Discerning Linkage-Based Algorithms Among Hierarchical Clustering Methods Margareta Ackerman and Shai Ben-David IJCAI 2011.

Discerning Linkage-Based Algorithms Among Hierarchical Clustering Methods

Margareta Ackermanand

Shai Ben-David

IJCAI 2011

Clustering is one of the most widely used tools for exploratory data analysis.

Social Sciences Biology Astronomy Computer Science ….

All apply clustering to gain a first understanding of the structure of large data sets.

The Theory-Practice Gap

Both statements still apply today.

“While the interest in and application of cluster analysis has been rising rapidly, the abstract nature of the tool is

still poorly understood” (Wright, 1973)

“There has been relatively little work aimed at reasoning about clustering independently of any

particular algorithm, objective function, or generative data model” (Kleinberg, 2002)

Bridging the Theory-Practice Gap:Previous work

• Axioms of clustering [(Kleinberg, NIPS 02), (Ackerman & Ben-David, NIPS 08), (Meila, NIPS 08)]

• Clusterability [(Balcan, Blum, and Vempala, STOC 08), (Ackerman & Ben-David, AISTATS 09) ]

There are a wide variety of clustering algorithms, which often produce very different clusterings.

How should a user decide which algorithm to use for a given application?

M. Ackerman, S. Ben-David, and D. Loker

Bridging the Theory-Practice Gap:Clustering algorithm selection

We propose a framework that lets a user utilize prior knowledge to select an algorithm

• Identify properties that distinguish between the input-output behaviour of different clustering algorithms

• The properties should be:1) Intuitive and “user-friendly”2) Useful for classifying clustering

algorithms

Our approach for clustering algorithm selection

• A property-based classification of partitional clustering algorithms (Ackerman, Ben-David, and Loker, NIPS ‘10)

• A characterizing of a single-linkage with the k-stopping criteria (Zadeh and Ben-David, UAI 09)

• A characterization of linkage-based clustering with the k-stopping criteria (Ackerman, Ben-David, and Loker, COLT ‘10)

Previous Work in Property-Based Framework

• Extend the above property-based framework to the hierarchical clustering setting

• Propose two intuitive properties that uniquely indentify hierarchical linkage-based clustering algorithms

• Show that common hierarchical algorithms, including bisecting k-means, cannot be simulated by any linkage-based algorithm

Our contributions

Outline

• Define Linkage-Based clustering• Introduce two new properties of

hierarchical clustering algorithms• Main result• Hierarchical clustering paradigms that are

not linkage-based• Conclusions

A set C_i is a cluster in a dendrogram D if there exists a node in the dendrogram so that C_i is the set of its leaf descendents.

Formal Setup: Dendrograms and clusterings

Dendrogram:

C = {C1, … , Ck} is a clustering in a dendrogram D if

– Ci is a cluster in D for all 1≤ i ≤ k, and

– clusters are disjoint, Ci∩Cj = Ø for all 1≤ i<j ≤k.

Formal Setup: Dendrograms and clusterings

Formal Setup: Hierarchical clustering algorithm

A Hierarchical Clustering Algorithm A mapsInput: A data set X with a distance function d,

denoted (X,d) toOutput: A dendrogram of X

An algorithm A is Linkage-Based if there exists alinkage-function l:{(X1, X2 ,d): d over X1uX2 }→ R+ such that for any (X,d), A(X,d) can be constructed asfollows: • Create a single-node tree for every elements of X

Linkage-Based Algorithm

An algorithm A is Linkage-Based if there exists alinkage-function l:{(X1, X2 ,d): d over X1uX2 }→ R+ such that for any (X,d), A(X,d) can be constructed asfollows: • Create a single-node tree for every elements of X• Repeat the following until a single tree remains:

Merge the pair of trees whose element sets are closest according to l.

Linkage-Based Algorithm

Ex. Single-linkage, average-linkage,complete linkage

Outline

• Define Linkage-Based clustering• Introduce two new properties of

hierarchical clustering algorithms• Main result• Hierarchical clustering paradigms that are

not linkage-based• Conclusions

Locality Informal Definition

If we select a set of disjoint clusters from a dendrogram, and run the algorithm on the union of these clusters, we obtain a result that is consistent with the original dendrogram.

D = A(X,d) D’ = A(X’,d)X’={x1, …, x6}

Outer ConsistencyA(X,d)

C

• The outer-consistent change makes the clustering C more prominent.

• If A is outer-consistent, then A(X,d’) will also include the clustering C.

C on dataset (X,d)C on dataset (X,d’)

Increase pairwise between-cluster

distances

Outline

• Define Linkage-Based clustering• Introduce two new properties of

hierarchical clustering algorithms• Main result• Hierarchical clustering paradigms that are

not linkage-based• Conclusions

Theorem: A hierarchical clustering function is

Linkage-Based if and only if

it is Local and Outer-Consistent.

Our Main Result

Recall direction: If A satisfies Outer-Consistency and Locality, then A is Linkage-Based.

Goal: Define a linkage function l so that the linkage-based clustering based on l outputs A(X,d) (for every X and d).

Brief Sketch of Proof

• Define an operator <A :

(X,Y,d1) <A (Z,W,d2) if when we run A on (XuYuZuW,d), where d extends d1 and d2, X and Y are merged before Z and W.

Brief Sketch of Proof

A(X,d)

Z W X Y

• Prove that <A can be extended to a partial ordering by proving that it is cycle-free

• This implies that there exists an order preserving function l that maps pairs of data sets to R+.

Outline

• Define Linkage-Based clustering• Introduce two new properties of

hierarchical clustering• Main result• Hierarchical clustering paradigms that are

not linkage-based• Conclusions

Hierarchical but Not Linkage-Based

• P -Divisive algorithms construct dendrograms top-down using a partitional 2-clustering algorithm P to determine how to split nodes.

• Many natural partitional 2-clustering algorithms satisfy the following property:

A partitional 2-clustering algorithm P is Context Sensitive if there exist d d’⊂ so that

P({x,y,z),d) = {x, {y,z}} and P({x,y,z,w} ,d’)= {{x,y}, {z,w}}.

Ex. K-means, min-sum, min-diameter, and further-centroids.

Hierarchical but Not Linkage-Based

• The input-output behaviour of some natural divisive algorithms is distinct from that of all linkage-based algorithms.

• The bisecting k-means algorithm, and other natural divisive algorithms, cannot be simulated by any linkage-based algorithm.

Theorem: If P is context-sensitive, then the P –divisive algorithm fails the locality property.

Outline

• Define Linkage-Based clustering• Introduce two new properties of

hierarchical clustering algorithms• Main result• Hierarchical clustering paradigms that are

not linkage-based• Conclusions

Conclusions

• We characterize hierarchical Linkage-Based clustering in terms of two intuitive properties.

• Show that some natural hierarchical algorithms have different input-output behavior than any linkage-based algorithm.

Locality

For any clustering C = {C1, … , Ck} in D = A(X,d),

• C is also a clustering in D’ = A(X’ = u Ci , d)

• C i roots the same sub-dendrogram in both D and D’

• For all x,y in X’, x occurs below y in D iff the same holds in D’.

D’ = A(X’,d)X’={x1, …, x6}

D = A(X,d)