Introduction to Information Retrieval

Flat Clustering

1
Introduction to Information Retrieval

Clustering: Definition

 (Document) clustering is the process of grouping a set of

documents into clusters of similar documents.
 Documents within a cluster should be similar.
 Documents from different clusters should be dissimilar.
 Clustering is the most common form of unsupervised
learning.
 Unsupervised = there are no labeled or annotated data.

2
Introduction to Information Retrieval

Data set with clear cluster structure

Propose algorithm
for finding the
cluster structure
in this example

3
Introduction to Information Retrieval

Classification vs. Clustering

 Classification: supervised learning
 Clustering: unsupervised learning
 Classification: Classes are human-defined and part of the
input to the learning algorithm.
 Clustering: Clusters are inferred from the data without
human input.
 However, there are many ways of influencing the outcome of
clustering: number of clusters, similarity measure,
representation of documents, . . .

4
Introduction to Information Retrieval

The cluster hypothesis

Cluster hypothesis. Documents in the same cluster behave

similarly with respect to relevance to information needs. All
applications of clustering in IR are based (directly or indirectly) on
the cluster hypothesis.

“closely associated documents tend to be relevant to the same

requests”.

5
Introduction to Information Retrieval

Applications of clustering in IR
Application What is Benefit
clustered?
Search result clustering search more effective
results information
presentation to user
Scatter-Gather (subsets alternative user
of) collection interface: “search
without typing”
Collection clustering collection effective information
presentation for
exploratory browsing
Cluster-based retrieval collection higher efficiency:
faster search
6
Introduction to Information Retrieval

Search result clustering for better navigation

7
Introduction to Information Retrieval

Scatter-Gather

8
Introduction to Information Retrieval

Clustering for improving recall

 To improve search recall:
 Cluster docs in collection a priori
 When a query matches a doc d, also return other docs in the
cluster containing d
 Hope: if we do this: the query “car” will also return docs
containing “automobile”
 Because the clustering algorithm groups together docs
containing “car” with those containing “automobile”.
 Both types of documents contain words like “parts”, “dealer”,
“mercedes”, “road trip”.

9
Introduction to Information Retrieval

Data set with clear cluster structure

Propose algorithm
for finding the
cluster structure
in this example

10
Introduction to Information Retrieval

Desiderata for clustering

 General goal: put related docs in the same cluster, put
unrelated docs in different clusters.
 How do we formalize this?
 The number of clusters should be appropriate for the data set
we are clustering.
 Initially, we will assume the number of clusters K is given.
 Later: Semiautomatic methods for determining K
 Secondary goals in clustering
 Avoid very small and very large clusters
 Define clusters that are easy to explain to the user
 Many others . . .
11
Introduction to Information Retrieval

Flat vs. Hierarchical clustering

 Flat algorithms
 Usually start with a random (partial) partitioning of docs into
groups
 Refine iteratively
 Main algorithm: K-means
 Hierarchical algorithms
 Create a hierarchy
 Bottom-up, agglomerative
 Top-down, divisive

12
Introduction to Information Retrieval

Hard vs. Soft clustering

 Hard clustering: Each document belongs to exactly one
cluster.
 More common and easier to do
 Soft clustering: A document can belong to more than one
cluster.
 Makes more sense for applications like creating browsable
hierarchies
 You may want to put sneakers in two clusters:
 sports apparel
 shoes
 You can only do that with a soft clustering approach.
 We will do flat, hard and exhaustive clustering only in this
class.

13
Introduction to Information Retrieval

Flat algorithms
 Flat algorithms compute a partition of N documents into a
set of K clusters.
 Given: a set of documents and the number K
 Find: a partition into K clusters that optimizes the chosen
partitioning criterion
 Global optimization: exhaustively enumerate partitions,
pick optimal one
 Not tractable
 Effective heuristic method: K-means algorithm

14
Introduction to Information Retrieval

Outline
 Clustering: Introduction

 Clustering in IR

 K-means

 Evaluation

 How many clusters?

15
Introduction to Information Retrieval

What is a good clustering?

 Internal criteria
 Example of an internal criterion: RSS in K-means
 But an internal criterion often does not evaluate the actual
utility of a clustering in the application.
 Alternative: External criteria
 Evaluate with respect to a human-defined classification

16
Introduction to Information Retrieval

External criteria for clustering quality

 Based on a gold standard data set, e.g., the Reuters
collection we also used for the evaluation of classification
 Goal: Clustering should reproduce the classes in the gold
standard
 (But we only want to reproduce how documents are divided
into groups, not the class labels.)
 First measure for how well we were able to reproduce the
classes: purity

17
Introduction to Information Retrieval

External criterion: Purity

 Ω= {ω1, ω2, . . . , ωK} is the set of clusters and

C = {c1, c2, . . . , cJ} is the set of classes.
 For each cluster ωk : find class cj with most members nkj in ωk
 Sum all nkj and divide by total number of points

18
Introduction to Information Retrieval

Example for computing purity

To compute purity: 5 = maxj |ω1 ∩ cj | (class x, cluster 1);

4 = maxj |ω2 ∩ cj | (class o, cluster 2); and 3 = maxj |ω3 ∩ cj |
(class ⋄, cluster 3). Purity is (1/17) × (5 + 4 + 3) ≈ 0.71.

19
Introduction to Information Retrieval

Rand index
 Definition:

 Based on 2x2 contingency table of all pairs of documents:

 TP+FN+FP+TN is the total number of pairs.

 There are pairs for N documents.
 Example: = 136 in o/⋄/x example
 Each pair is either positive or negative (the clustering puts
the two documents in the same or in different clusters) . . .
 . . . and either “true” (correct) or “false” (incorrect): the
clustering decision is correct or incorrect.
20
Introduction to Information Retrieval

Rand Index: Example

As an example, we compute RI for the o/⋄/x example. We first
compute TP + FP. The three clusters contain 6, 6, and 5 points,
respectively, so the total number of “positives” or pairs of
documents that are in the same cluster is:

Of these, the x pairs in cluster 1, the o pairs in cluster 2, the ⋄

pairs in cluster 3, and the x pair in cluster 3 are true positives:

Thus, FP = 40 − 20 = 20. FN and TN are computed similarly.

21
Introduction to Information Retrieval

Two other external evaluation measures

 Two other measures

 Normalized mutual information (NMI)
 How much information does the clustering contain about the
classification?
 Singleton clusters (number of clusters = number of docs) have
maximum MI
 Therefore: normalize by entropy of clusters and classes
 F measure
 Like Rand, but “precision” and “recall” can be weighted

22
Introduction to Information Retrieval

Rand measure for the o/⋄/x example

(20 + 72)/(20 + 20 + 24 + 72) ≈ 0.68.

23
Introduction to Information Retrieval

Evaluation results for the o/⋄/x example

All four measures range from 0 (really bad clustering) to 1 (perfect

clustering).

24
Introduction to Information Retrieval

25
Introduction to Information Retrieval

Adjusted Rand Index (ARI)

26
Introduction to Information Retrieval

F-Score

27
Introduction to Information Retrieval

Dunn Measure (Internal Measure)

http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.90.9770&rep=rep1&type=pdf

28
Introduction to Information Retrieval

https://link.springer.com/content/pdf/10.1007%2Fs00357-006-0017-z.pdf
29
Introduction to Information Retrieval

30
Introduction to Information Retrieval

Outline
❶ Recap

❷ Clustering: Introduction

❸ Clustering in IR

❹ K-means

❺ Evaluation

❻ How many clusters?

31
Introduction to Information Retrieval

K-means

 Perhaps the best known clustering algorithm

 Simple, works well in many cases
 Use as default / baseline for clustering documents

32
Introduction to Information Retrieval

Document representations in clustering

 Vector space model

 As in vector space classification, we measure relatedness
between vectors by Euclidean distance . . .
 . . .which is almost equivalent to cosine similarity.
 Almost: centroids are not length-normalized.

33
Introduction to Information Retrieval

K-means
 Each cluster in K-means is defined by a centroid.
 Objective/partitioning criterion: minimize the average
squared difference from the centroid
 Recall definition of centroid:

where we use ω to denote a cluster.

 We try to find the minimum average squared difference by
iterating two steps:
 reassignment: assign each vector to its closest centroid
 recomputation: recompute each centroid as the average of
the vectors that were assigned to it in reassignment
34
Introduction to Information Retrieval

K-means algorithm

35
Introduction to Information Retrieval

Worked Example: Set of to be clustered

36
Introduction to Information Retrieval

Worked Example: Random selection of initial

centroids

Exercise: (i) Guess what the

optimal clustering into two clusters is in this case; (ii) compute the
centroids of the clusters
37
Introduction to Information Retrieval

Worked Example: Assign points to closest center

38
Introduction to Information Retrieval

Worked Example: Assignment

39
Introduction to Information Retrieval

Worked Example: Recompute cluster centroids

40
Introduction to Information Retrieval

Worked Example: Assign points to closest centroid

41
Introduction to Information Retrieval

Worked Example: Assignment

42
Introduction to Information Retrieval

Worked Example: Recompute cluster centroids

43
Introduction to Information Retrieval

Worked Example: Assign points to closest centroid

44
Introduction to Information Retrieval

Worked Example: Assignment

45
Introduction to Information Retrieval

Worked Example: Recompute cluster centroids

46
Introduction to Information Retrieval

Worked Example: Assign points to closest centroid

47
Introduction to Information Retrieval

Worked Example: Assignment

48
Introduction to Information Retrieval

Worked Example: Recompute cluster centroids

49
Introduction to Information Retrieval

Worked Example: Assign points to closest centroid

50
Introduction to Information Retrieval

Worked Example: Assignment

51
Introduction to Information Retrieval

Worked Example: Recompute cluster centroids

52
Introduction to Information Retrieval

Worked Example: Assign points to closest centroid

53
Introduction to Information Retrieval

Worked Example: Assignment

54
Introduction to Information Retrieval

Worked Example: Recompute cluster centroids

55
Introduction to Information Retrieval

Worked Example: Assign points to closest centroid

56
Introduction to Information Retrieval

Worked Example: Assignment

57
Introduction to Information Retrieval

Worked Example: Recompute cluster caentroids

58
Introduction to Information Retrieval

Worked Ex.: Centroids and assignments after

convergence

59
Introduction to Information Retrieval

K-means is guaranteed to converge: Proof

 RSS = sum of all squared distances between document vector
and closest centroid
 RSS decreases during each reassignment step.
 because each vector is moved to a closer centroid
 RSS decreases during each recomputation step.
 see next slide
 There is only a finite number of clusterings.
 Thus: We must reach a fixed point.
 Assumption: Ties are broken consistently.

60
Introduction to Information Retrieval

Recomputation decreases average distance

– the residual sum of squares (the “goodness”
measure)

The last line is the componentwise definition of the centroid! We

minimize RSSk when the old centroid is replaced with the new
centroid. RSS, the sum of the RSSk , must then also decrease
during recomputation.
61
Introduction to Information Retrieval

K-means is guaranteed to converge

 But we don’t know how long convergence will take!

 If we don’t care about a few docs switching back and forth,
then convergence is usually fast (< 10-20 iterations).
 However, complete convergence can take many more
iterations.

62
Introduction to Information Retrieval

Optimality of K-means

 Convergence does not mean that we converge to the optimal

clustering!
 This is the great weakness of K-means.
 If we start with a bad set of seeds, the resulting clustering
can be horrible.

63
Introduction to Information Retrieval

Convergence Exercise: Suboptimal clustering

 What is the optimal clustering for K = 2?

 Do we converge on this clustering for arbitrary seeds di , dj?

64
Introduction to Information Retrieval

Initialization of K-means
 Random seed selection is just one of many ways K-means
can be initialized.
 Random seed selection is not very robust: It’s easy to get a
suboptimal clustering.
 Better ways of computing initial centroids:
 Select seeds not randomly, but using some heuristic (e.g., filter
out outliers or find a set of seeds that has “good coverage” of
the document space)
 Use hierarchical clustering to find good seeds
 Select i (e.g., i = 10) different random sets of seeds, do a K-
means clustering for each, select the clustering with lowest RSS

65
Introduction to Information Retrieval

Time complexity of K-means

 Computing one distance of two vectors is O(M).
 Reassignment step: O(KNM) (we need to compute KN
document-centroid distances)
 Recomputation step: O(NM) (we need to add each of the
document’s < M values to one of the centroids)
 Assume number of iterations bounded by I
 Overall complexity: O(IKNM) – linear in all important
dimensions
 However: This is not a real worst-case analysis.
 In pathological cases, complexity can be worse than linear.
 K-medoids
66
Introduction to Information Retrieval

How many clusters?

 Number of clusters K is given in many applications.
 E.g., there may be an external constraint on K. Example: In the
case of Scatter-Gather, it was hard to show more than 10–20
clusters on a monitor in the 90s.
 What if there is no external constraint? Is there a “right”
number of clusters?
 One way to go: define an optimization criterion
 Given docs, find K for which the optimum is reached.
 What optimiation criterion can we use?
 We can’t use RSS or average squared distance from centroid as
criterion: always chooses K = N clusters.

67
Introduction to Information Retrieval

Simple objective function for K (1)

 Basic idea:
 Start with 1 cluster (K = 1)
 Keep adding clusters (= keep increasing K)
 Add a penalty for each new cluster
 Trade off cluster penalties against average squared distance
from centroid
 Choose the value of K with the best tradeoff

68
Introduction to Information Retrieval

Simple objective function for K (2)

 Given a clustering, define the cost for a document as
(squared) distance to centroid
 Define total distortion RSS(K) as sum of all individual
document costs (corresponds to average distance)
 Then: penalize each cluster with a cost λ
 Thus for a clustering with K clusters, total cluster penalty is
Kλ
 Define the total cost of a clustering as distortion plus total
cluster penalty: RSS(K) + Kλ
 Select K that minimizes (RSS(K) + Kλ)
 Still need to determine good value for λ . . .
69
Introduction to Information Retrieval

Finding the “knee” in the curve

Pick the number of

clusters where curve
“flattens”. Here: 4 or 9.

70
Introduction to Information Retrieval

Silhouette method
• a(i) is the average distance between i and all other
data within the same cluster.
• b(i) is the smallest average distance of i to all points
in any other cluster, of which i is not a member.

Score is 0 for clusters

with size = 1

71
Introduction to Information Retrieval

k-medoids
1. Initialize: select k of the n data points as the medoids

2. Associate each data point to the closest medoid.

3. While the cost of the configuration decreases:

3.1 For each medoid m, for each non-medoid data point o:
3.1.1. Swap m and o, associate each data point to the
closest medoid, recompute the cost (sum of
distances of points to their medoid)
3.1.2. If the total cost of the configuration increased in the
previous step, undo the swap

72
Introduction to Information Retrieval

EM Algorithm

73