-60 0
-200 -150 -100 -50 0 50 100 150 200
 23

Initialization
•  Many heuristics
–  Random: K random points in the dataset

–  Farthest First:
•  Pick the first center at random
•  Pick the ith center as the point “farthest away”
from the last (i-1) centers

–  K-means++: (see paper)

•  Nice theoretical guarantees on quality of clustering
 24

Stopping
•  Alternate E and M steps till the cluster
representatives do not change.

•  Guaranteed to converge
–  To a local optima …
–  … but not necessarily to a global optima

•  Finding the optimal solution (with least RSS(C)) is NP-

hard, even for 2 clusters.
 25

Where k-means fails …

y y

x
 26

Scaling / changing features can help

y R

0.5 * x
 27

Limitations of k-means
•  Scaling/changing the feature space can
change the solution.
•  Cluster points into spherical regions.

•  Number of clusters should be known

apriori
 28

Outline
•  K-means Clustering

•  Distance Metrics

•  Using distance metrics for clustering

–  K-medoids
–  Hierarchical Clustering

 29

Distance Metrics
•  Function d that maps pairs of points x, y to
real numbers (usually between 0, 1)

•  Symmetric: d(x,y) = d(y,x)

•  Triangle Inequality: d(x,y) + d(y,z) ≥ d(x,z)

•  Choice of distance metric is usually

application dependent
 30

Euclidean Distance
!−! ! = !! − !! ! !
!

!

•  Straight line distance between two points
x = [x1, x2, …, xd] and y = [y1, y2, …, yd]

•  K-means minimizes the sum of the Euclidean

distances between the points and the centers
–  We use the mean as a center
 31

Minkowski (Lp) Distance

!
!
!
!! = ! !! − !! ! !

!

•  L2 = ?
 32

Minkowski (Lp) Distance

!
!
!
!! = ! !! − !! ! !

!

•  L2 = Euclidean
•  L1 = ?
 33

Minkowski (Lp) Distance

!
!
!
!! = ! !! − !! ! !

!

•  L1 = city block / Manhalan

•  L∞ = ?
 34

Vector-based Similarities
•  Cosine Similarity (inverse of a distance)
Dot Product

! !! ∙ !!
!

! !
! !! ! !!

L2 Norm

–  can be used in conjunction with TFIDF scores

 35

Vector-based Similarities
•  Pearson’s Correlation Coefficient
–  cosine similarity on mean normalized vectors

Mean of xi’s Dot Product

! !! − ! ∙ (!! − !)
!

! !
! !! − ! ! !! − !

L2 Norm
 36

Set-based Distances
•  Let A and B be two sets.

|! ∩ !|
Jaccard !, ! = ! !

|! ∪ !|

–  Again, a measure of similarity (inverse of distance)

 37

Scaling / Changing features …

•  … can be thought of as using a different
distance function.

•  How do we cluster for general distance

functions?
 38

Outline
•  K-means Clustering

•  Distance Metrics

•  Using distance metrics for clustering

–  K-medoids
–  Hierarchical Clustering
 39

K-means for general distance functions?

•  Mean of a set of points does not always make

sense.
–  Mean of a set of movies or a set of documents?

•  Mean m of a set of points P minimizes the sum

of Euclidean distances between m and every
point in P
–  Best cluster representative under Euclidean Distance

–  The above is not true for a general distance metric.

 40

K-medoids
•  Allows a general distance metric d(x,y).

•  Same algorithm as K-means …

•  … but we don’t pick the new centers using
mean of the cluster.
 41

K-medoids
–  Pick some initialization for cluster
representatives µ0.
–  E-step:
Assign points to the closest representative in µi.
–  M-step:
Recompute the representatives µi+1 as the medoid,
or one of the points in the cluster with the
minimum distance from all the other points.
 42

Medoid
•  m is the medoid of a set of points P if

! = ! argmin !(!, !) !
!!∈!! !∈!

Point that minimizes the sum of distances to
all other points in the set.
 43

Computing the medoid

! = ! argmin !(!, !) !
!!∈!! !∈!

•  Need to compute all |P|2 distances.

•  In comparison, computing the mean in k-

means only requires computing d averages
involving |P| numbers each.
 44

K-medoids summary
•  Same algorithm as K-means, but uses medoids
instead of means

•  Centers are always points that appear in the

original dataset

•  Can use any distance measure for clustering.

•  Still need to know the number of clusters a priori…

 45

Outline
•  K-means Clustering

•  Distance Metrics

•  Using distance metrics for clustering

–  K-mediods
–  Hierarchical Clustering
 46

Hierarchical Clustering
•  Rather than compute a single clustering,
compute a family of clusterings.

•  Can choose the clusters a posteriori.

 47

Agglomerative Clustering
•  Initialize each point to its own cluster

•  Repeat:
–  Pick the two clusters that are closest
–  Merge them into one cluster
–  Stop when there is only one cluster left
 48

Example
Step 1: {1} {2} {3} {4} {5} {6} {7}
Step 2: {1} {2, 3} {4} {5} {6} {7}
Step 3: {1, 7} {2, 3} {4} {5} {6}
Step 4: {1, 7} {2, 3} {4, 5} {6}
Step 5: {1, 7} {2, 3, 6} {4, 5}
Step 6: {1, 7} {2, 3, 4, 5, 6}
Step 7: {1, 2, 3, 4, 5, 6, 7}

Example based on Ryan Tibshirani’s slides
 49

Dendrogram Entire dataset

Height of a node is
proportional to distance
between children clusters Each node
is a
cluster

Individual points in
the dataset
 50

Dendrogram

A horizontal cut in the

dendrogram results in a
clustering
 51

Distance between clusters

Step 1: {1} {2} {3} {4} {5} {6} {7}
Step 2: {1} {2, 3} {4} {5} {6} {7}
Step 3: {1, 7} {2, 3} {4} {5} {6}
Step 4: {1, 7} {2, 3} {4, 5} {6}

What are the next two closest
clusters?
 52

Single Linkage
!!"#$%& (!! , !! ) = ! min !(!, !)!
!∈!! ,!∈!!

Distance between two clusters is

the distance between the two
closest points in the clusters.

{6} is closer to {4,5} than {2,3}
according to single linkage
 53

Complete Linkage
!!"#$%&'& (!! , !! ) = ! max !(!, !)!
!∈!! ,!∈!!

Distance between two clusters is

the distance between the two
farthest points in the clusters.

{6} is closer to {2,3} than {4,5}
according to complete linkage
 54

Single vs Complete Linkage

4

3 1

2
> 1
1

0
0 1 2 3 4 5 6 7
 55

Single Linkage
4

3 1

2
> 1
1

0
0 1 2 3 4 5 6 7
 56

Single Linkage
4

0
0 1 2 3 4 5 6 7

Chaining: Single linkage can result in clusters

that are spread out and not compact
 57

Complete Linkage
4

3 3

2
< 2
1

0
0 1 2 3 4 5 6 7
 58

Complete Linkage
4

0
0 1 2 3 4 5 6 7

Complete linkage returns more compact clusters

in this case.
 59

Single vs Complete Linkage

3

1.02 5.02 6.99

1
4 6

0
0 1 2 3 4 5 6 7 8
 60

In both cases …
3

1.02 5.02 6.99

1
4 6

0
0 1 2 3 4 5 6 7 8
 61

Single Linkage
3

1.02 5.02 6.99

1
4 6

0
0 1 2 3 4 5 6 7 8
 62

Complete Linkage
3

1.02 5.02 6.99

1
4 6

0
0 1 2 3 4 5 6 7 8

Complete Linkage is sensitive to outliers.

 63

Average Linkage
!!!! ,!!!! !(!, !)
!!"# (!! , !! ) = ! ! !!
|!! | ∙ |!! |

Distance between two clusters is
the average distance between
every pair of points in the
clusters.
 64

Hierarchical Clustering summary

•  Create a family of hierarchical clusterings
–  Visualized using a dendrograms
–  Users can choose number of clusters after
clustering is done.

•  Can use any distance function

•  Different choices for measuring distance

between clusters.