L2 and L1 Criteria for K-Means Bilinear Clustering B. Mirkin School of Computer Science Birkbeck...

L2 and L1 Criteria for K-Means Bilinear Clustering

B. MirkinSchool of Computer ScienceBirkbeck College, University of London

Advert of a Special Issue: The Computer Journal, Profiling Expertise and Behaviour: Deadline 15 Nov. 2006. To submit, http:// www.dcs.bbk.ac.uk/~mark/cfp_cj_profiling.txt

Outline: More of Properties than MethodsClustering, K-Means and IssuesData recovery PCA model and clusteringData scatter decompositions for L2 and L1 Contributions of nominal featuresExplications of Quadratic criterionOne-by-one cluster extraction: Anomalous patterns and iK-MeansIssue of the number of clustersComments on optimisation problemsConclusion and future work

WHAT IS CLUSTERING; WHAT IS DATAK-MEANS CLUSTERING: Conventional K-Means; Initialization of K-Means; Intelligent K-Means; Mixed Data; Interpretation AidsWARD HIERARCHICAL CLUSTERING: Agglomeration; Divisive Clustering with Ward Criterion; Extensions of Ward ClusteringDATA RECOVERY MODELS: Statistics Modelling as Data Recovery;

Data Recovery Model for K-Means; for Ward; Extensions to Other Data Types; One-by-One ClusteringDIFFERENT CLUSTERING APPROACHES: Extensions of K-Means; Graph-Theoretic Approaches; Conceptual Description of ClustersGENERAL ISSUES: Feature Selection and Extraction; Similarity on Subsets and Partitions; Validity and Reliability

Clustering, K-Means and IssuesBilinear PCA model and clusteringData Scatter Decompositions: Quadratic and Absolute Contributions of nominal featuresExplications of Quadratic criterionOne-by-one cluster extraction: Anomalous patterns and iK-MeansIssue of the number of clustersComments on optimisation problemsConclusion and future work

Example: W. Jevons (1857) planet clusters, updated (Mirkin, 1996)

Pluto doesn’t fit in the two clusters of planets: originated another cluster (2006)

Clustering algorithms

Nearest neighbour Ward’s Conceptual clustering

K-means Kohonen SOM Spectral clustering ………………….

K-Means: a generic clustering method

Entities are presented as multidimensional points (*) 0. Put K hypothetical centroids (seeds) 1. Assign points to the centroids according to minimum distance rule 2. Put centroids in gravity centres of thus obtained clusters 3. Iterate 1. and 2. until convergence

K= 3 hypothetical centroids (@)

* *

* * * * * * * * @ @

@** * * *

K-Means: a generic clustering method

Entities are presented as multidimensional points (*) 0. Put K hypothetical centroids

(seeds) 1. Assign points to the centroids according to Minimum distance rule 2. Put centroids in gravity centres of thus obtained clusters 3. Iterate 1. and 2. until convergence

* *

* * * * * * * * @ @

@** * * *

K-Means: a generic clustering method

Entities are presented as multidimensional points (*) 0. Put K hypothetical centroids (seeds) 1. Assign points to the centroids according to Minimum distance rule 2. Put centroids in gravity centres of thus obtained clusters 3. Iterate 1. and 2. until convergence

* *

* * * * * * * * @ @

@** * * *

K-Means: a generic clustering method

Entities are presented as multidimensional points (*) 0. Put K hypothetical centroids (seeds) 1. Assign points to the centroids according to Minimum distance rule 2. Put centroids in gravity centres of thus obtained clusters 3. Iterate 1. and 2. until convergence 4. Output final centroids and clusters

* * @ * * * @ * * * *

** * * *

@

Advantages of K-Means Models typology building Computationally effective Can be utilised incrementally, `on-line’

Shortcomings (?) of K-Means Initialisation affects results Convex cluster shape

Initial Centroids: Correct

Two cluster case

Initial Centroids: Correct

Initial Final

Different Initial Centroids

Different Initial Centroids: Wrong

Initial Final

Issues:K-Means gives no advice on:

* Number of clusters* Initial setting* Data normalisation* Mixed variable scales* Multiple data sets

K-Means gives limited advice on: * Interpretation of

resultsThese all can be addressed with the data recovery approach

Clustering, K-Means and IssuesData recovery PCA model and clusteringData Scatter Decompositions: Quadratic and Absolute Contributions of nominal featuresExplications of Quadratic criterionOne-by-one cluster extraction: Anomalous patterns and iK-MeansIssue of the number of clustersComments on optimisation problemsConclusion and future work

Data recovery for data mining (discovery of patterns in data)

Type of Data Similarity Temporal Entity-to-feature

Type of Model Regression Principal

components Clusters

Model:Data = Model_Derived_Data + ResidualPythagoras:

|Data|m = |Model_Derived_Data|m + |Residual|m

m =1, 2. The better fit, the better the model: a natural source of

optimisation problems

K-Means as a data recovery method

Representing a partition

Cluster k:

Centroid

ckv (v - feature)

Binary 1/0 membership

zik (i - entity)

Basic equations (same as for PCA, but score vectors zk constrained to be binary)

y – data entry, z – 1/0 membership, not score

c - cluster centroid, N – cardinality

i - entity, v - feature /category, k - cluster

,1

ivikz

kvc

K

kivy

Clustering: general and K-MeansBilinear PCA model and clusteringData Scatter Decompositions L2 and L1Contributions of nominal featuresExplications of Quadratic criterionOne-by-one cluster extraction: Anomalous patterns and iK-MeansIssue of the number of clustersComments on optimisation problemsConclusion and future work

Quadratic data scatter decomposition (classic)

K-means: Alternating LS minimisation y – data entry, z – 1/0 membership

c - cluster centroid, N – cardinality

i - entity, v - feature /category, k - cluster

K

k Si

V

vkviv

V

vk

K

kkv

N

i

V

viv

k

cyNcy1 1

2

1 1

2

1 1

2 )(

,1

ivikz

kvc

K

kivy

Absolute Data Scatter Decomposition (Mirkin 1997)

K

k Si

V

vkviv

V

vkvkv

K

kiv

Si

N

i

V

viv

k

kv

cy

cnyy

1 1

1 11 1

||

|)|||2(||

00:

00:

kvivkvk

kvkvivkkv cycSi

ccySiS

Ckv are medians

OutlineClustering: general and K-MeansBilinear PCA model and clusteringData Scatter Decompositions L2 and L1 Implications for data pre-processingExplications of Quadratic criterionOne-by-one cluster extraction: Anomalous patterns and iK-MeansIssue of the number of clustersComments on optimisation problemsConclusion and future work

Meaning of the Data scatter

m=1,2; The sum of contributions of features – the basis for feature pre-processing (dividing by range rather than std) Proportional to the summary variance (L2) / absolute deviation from the median (L1)

V

v

N

i

miv

N

i

V

v

miv

m yyD1 11 1

||||

Standardisation of features

Yik = (Xik –Ak)/Bk

X - original data Y – standardised data i – entities k – features Ak – shift of the origin, typically, the average Bk – rescaling factor, traditionally the

standard deviation, but range may be better in clustering

Normalising by std

decreases the effect of more useful feature 2

by rangekeeps the effect of distribution shape in T(Y):B = range*#categories (for L2 case)(under the equality-of-variables assumption)

B1=Std1 << B2= Std2

Data standardisation

Categories as one/zero variablesSubtracting the averageAll features: Normalising by rangeCategories - sometimes by the number of them

Illustration of data pre-processing Mixed scale data table

Conventional quantitative coding + … data standardisation

No normalisation

Tom Sawyer

Z-scoring (scaling by std)

Tom Sawyer

Normalising by range*#categories

Tom Sawyer

OutlineClustering: general and K-MeansBilinear PCA model and clusteringData Scatter Decompositions: Quadratic and Absolute Contributions of nominal featuresExplications of Quadratic criterionOne-by-one cluster extraction: Anomalous patterns and iK-MeansIssue of the number of clustersComments on optimisation problemsConclusion and future work

Contribution of a feature F to a partition (m=2)

Proportional to correlation ratio 2 if F is quantitative a contingency coefficient between cluster

partition and F, if F is nominal:Pearson chi-square (Poisson normalised)

Goodman-Kruskal tau-b (Range normalised)

Fv

k

K

kkvNc

1

2Contrib(F) =

Contribution of a quantitative feature to a partition (m=2)

Proportional to correlation ratio 2 if F is quantitative

2

1

222 /)(

K

kkkpNN

Contribution of a pair nominal feature – partition, L2 case

Proportional to a contingency coefficient Pearson chi-square (Poisson normalised)

Goodman-Kruskal tau-b (Range normalised) Bj=1

Still needs be normalised by the square root of #categories, to balance the contribution of a numerical feature

2/

2)

,( jBkpjk jpkpkjpNContr

jj pB

Contribution of a pair nominal feature – partition, L1 case

A highly original contingency coefficient Still needs be normalised by square root of #categories, to balance the contribution of a numerical feature

|2|

),(

),(

/2[/

),(

)1/(2 kpkvp

vk

vkvp

kp

kvpKNContr

vkvkvk ppp

||/|2|),(kvckpkvpNvkContr

Clustering: general and K-MeansBilinear PCA model and clusteringData Scatter Decompositions: Quadratic and Absolute Contributions of nominal featuresExplications of Quadratic criterionOne-by-one cluster extraction: Anomalous patterns and iK-MeansIssue of the number of clustersComments on optimisation problemsConclusion and future work

Equivalent criteria (1)

A. Bilinear residuals squared MIN

Minimizing difference between data andcluster structureB. Distance-to-Centre Squared MIN

Minimizing difference between data andcluster structure

N

i Vv

ive1

2

K

kk

Si

icdWk1

2 ),(

Equivalent criteria (2)

C. Within-group error squared MIN

Minimizing difference between data andcluster structureD. Within-group variance Squared MIN

Minimizing within-cluster variance

2

1

)( iv

N

ikv

Vv Si

yck

K

kkk SS

1

2 )(||

Equivalent criteria (3)E. Semi-averaged within distance squared MIN

Minimizing dissimilarities within clustersF. Semi-averaged within similarity squared

MAX

Maximizing similarities within clusters

||/),(1 ,

2k

K

k Sji

Sjidk

jijiawhereSjia k

K

k Sji k

,),(|,|/),(1 ,

Equivalent criteria (4)

G. Distant Centroids MAX

Finding anomalous typesH. Consensus partition MAX

Maximizing correlation between sought partition and given variables

||1

2k

K

k Vv

kv Sc

),(1

vSV

v

Equivalent criteria (5)

I. Spectral Clusters MAX

Maximizing summary Raileigh quotient over binary vectors

K

kk

Tkk

TTk zzzYYz

1

/

Gower’s controversy: 2N+1 entities

Two-cluster possibilities W(c1,c2/c3)= N2 d(c1,c2) W(c1/c2,c3)= N d(c2,c3) W(c1/c2,c3)=o(W(c1,c2/c3))

Separation over grand mean/median rather than just over distances (in the most general d setting)

c1 c2 c3

N N 1

OutlineClustering: general and K-MeansBilinear PCA model and clusteringData Scatter Decompositions: Quadratic and Absolute Contributions of nominal featuresExplications of Quadratic criterionOne-by-one cluster extraction strategy: Anomalous patterns and iK-MeansComments on optimisation problemsIssue of the number of clustersConclusion and future work

PCA inspired Anomalous Pattern Clustering

yiv =cv zi + eiv,

where zi = 1 if iS, zi = 0 if iS

With Euclidean distance squared

Si

V

vSviv

V

vSSv

N

i

V

viv cyNcy

1

2

1

2

1 1

2 )(

Si

SSS

N

i

cidNcdid ),()0,()0,(1

cS must be anomalous, that is,

interesting

Spectral clustering can be not optimal (1)

Spectral clustering (becoming popular, in a different setting): Find maximum eigenvector by

maximising

over all possible x

Define zi=1 if xi>a ; zi=0, if xi a, for some a

xxxYYx TTT /

Spectral clustering can be not optimal (2)

Example (for similarity data):

2 6 19 20

3 4

5

1

z | 0.681 0.260 0.126 0.168 |i | 1 2 3-5 6-20

This cannot be typical…

Initial setting with Anomalous Pattern Cluster

Tom Sawyer

Anomalous Pattern Clusters: Iterate

0

Tom Sawyer

12

3

iK-Means:Anomalous clusters + K-meansAfter extracting 2 clusters (how one can know that 2 is right?)

Final

iK-Means:Defining K and Initial Setting with Iterative Anomalous Pattern Clustering

Find all Anomalous Pattern clusters Remove smaller (e.g., singleton) clusters Put the number of remaining clusters as K and initialise K-Means with their centres

OutlineClustering: general and K-MeansBilinear PCA model and clusteringData Scatter Decompositions: Quadratic and Absolute Contributions of nominal featuresExplications of Quadratic criterionOne-by-one cluster extraction: Anomalous patterns and iK-MeansIssue of the number of clustersComments on optimisation problemsConclusion and future work

Study of eight Number-of-clusters methods (joint work with Mark Chiang):

• Variance based:Hartigan (HK)

Calinski & Harabasz (CH) Jump Statistic (JS)• Structure based:

Silhouette Width (SW)• Consensus based:

Consensus Distribution area (CD)Consensus Distribution mean (DD)

• Sequential extraction of APs (iK-Means):Least Square (LS)Least Moduli (LM)

Experimental resultsat 9 Gaussian clusters (3 size patterns), 1000 x 15 data

Estimated number of clusters

Adjusted Rand Index

Large spread

Small spread

Large spread

Small spread

HK

CH

JS

SW

CD

DD

LS

LM

1-time winner 2-times winner 3-times winner

Two winners counted each time

Clustering: general and K-MeansBilinear PCA model and clusteringData Scatter Decompositions: Quadratic and Absolute Contributions of nominal featuresExplications of Quadratic criterionOne-by-one cluster extraction: Anomalous patterns and iK-MeansIssue of the number of clustersComments on optimisation problemsConclusion and future work

Some other data recovery clustering models

Hierarchical clustering: Ward agglomerative and Ward-like divisive, relation to wavelets and Haar basis (1997) Additive clustering: partition and one-by-one clustering (1987) Biclustering: box clustering (1995)

Hierarchical clustering for conventional and spatial data

Model: Same

Cluster structure: 3-valued z representing a splitA split S=S1+S2 of a node S in children S1, S2:

zi = 0 if i S, = a if iS1= -b if iS2If a and b taken to z being centred, thenode vectors for a hierarchy formorthogonal base (an analogue to SVD)

,1

ivikkv

K

kiv ezcy

S1 S2

S

Last:Data recovery K-Means-wise model is an

adequate tool that involves wealth of interesting criteria for mathematical investigation

L1 criterion remains a mystery, even for the most popular method, the PCA

Greedy-wise approaches remain a vital element both theoretically and practically

Evolutionary approaches started sneaking in: should be given more attention

Extending the approach to other data types is a promising direction