Incremental Clustering

 Previous clustering algorithms worked in

“batch” mode: processed all points at
essentially the same time.
 Some IR applications cluster an incoming
document stream (e.g., topic tracking).
 For these applications, we need incremental
clustering algorithms.
Incremental Clustering Issues
 How to be efficient? Should all
documents be cached?
 How to handle or support concept drift?
 How to reduce sensitivity to ordering?
 Goals:
 minimize the maximum cluster diameter
 minimize the number of clusters given a
fixed diameter
 Incremental Clustering Model
[Charikar et al. 1997]
 Extension to HAC as follows:
 Incremental Clustering: “for an update sequence
of n points in M, maintain a collection of k
clusters such that as each one is presented,
either it is assigned to one of the current k
clusters or it starts off a new cluster while two
existing clusters are merged into one.”
 Maintains a HAC for points added up until
current time.
M. Charikar, C. Chekuri, T. Feder, R. Motwani. “Incremental Clustering and Dynamic
Information Retrieval”, Proc. 29th Annual ACM Symposium on Theory of Computing,
1997.
 Doubling Algorithm (a=b=2)
1. Assign first k+1 points to k+1 clusters with each
point as centroid, d1=distance between closest
two points.
2. Do while more points
1. dt+1 = bdt
2. Merge clusters until all clusters in some new cluster:

1. Pick an arbitrary cluster; merge all clusters within dt+1 of centers

2. Remove selected clusters from old clusters
3. Calculate the centroid for the new cluster
3. Update clusters while number of clusters <=k:
1. Assign new point to closest cluster if within adt+1 of center;
otherwise create new cluster.
 Example:Plot -- Incremental
50
45 11
2
40
35 15
9
30 5 1
14 7 10
25
8 6
20 16
4 3 13
15
10 12
5
0
0 10 20 30 40 50
 Example:Doubling Merge
d2=24.08
50
45 11
2
40
35 15 X
9
30 5 1
14 7 10
25
8 6
20 16
4 3 13
15
10 12
5
0
0 10 20 30 40 50
 Example:Doubling Update
d2=24.08
50
45 11
2
40
35 15 X
9
30 5 1
14 7 10
25
8 6
20 16
4
15 X 3 13
10 12
5
0
0 10 20 30 40 50
 Example:Doubling Update
d2=24.08
50
45 11
2
40
35 15 X
9
30 5 1
14 7 10
25
8 6
20 X 16
4 3 13
15
10 12
5
0
0 10 20 30 40 50
 Example:Doubling Update
d2=24.08
50
45 11
2
40
35 15
5
9 X
30 1
14 7 10
25
8 6
20 X 16
4 3 13
15
10 12
5
0
0 10 20 30 40 50
 Example:Doubling Solution
50
45 11
2
40
35 15
9
30 5 1
14 7 10
25
8 6
20 16
4 3 13
15
10 12
5
0
0 10 20 30 40 50
Clique Partition Background
 A clique in G=(V,E) is a subset V’ of V
s.t. every two vertices in V’ are joined
by an edge in E.
 A clique partition for G is a partition of
V into disjoint subsets V1…Vk s.t. for
1<=I<=k, the subgraph induced by Vi
is a complete graph.
 Clique Partition Algorithm
1. Assign first k+1 points to k+1 clusters with each point as
centroid, d1=distance between closest two points.
2. Do while more points
1. dt+1 = 2dt

2. Merge clusters:

1. Compute minimum clique partition from dt+1 threshold graph

2. Merge clusters in each clique
3. In each new cluster, arbitrarily assign one of the existing centers
as the center for the new cluster
3. Update clusters while number of clusters <=k:
1. Assign new point to a cluster if within dt+1 of center of it or sub-
clusters; otherwise create new cluster.
 Example: CP: Merge d1=12.04
50
45 11
2
40
35 15
9
30 5 1
14 7 10
25
8 6
20 16
4 3 13
15
10 12
5
0
0 10 20 30 40 50
 Example: CP: Update
d2=24.08
50
45 11
2
40
35 15
9
30 5 1
14 7 10
25 23
8 6
20 16
4 3 13
15
10 12
5
0
0 10 20 30 40 50
 Web Document Clustering
Applications
 Organizing search engine retrieval results
 Meta-search engine that hierarchically clusters of
results: Vivisimo
 Meta-search engine that graphically displays
clusters of results: Kartoo
 Detecting redundancy (e.g., mirror sites or
moved or re-formatted documents)
 User interest profiles (aka filtering)
Vivisimo: Result Organization
Kartoo: Visual Clustering
 Detecting Mirrors/Subsumed
Web Documents
Resemblance assesses similarity between two
documents.
| S ( A)  S ( B) |
r ( A, B) 
| S ( A)  S ( B) |

Containment assesses how A is a subset of B.

| S ( A)  S ( B) |
c( A, B) 
| S ( A) |
A.Z. Broder, S.C. Glassman, M.S. Manasse, G. Zweig, “Syntactic Clustering
of the Web”, Proceedings of WWW6, 1997.
 Computing R and C
 S(D,w) (shingle) is the set of all unique
contiguous subsequences of length w in
document D.
 S(D) is S(D,w) for a fixed size w.
 To reduce the storage and computation, we
can sample the shingles for each doc:
 First s: MINs(W)
 Every mth: MODm(W)
 Estimating R & C from a
Portion of a Document
Keep a sketch of each document D, which consists of F(D)
and/or V(D) .
 : U  U is a random permutatio n
F ( A)  MIN s ( ( S ( A)))
V ( A)  MODM ( ( S ( A)))
| MIN s ( F ( A)  F ( B))  F ( A)  F ( B) |
r ( A, B) 
| MIN s ( F ( A)  F ( B)) |
V ( A)  V ( B)
r ( A, B) 
V ( A)  V ( B)
V ( A)  V ( B)
c( A, B) 
V ( A)
 Web Clustering with R & C
 w=10, m=25, s=50?, threshold=.5
 Pre-process documents

1. For each doc, calculate a sketch

2. Sort pairs of <shingle,docid>, removing lexically-

equivalent and shingle-equivalent docs
3. Compute list of doc pairs with # of shared
shingles, ignoring very common shingles
4. Generate clusters
1. if r(A,B) > threshold, then add link A<->B
2. Produce connected components using union-find
Web Clustering Results 1997
 30M web pages, 150 GBytes
 600M shingles
 3.6M clusters of 12.3M docs
 2.1M clusters of 5.3M identical docs
 Took 10.5 CPU days to compute
 Web Applications of
Resemblance Clusters
 Find URL similar to …
 relies on fixed threshold and requires URLs to
have been processed
 WWW Lost and Found
 requires keeping some historical sketch info
 Remove similar docs from search results