explorations

64 Pages·2014·3.91 MB·English

Volume 15, Number 2 Published by the Association for Computing Machinery
December 2013 Special Interest Group on Knowledge Discovery and Data Mining
SIGKDD
explorations
TABLE OF CONTENTS
Position Paper
1 Clustering high dimensional data: Examining differences and
commonalities between subspace clustering and text clustering –
A position paper
Hans-Peter Kriegel, Eirini Ntoutsi
Contributed Articles
9 Mining Text and Social Streams: A Review
Charu C. Aggarwal
20 Mining Social Media with Social Theories: A Survey
Jiliang Tang, Yi Chang, Huan Liu
30 Brain Network Analysis: a Data Mining Perspective
Xiangnan Kong, Philip S. Yu
39 Predictive Analysis of Engine Health for Decision Support
Shubhabrata Mukherjee, Aparna S. Varde, Giti Javidi, Ehsan Sheybani
49 OpenML: networked science in machine learning
Joaquin Vanschoren, Jan N. van Rijn, Bernd Bischl, Luis Torgo
Editor-in-Chief:
Bart Goethals
Associate Editors:
Charu Aggarwal
Srinivasan Parthasarathy
Ankur Teredesai
http://www.kdd.org/explorations/

Clustering high dimensional data: Examining differences
and commonalities between subspace clustering and text
clustering - A position paper
Hans-Peter Kriegel Eirini Ntoutsi
Ludwig-Maximilians-UniversityofMunich Ludwig-Maximilians-UniversityofMunich
Germany Germany
[email protected]fi.lmu.de [email protected]fi.lmu.de
ABSTRACT bothclustersandthedimensionsuponwhichtheseclusters
are defined. To this end, both domains evaluate the impor-
The goal of this position paper is to contribute to a clear
tance of the dimensions for the clustering task and appro-
understandingofthecommonalitiesanddifferencesbetween
priatelyincorporatethisimportanceinthedistancefunction
subspace clustering and text clustering. Often text data is
and consequently, in the clustering algorithm. Besides the
foistedasanidealfitforsubspaceclusteringduetoitshigh
common goal and the similar approach though, there are
dimensionalnatureandsparsityofthedata. Indeed,thear-
fundamental differences between the two domains. To the
eas of subspace clustering and text clustering share similar
best of our knowledge, no work that elaborates on how the
challengesandthesamegoal,thesimultaneousextractionof
two areas are related has been proposed so far. Our goal of
bothclustersandthedimensionswheretheseclustersarede-
thisworkistoserveasapointforfruitfulexchangeofideas
fined. However, there are fundamental differences between
and techniques between the two domains.
thetwoareasw.r.tobjectfeaturerepresentation,dimension
The rest of the paper is organized as follows: A short
weighting and incorporation of these weights in the dissim-
overviewofsubspaceclusteringandtextclusteringdomains
ilarity computation. We make an attempt to bridge these
is presented in Sections 2 and 3, respectively. The differ-
twodomainsinordertofacilitatetheexchangeofideasand
ences and commonalities between the two domains are dis-
best practices between them.
cussedinSection4andrefertothefollowingaspects: object-
feature representation (Section 4.1), dimension weighting
Keywords
(Section 4.2) and incorporating dimension weights in the
high dimensional data, subspace clustering, text clustering similarity/distance function (Section 4.3). In Section 5, we
overview a few existing approaches that make use of con-
1. INTRODUCTION cepts from both domains. Section 6 concludes our work.
SubspaceClustering isanewresearchareareceivingalotof
attention from the community [23; 18; 19; 8]. In contrast 2. SUBSPACE CLUSTERING IN A NUT-
to traditional clustering that partitions objects in the full
SHELL
dimensional feature space, the subspace clustering methods
simultaneously partition both objects and dimensions. A Theareaofsubspaceclusteringhaslatelyemergedasasolu-
subspaceclusterisdefinedintermsofbothitsmembersand tiontotheproblemofhighdimensional data,asitisdifficult
the dimensions where these members are grouped together. tofindmeaningfulclustersinhundredsoreventhousandsof
dimensions. Different features might be relevant for differ-
As with traditional full dimensional clustering, subspace
entclustersandtherefore,thegoalofsubspaceclusteringis
clustering is not confined to specific data types or appli-
to find both the cluster members and the dimensions upon
cationdomains. Relatedworkoftenmentionstext data asa
which these members form a cluster.
candidate application for subspace clustering [23; 8] due to
the high dimensionality and sparsity of the text data [26]. Thisisincontrasttotraditionalclusteringthatsearchesfor
Typically, documents are represented as vectors in a very clusters in the full dimensional feature space [15; 13]. Also,
highdimensionalspacewheretheentriesofthisvectorrep- this is different from global dimensionality reduction tech-
resent words appearing in the document collection. niques like PCA [16] that reduce the dimensionality of the
featurespaceandsearchforclustersinthereduced(though
Text data have been studied extensively in the fields of In-
still full) dimensional space. Several subspace clustering
formation Retrieval and Text Mining. The text clustering
methods have been proposed in the literature so far (for a
domain [4] is among the most important ones with a lot of
comprehensiveoverviewoftheareaseee.g.,[23],[18],[21]).
applications like organization, summarization and indexing
ofdocumentcontent. Nowadays,itisaveryactiveresearch A crucial point in subspace clustering is the decision about
field due to the abundance of textual data. the relevant dimensions that should be considered for clus-
tering. Givenad-dimensionaldataset,thenumberofpossi-
Both subspace clustering and text clustering domains face
blesubspacesis2d−1,thereforeitisinfeasibletoexamineall
the challenge of high dimensionality and aim at extracting
possiblesubspacesforclusters. Thesolutionistoefficiently
navigatethroughthesearchspaceofallpossiblesubspaces;
tothisend,twodifferentapproacheshavebeenproposedin
SIGKDD Explorations Volume 15, Issue 2 Page 1
the literature [18]: • Stopword removal: A stopword is a word which is not
thought to convey any meaning as a dimension in the
• Bottom-up approaches: They start with 1–D sub- vectorspace(i.e.,withoutcontext),e.g.,thewords“the”,
spaces (i.e., single dimensions) and iteratively merge “and”. Usually the document words are compared with
lower dimensional subspaces in order to compute respect to a known list of stopwords and the detected
higher dimensional subspaces. The downward-closure stopwords are removed from the document.
property of Apriori [6] is usually employed to prune
non-appropriatesubspaces. Thekeyinthiscategoryis • Rare words removal: Rare words, i.e., words that ap-
themergingprocedure,i.e.,howlowdimensionalsub- pearinveryfewdocumentsareusuallyremovedsincethey
spaces would be merged to yield higher dimensional are considered not to capture much information about
subspaces. Once the appropriate subspaces are de- somecategoryofdocuments[30]. Athresholdonthenum-
tected, the clustering algorithm is applied over each berofdocumentsisusedthatdetermineswhetheraword
subspace similarly to full dimensional clustering. is rare and should be removed from the feature space as
an outlier.
To this category belong methods like CLIQUE [5],
MAFIA[22]andSUBCLU[17]. Thesemethodsresult Althoughpreprocessingresultsinsomesortofdimensional-
inoverlappingclusters;anobjectmightbeassignedto ity reduction, the number of dimensions remains very high
more than one clusters defined in different subspaces (hundreds or thousands of keywords1). To deal with this
of the original feature space. problem several dimensionality reduction techniques have
beenproposedwhichcanbedistinguishedintofeaturetrans-
• Top-downapproaches: Theystartsearchinginthefull
formation and feature selection techniques. The feature
dimensional space and iteratively learn the “correct”
transformation techniques try to reduce the dimensional-
subspace for each point or cluster. The key in this
itytoafewernewdimensionswhicharecombinationsofthe
category is how to learn the “correct” subspace for a
original dimensions. In this category belongs methods like
point or cluster.
Principal Component Analysis (PCA) [16] and Latent Se-
Representatives of this category are methods like mantic Indexing (LSI) [20]. Because they are global meth-
PROCLUS[3],DOC[25],COSA[11]andPreDeCon[9]. ods though, they might be problematic for topics defined
upon different keywords. Feature selection methods aim at
Top-downsubspaceclusteringmethodsareclosertothetext removing dimensions which seem irrelevant for modeling,
clusteringmethodssincetheysimultaneouslysearchforthe e.g.,wordsappearingveryofteninacollectionorveryrarely,
best non-overlapping partitioning of the data and the best akaoutliers. Nevertheless,theresultingfeaturespaceisstill
subspace for each partition. Therefore, we refer to the top- highdimensionalandthereforedimensions’weightingtakes
down subspace clustering methods from now on. place as we will explain in the following sections.
Typical examples of distance-based text clustering algo-
3. TEXTCLUSTERINGINANUTSHELL rithms are hierarchical clustering and k-Means [4], while
The goal of text clustering is to organize documents into cosine similarity is the commonly used similarity function.
clusters of similar content, the so-called topics. Documents
aretypicallyrepresentedintermsoftheircomponentwords, 4. DIFFERENCES BETWEEN SUBSPACE
throughthevectorspacemodel [27]. Usually,noinformation
CLUSTERING AND TEXT CLUSTER-
abouttheorderofthewordsisconsideredinthismodeland
ING
thusthismodelisalsoknownasthebagofwords representa-
tion model. Although more elaborate representations have In the previous sections, we presented a short introduction
beenproposedsuchasmulti-wordtermsandN-Grams[28], totheareasofsubspaceclusteringandtextclustering. Here-
thevectorspacemodelremainsthemostpopularone. This after,weshowhowthetwoareasarerelatedbypointingout
representation though results in a high dimensional feature their commonalities and differences.
space, where features correspond to words appearing in the As already mentioned, both areas deal with high dimen-
document collection. Moreover, only a small subset of all sional data: Subspace clustering targets high dimensional
wordsappearinginthecollectionappearsineachdocument. data, though not focusing on specific application domains
As a result, the document representation is very sparse. or data types. The data instances are described in terms of
Typically,beforeapplyinganyminingtechniqueinthetext all dimensions, in the full (high) dimensional feature space.
collection, preprocessing takes place in order to remove Thevastmajorityofthealgorithmsdoesnotdealwithmiss-
non–informative words. Preprocessing consists of several ing values, though recently fault tolerant subspace cluster-
steps [7]: ing[12]hasbeenproposedtoaccommodatedatawithmiss-
ingvalues. Fromatextclustering perspective,thekeywords
• Filtering: Special characters and punctuation are re- ofadocumentcomprisethedimensionsofthefeaturespace
moved since they are considered not to hold any discrim- and documents are points or vectors in this space. There
inative power. are hundreds or thousands of keywords (thus, high dimen-
sional space) and the vectors describing each document are
• Stemming: Thewordsarereducedtotheirbaseformor
very sparse (most of the entries are null). A null entry in
stem or root with the goal of eliminating multiple occur-
this case might mean that the corresponding keyword is ir-
rencesofaword(e.g.,thewords“fish”,“fishing”,“fisher”,
relevant (e.g., consider the word “Ukraine” in a document
“fished”areallderivedfromthesameroot“fish”). Stem-
ming is language dependent. The Porter’s algorithm [24] 1The terms “word”, “keyword” “term” are used inter-
is the best known stemmer for the English language. changeably.
SIGKDD Explorations Volume 15, Issue 2 Page 2
about the importance of vegetables in our daily diet) or The matrix V is the result of the vector space model rep-
the corresponding word might be missing because it is im- resentation, called document–term matrix in this settings.
plied by the context (e.g., in a document about “Greece” Each entry v in the matrix corresponds to the value of
i,j
theword“country”mightnotbereportedsinceitisknown keyword A ∈ A in document p ∈ D. Usually v equals
j i i,j
that Greece is a country). the number of times that keyword A appears in document
j
Note also that both areas share the same goal, namely the pi. Topreventbiasingtowardslongerdocuments, thenum-
simultaneous partitioning of the data points and the di- berofoccurrencesofakeywordinadocumentisnormalized
mensions (points/ instances and dimensions/ features, re- with respect to the total number of keyword occurrences in
spectivelyaccordingtothesubspaceclusteringterminology; the document, resulting in the so-called term frequency:
documentsandkeywords,respectivelyaccordingtothetext
Definition 1. Term Frequency (TF)
clusteringterminology). Inbothcases,thenotionofaclus-
The term frequency of a term/keyword A ∈ A in a docu-
terincludes,exceptfortheclustermembers,andthedimen- j
ment p ∈D is defined as follows:
sions where these members are similar enough to form the i
csulubsstpear.ceInclusustbesrpsa,cwehcelruesatseriinngt,exthtecsleusctleursitnegrsasarteopkincosw. n as TFj,i = (cid:80)nkjn,ik,i
The problem of simultaneously partitioning both the data
where the numerator represents the number of occurrences
andthedimensionsistackledinasimilarwaybybothsub-
of keyword A in document p and the denominator repre-
j i
spaceclusteringandtextclusteringareas. Inparticular,the
sents the occurrences of all keywords A ∈A in p .
k i
solutioninvolvesanappropriateweightingofthedimensions
according to their relevance for the clustering task and the TheTFscoreexpresseshowimportantakeywordiswithina
incorporationoftheseweightsinthedissimilarityfunction2 document. Itsvalueslieinthe[0,1]rangewithlargervalues
and consequently, in the clustering algorithm. indicatingmoreimportantkeywords. Hereafter,weconsider
We present the commonalities and differences between the theentriesofmatrixV tobetheTFvaluesofthekeywords
two domains with respect to the following aspects: (i) the in the documents of the collection. Note that since not all
object–feature representation (Section 4.1), (ii) the weight- keywords in the collection appear in each single document,
ing of the dimensions (Section 4.2) and (iii) the incorpora- there are a lot of null entries in this matrix corresponding
tion of dimensions’ weighting in the dissimilarity functions to non-appearing words in a document.
andconsequently,intheclusteringalgorithms(Section4.3).
4.1.3 Discussiononobject-featurerepresentationdif-
4.1 Object–featurerepresentationdifferences ferences
Let D = {p ,p ,...,p } be a dataset of n objects and let There is a crucial difference in the object-feature represen-
1 2 n
A=(A , A , ..., A ) be the d-dimensional feature space. tation of the two domains. In subspace clustering, there
1 2 d
LetV bea|D|×|A|matrix,referredhereafterastheobject– are no semantics on the values of the dimensions and all
feature value matrix, where the rows are the objects and values are considered to be of the same importance. On
the columns are the dimensions. Each entry v in this thecontrary,intextclusteringgreatervaluesindicatemore
i,j
matrix corresponds to the value of the object p ∈D in the important dimensions/keywords. For example, consider a
i
dimension A ∈A. We explain hereafter what these values keyword A appearing in two documents p ,p with term
j 1 1 2
are for each domain. frequencies 0.8,0.2, respectively. Since 0.8>0.2, A is con-
1
sideredtobemoreimportantfordocumentp comparedto
1
4.1.1 Object–featurerepresentationinsubspaceclus-
document p . Such a value differentiation though does not
2
tering
take place in the subspace clustering domain.
Insubspaceclustering,eachobjectisconsideredtohaveval- Although, as already mentioned, recently subspace cluster-
ues for all dimensions3. So, the construction of the object– ing over missing data has been proposed [12], missing data
feature value matrix V is rather straightforward. The rows are conceptually different from non–appearing words in a
correspond to the objects in the database and the columns document. A missing value in subspace clustering indicates
correspondtothedimensions. Eachentryvi,j inthematrix a value that is unknown rather than a value that is not im-
containsthevalueoftheobjectpi ∈DindimensionAj ∈A. portantforthedescriptionofadocumentorredundantw.r.t.
The original values of the dimensions may refer to different alreadyexistingwordsinthedocument(asitisusuallythe
scalesofreference. Topreventattributeswithinitiallylarge case for non appearing words in a document).
ranges (e.g., salary) from outweighing attributes with ini-
4.2 Dimensionweightingdifferences
tialsmallerranges(e.g.,age),anormalization processtakes
place prior to clustering. In this way, different dimensions Bothsubspaceclusteringandtextclusteringdomainsrelyon
can be compared meaningfully. somenotionofweightingforthedifferentdimensionsbased
on their importance for the clustering task.
4.1.2 Object–featurerepresentationintextclustering
Insubspaceclustering,theimportantdimensionsarelearned
In text clustering, the database D corresponds to a collec- either for an instance/object (instance–based approaches)
tion of documents and the dimensions A correspond to the or for a cluster (cluster–based approaches). Representative
distinctkeywordsinthiscollection. Weassumethatthepre- weightingschemesforbothapproachesarepresentedinSec-
processing step (c.f., Section 3) has been already applied. tion 4.2.1. In text clustering, the most well known weight-
2We use the term “dissimilarity” to refer to either distance ing schema is inverse document frequency (IDF) described
or similarity function. in Section 4.2.2.
3Recently,subspaceclusteringoverdatawithmissingvalues Wecanmodelthedimensionweightingintermsofamatrix
has been considered [12]. W referred to hereafter as the object–feature weight matrix.
SIGKDD Explorations Volume 15, Issue 2 Page 3
Inthegeneralcase,W isa|D|×|A|matrixwheretherows number of clusters k is given as input to the algorithms
correspondtoobjectsandthecolumnscorrespondtodimen- of this category. The definitions and details below are
sions. Each entry w in this matrix represents the weight from the algorithm PROCLUS [3]. Each cluster (called
i,j
(or importance) of the dimension A ∈ A for the object projected cluster) is represented by its medoid and it
j
p ∈D. Weexplainhereafterhowtheseentriesarefilledfor is assigned a subspace of preferred4 dimensions. Let
i
each domain. C ={c ,c ,...,c }beasetofk medoids(formoredetails
1 2 k
on the initial selection and refinement of this set, please
4.2.1 Dimensionweightinginsubspaceclustering
refer to [3]).
The importance of a dimension is evaluated either with re-
Locality of a cluster: ThelocalityL ofaclusterc ∈C
i i
specttosomeinstances(instance–basedapproaches)orwith
isdefinedasthesetofobjectsinDthatarewithindistance
respect to some cluster (cluster–based approaches). Both
δ fromitsmedoidc . Thedistanceiscomputedinthefull
i i
approaches rely on the so called “locality assumption” ac-
dimensionalspace,whereasthedistancethresholdδ isthe
i
cording to which, the subspace preference for an object or
minimum distance of c from any other medoid c ∈ C ,
i j
a cluster can be learned from its local neighborhood in the
i.e., δ =min{dist(c ,c )}, i(cid:54)=j.
i i j
full dimensional space. We describe each case below.
Preferred dimensions of a cluster: For each medoid
c , the average distance between c and the objects in its
(i) Dimension weighting in instance–based approaches i i
locality L is computed along each dimension A ∈A, de-
The preferred dimensions subspace is defined per instance i j
noted by X . If this value is as small as possible w.r.t.
and is learned by evaluating the local neighborhood of i,j
the statistical expectation, then A is a preferred dimen-
the instance/object in the full dimensional feature space. j
sion for cluster c . The statistical expectation is evaluated
The definitions and details below are from the algorithm i
in terms of the mean Y and the standard deviation σ .
PreDeCon [9]. i i
ThevalueZ = Xi,j−Yi iscomputedwhichindicateshow
Localityofaninstance: Foranobjectp∈D,itslocality i,j σi
the average distance between cluster c and its locality L
or neighborhood (N (p)) consists of all objects that fall i i
ε in dimension A is related to the average distance in all
within distance (cid:15) from p in the full dimensional space. j
dimensions. The lowest Z values are picked leading to a
i,j
Preferred dimensions of an instance: Roughlyspeak- totalofk∗l dimensions,wherel istheaveragedimension-
ing, an object prefers a dimension if it builds “com- ality per cluster given as input to the algorithm.
pact”neighborhoods,i.e.,neighborhoodsofsmallvariance,
After the weighting, each cluster in C is assigned a sub-
across this dimension. The variance in the neighborhood
space of preferred dimensions. Usually a bitvector of all
of p along some dimension A is defined as follows:
j dimensions is used to model the preferences and to distin-
guishbetweenpreferred(value1)andnonpreferred(value
Definition 2. Variance along a dimension
0) dimensions for a cluster [2].
Letp∈D andε∈R. Thevariance in the neighborhood of
p N (p) along a dimension A ∈A is given by: The objects assigned to a cluster “inherit” the subspace
ε j
preferences of the cluster, therefore each object can be as-
Var (N (p))= (cid:80)q∈Nε(p)(distAj(p,q))2 sumed to have a bitvector of dimension preferences. This
Aj ε |N (p)| way,wecanfilltheentriesoftheobject–featureweightma-
ε
trix W with values 1 and 0 indicating preferred and non–
where dist (p,q) is the distance of p,q in dimension A .
Aj j preferred,respectively,dimensionsforanobject. Notethat
this way the objects belonging to the same cluster, would
Asmallvariance,withrespectto agivenvariancethresh- have the same bitvectors of subspace preferences.
old δ, indicates a preferable dimension. For each p ∈ D,
its subspace preference vector w¯p is built [9] which distin- 4.2.2 Dimensionweightingintextclustering
guishesbetweenpreferableandnon-preferabledimensions.
In text clustering, the importance of a keyword is evalu-
ated in terms of the whole collection of documents, rather
Definition 3. Subspace preference vector
thanperdocumentorpertopic/cluster. Usually,theInverse
Let p ∈ D, δ ∈ R and κ ∈ R be a constant with κ (cid:29) 1.
Document Frequency (IDF) is employed towards this end.
The subspace preference vector of p is defined as:
Definition 4. Inverse Document Frequency (IDF)
w¯ =(w ,w ,...w )
p 1 2 d The inverse document frequency of a keyword A in a col-
j
where: lection of documents D is given by:
(cid:26) 1 if Var (N (p))>δ |D|
wi = κ if VarAAii(Nεε(p))≤δ IDFAj = |{pi ∈D:Aj ∈pi}|
wherethedenominatorrepresentsthenumberofdocuments
Thevaluesofthesubspacepreferencevectoraretheentries
in D which contain the specific keyword/dimension A and
of the object–feature weight matrix W: k-value entries in- j
|D| is the cardinality of the collection.
dicate preferable dimensions while 1-value entries indicate
non-preferable ones.
The basic intuition behind IDF is that common keywords
(i.e., keywords appearing very often in the collection) have
(ii) Dimension weighting in cluster–based approaches
nodiscriminativepowerandthus,theyareassignedasmall
The subspace of the preferred dimensions is defined per
cluster and is learned by evaluating the local neighbor- 4Weusebothterms“preferred”and“projected”todescribe
hood of the cluster in the full dimensional space. The a dimension that is important.
SIGKDD Explorations Volume 15, Issue 2 Page 4
IDF score. Larger IDF scores indicate more discriminative Definition 6. Preference weighted distance
keywords. Let p,q ∈ D, then their preference weighted distance is
UsingtheIDFscoresofthekeywordswecanfilltheentriesof given by:
theobject–featureweightmatrixW. NotethatsinceIDFis (cid:118)
definedperkeyword,eachcolumnofthematrixcorrespond- dist (p,q)=(cid:117)(cid:117) (cid:88) 1 ·(V[p,A ]−V[q,A ])2
ing to a single keyword would be filled with the same IDF p (cid:116) W[p,Aj] j j
Aj∈A
value.
Theaboveformulaconsidersonlythepreferencesofp. The
4.2.3 Discussionondimensionweightingdifferences symmetric version is:
Although, the weighting of the dimensions is performed in
dist(p,q)=max(dist (p,q),dist (p,q))
p q
bothsubspaceclusteringandtextclusteringdomains,there
are core differences in the weighting schemes. where V[p,Aj] (V[q,Aj]) is the value of object p (q, re-
In subspace clustering, the dimension weighting is local re- spectively)indimensionAj andW[p,Aj](W[q,Aj])isthe
lyingontheneighborhoodofeachobjectorcluster. Intext weight of Aj for point p (q, respectively).
clustering, the weighting of the keywords is global and is
Note, that the dimension weights are either 1 or κ (κ>>
based upon the whole collection of documents.
1). Therefore, the similarity function weights attributes
Also, the dimension weights in subspace clustering act
withlowvariance(dimensionweightκ)considerablylower
mainly as a filter in order to distinguish between preferred
(by a factor 1/κ) than attributes with a high variance (di-
and non–preferred dimensions. For example, PreDeCon [9]
mension weight 1).
uses the dimension weights κ and 1 to model preferred
andnon-preferreddimensions,respectively. Similarly,PRO- (ii) Dissimilarity in cluster–based approaches We
CLUS[3]employstwodimensionweights,0and1with1in- adapt the distance function of PROCLUS to the generic
dicating a preferred dimension. That is, the actual weights diss() function notation.
ofthedimensionsarenotconsideredinsubspaceclustering,
Definition 7. Projected distance
but rather the information about which dimension is pre-
Let p,q ∈ D with q being the medoid of a cluster. The
ferred or not. This is in contrast to IDF weighting in text.
projecteddistanceofpwithrespecttothemedoidqisgiven
4.3 Incorporating dimension weighting in by:
similarity/distance computation differ- (cid:80) W[q,A ]∗|V[p,A ]−V[q,A ]|
ences dist(p,q)= Aj∈A j j j
|{A :W[q,A ]=1}|
j j
The weighting of the dimensions is incorporated in the dis-
W[q,A ] is the importance of dimension d for the cluster
similarityfunctioninbothsubpsaceclusteringandtextclus- j
representedbythemedoidq. Notethatinthiscase,weights
tering domains. We model this contribution in terms of a
take values in {0,1}. That is, the non important dimen-
generic dimension–weighted dissimilarity function.
sionsarenotconsideredatallduringdistancecomputation
and the important ones are equally taken into account.
Definition 5. Dimension–weighted dissimilarity
Let p,q ∈D. We denote by V[p,Aj] (V[q,Aj]) the value of 4.3.2 Dissimilaritycomputationintextclustering
object p (q, respectively) with respect to dimension A and
by W[p,A ] (W[q,A ]) the importance of A for the ojbject TF ×IDF is the most common schema for combining key-
j j j
p (q, respectively). The dissimilarity between p and q is a word weights and their values across different documents.
combination of their values and dimensions’ weights: According to this schema, the value of a keyword in a doc-
ument(TF)ismultipliedbytheimportanceofthekeyword
diss(p,q)=Aggr f(V[p,A ],V[q,A ],W[p,A ],W[q,A ])
Aj∈A j j j j in the whole collection (IDF).
The function f() combines the value and weights in each Definition 8. TF×IDF score
dimension. The total score is an aggregation of the cor- Let p ∈ D be a document and let A ∈ A be one of its
j
responding dimensions’ scores, through some aggregation
keywords. The TF×IDF score of A in p is given by:
function Aggr(). j
(TF ×IDF) =TF ×IDF
p,Aj p,Aj Aj
Wealreadydiscussedwhattheobjectvaluesanddimension Typically, the dissimilarity function is applied upon the
weightsstandforineachdomainandhowtheobject–feature TFIDFvaluesofthedocuments. Acommonlyuseddissimi-
value matrix V and the object–feature weight matrix W larityfunctionisthecosinesimilaritythatfindsthecosineof
are filled. We explain hereafter how the diss() function is the angle between the two documents. It becomes 1 if the
instantiated per domain. documents are identical and 0 if there is nothing in com-
mon between them (i.e., the vectors are orthogonal to each
4.3.1 Dissimilarity computation in subspace cluster-
other).
ing
The cosine similarity can be re-written in terms of the
We distinguish between instance–based and cluster–based generic formula diss() as follows:
approaches, as with the weighting of dimensions case (cf.,
Definition 9. Cosine similarity
Section 4.2).
Let p,q∈D be two documents. Their similarity is defined
as:
(i) Dissimilarity in instance–based approaches We
addisasp(t)tfhuencdtiisotnanncoetafutinocnt.ionofPreDeCon[9]tothegeneric cos(p,q)=(cid:114)(cid:80)Aj∈A(cid:80)(AVj[p∈,AAjV]∗[pW,A[pj,]A∗Wj])[2p∗,A(cid:114)j(cid:80)]∗AVj[∈q,AAj(V]∗[Wq,A[qj,A]∗jW][q,Aj])2
SIGKDD Explorations Volume 15, Issue 2 Page 5
Note that the weight of a keyword A is defined over the words that frequently occur within the cluster are used for
j
whole collection D thus, documents p and q share the same theassignmentprocess. Thenumberofwords,i.e.,projected
weight for this keyword, i.e.,W[p,A ]=W[q,A ]=IDF . dimensions for each cluster, is reduced at each iteration by
j j Aj
a geometric factor. In their experiments, they started with
4.3.3 Discussion on incorporation of dimension a projected dimensionality of 500 words which was finally
weightingindistancefunctiondifferences reduced to 200 words. This work introduces subspace clus-
Although,bothdomainsconsiderobjects’valuesanddimen- tering concepts to text clustering, in particular, the local
sions’weightsfordissimilaritycomputation,thereisaclear (within each cluster) weighting of the words and selection
difference between the two domains. The difference lies in of the most prominent ones for cluster centroid representa-
the fact that in text clustering the value of a dimension/ tion. TheproposedalgorithmresemblesPROCLUS[3](c.f.,
keyword in a document, i.e.,TF, also expresses some notion Section4.2.1),howeverthedistancefunctionandtheselec-
of importance for the keyword in the document. This is in tion of projected dimensions at each iteration is adopted to
contrast to subspace clustering where all values are treated the text domain. In particular, cosine function is used for
similarly and there is no importance discrimination. dissimilarity assessment and the selection of projected di-
mensions is done on the basis of their occurrences in each
This fact is also reflected in the chosen dissimilarity func-
cluster.
tions. In subspace clustering, where the values carry no
semantics on their importance, the value difference in each In [14], the authors propose a modification of W–k–
dimension(e.g.,,absoluteasinPROCLUS[3]orsquaredas Means [10], that calculates cluster specific weights for each
in PreDeCon [9]) is considered. In text clustering, though, feature during the clustering process, for text documents.
thewidelyusedcosinesimilarityfunctionmultipliestheac- Totallym×k weightsareproducedbythealgorithmwhere
tual object values at each dimension, so that higher values m is the number of features and k the number of clusters.
result in higher scores. For example, if we consider two ob- Theinitialweightsarerandomlygeneratedandrefineddur-
jectswithvalues0.8and0.6foraspecificdimension(case1), ing the iteration phase, similarly to K–Means. Based on
thecontributionofthisdimensiontothedissimilarityscore theseweights,thesubsetofkeywordsthatcanserveasclus-
willbe0.8−0.6=0.2forPROCLUSand(0.8−0.6)2 =0.04 ter label can be identified. In their experiments, the pro-
for PreDeCon, whereas for the cosine similarity it will be posedsubspacemethodperformedbetterthanstandardK–
0.8∗0.6=0.48. For two other objects, with values 0.4,0.2 Means and Bisecting–KMeans [29].
inthesamedimension(case2),theresultwouldbethesame There are also methods that extract an initial clustering
forPROCLUSandPreDeConsincetheyrelyontheirvalue with respect to the whole feature space of documents (usu-
difference which is again 0.2, however it will be different ally through TF ×IDF weighting) and refine the cluster-
forcosinesimilarity;thecontributionofthisdimensionthis ingresultbyapplyingclusteringagaininsideeachextracted
time would be: 0.4∗0.2 = 0.08 counting for the fact that clusterbutusingarefinedfeaturespace. Forexample,in[31]
the actual object values are lower in case 2 than in case 1. the authors propose the extraction and maintenance of a
Also, in subspace clustering the dimension weighting pre- two level hierarchy of global and local topics: the global
dominantlyindicateswhetherthecorrespondingobjectval- topics are extracted by applying clustering over the whole
ues should be considered or not. For example, in PRO- collectionofdocumentsinthefeaturespacederivedfromthe
CLUS [3] only projected dimensions (having weight 1) are whole collection through TF ×IDF. To extract the local
considered during dissimilarity assessment and there is no topics, clustering is applied again inside each global clus-
specialweightingperpreferreddimension. InPreDeCon[9] terusingtheclusterpopulationtogeneratethenewcluster
also,thevaluedifferencesinthepreferreddimensions(those specificfeaturespace(theIDF scoresarebasedontheclus-
having weight κ) are down-weighted by 1/κ, but again this terpopulation,ratherthanonthewholecollection),instead
holds for all preferred dimensions. On the contrary, in text of using the generic feature space extracted from the whole
clusteringtheactualweightsofthedimensionsasexpressed dataset. This way, the feature space is refined per cluster.
by their IDF values over the whole collection of documents Discussion Although there are some methods for text
are considered and contribute proportionally to the dissim- clustering that make use of subspace clustering concepts,
ilarity score. like [1], [14] and methods that refine the feature space
through global and local weighting like [31], there are still
manythingsthatthetwodomainscouldexchangeandben-
5. TOWARDS COMBINING THE BEST OF
efit from each other, like for example a concurrent incorpo-
BOTHWORLDS ration of both global and local weighing of features in the
Intheprevioussection,weelaboratedonthedifferencesand clustering process. There is a bunch of subspace clustering
commonalitiesbetweensubspaceclusteringandtextcluster- algorithmsintheliterature,athoroughreportontheirper-
ingdomainswithrespecttodatarepresentation,dimension formance over text data would be very useful as a starting
weighting and incorporation of these weights in the dissim- point. Of course, the weighting of the dimensions and the
ilarity functions. In this section, we overview approaches distancefunctionshouldrespectthepeculiaritiesofthetext
that make use of concepts from both domains. data as discussed in Section 4. There are also other fields
In [1], the authors study the problem of semi-supervised which resemble the semantics of the data values in text;
textclassificationanduseclusteringtocreatethesetofcat- forexample,inrecommendationapplicationshigherratings
egoriesfortheclassificationofdocuments. Toimproveclus- indicatemoredesireditems,e.g.,moviesorrestaurants. Sub-
teringquality,theyuseamodifiedversionofK–Meanswhere space clustering is relevant to such kind of data too, since
they iteratively refine both the clusters and the dimensions groups of users with different item preferences might exist,
insideeachcluster. Inparticular,ateachiterationtheyfilter however as with text, the data semantics should be also
outsomeofthewordsofthefeaturespaceensuringthatonly taken into account.
SIGKDD Explorations Volume 15, Issue 2 Page 6
6. CONCLUSIONS 8. REFERENCES
Weoutlinedthedifferencesandcommonalitiesbetweensub-
spaceclusteringandtextclusteringandwearguedthattext [1] C. Aggarwal, S. Gates, and P. Yu. On using par-
data is not a straightforward application domain for sub- tial supervision for text categorization. Knowledge and
space clustering as it is often suggested in the literature. Data Engineering, IEEE Transactions on, 16(2):245–
Althoughbothdomainssharethesamegoal,theconcurrent 255, Feb 2004.
extraction of clusters and dimensions where these clusters
are formed, and deal with similar challenges, high dimen- [2] C. C. Aggarwal, J. Han, J. Wang, and P. S. Yu. A
sional and sparse data, there are key differences between frameworkforprojectedclusteringofhighdimensional
the two domains which we summarize bellow: data streams. In VLDB, pages 852–863, 2004.
[3] C.C.Aggarwal,C.M.Procopiuc,J.L.Wolf,P.S.Yu,
• Theobjectvaluesintextclusteringalsoreflecttheim-
andJ.S.Park.Fastalgorithmsforprojectedclustering.
portanceofakeywordwithinadocument(TFscores).
In Proceedings of the ACM International Conference
onManagementofData(SIGMOD),Philadelphia,PA,
• Missing data in subspace clustering imply unknown
1999.
data, whereas in text clustering a non-appearing key-
word indicates words that are not important or are [4] C.C.AggarwalandC.Zhai.Asurveyoftextclustering
probably redundant for the description of the docu- algorithms. In C. C. Aggarwal and C. Zhai, editors,
ment. Mining Text Data, pages 77–128. Springer, 2012.
• Bothdomainsusesomenotionofdimensionweighting [5] R. Agrawal, J. Gehrke, D. Gunopulos, and P. Ragha-
tocountforthedifferentimportanceofthedimensions van.Automaticsubspaceclusteringofhighdimensional
for the clustering task. data for data mining applications. In Proceedings of
the ACM International Conference on Management of
• In subspace clustering, dimension weighting is local Data (SIGMOD), Seattle, WA, 1998.
and is derived with respect to the local neighborhood
[6] R. Agrawal and R. Srikant. Fast algorithms for mining
of a point or a cluster.
association rules in large databases. In Proceedings of
the 20th International Conference on Very Large Data
• In text clustering, dimension weighting is global and
Bases,VLDB,pages487–499,SanFrancisco,CA,USA,
is defined with respect to the whole collection of doc-
1994. Morgan Kaufmann Publishers Inc.
uments.
[7] N.O.AndrewsandE.A.Fox.Recentdevelopmentsin
• Dimension weighting in subspace clustering mainly
document clustering. Technical report, Computer Sci-
filters important from non-important dimensions by
ence, Virginia Tech, 2007.
assigning the same weight to all important/non-
important dimensions. [8] I. Assent. Clustering high dimensional data. Wiley In-
terdisciplinary Reviews: Data Mining and Knowledge
• Intextclustering,theactualweightsofthedimensions Discovery, 2(4):340–350, 2012.
are used in the dissimilarity function.
[9] C. Bo¨hm, K. Kailing, H.-P. Kriegel, and P. Kro¨ger.
• The dissimilarity function in text clustering also con- Density connected clustering with local subspace pref-
siderstheimportanceofakeywordwithinadocument erences. In Proceedings of the 4th IEEE International
(TFscores),soasmoreimportantkeywordsaregiven Conference on Data Mining (ICDM), Brighton, UK,
higher scores compared to less important ones. pages 27–34, 2004.
[10] E. Y. Chan, W. K. Ching, M. K. Ng, and J. Z.
Our goal of this work is to point out the differences be-
Huang. An optimization algorithm for clustering using
tweenthetwodomainsandshowtheircommonalities. Text
weighted dissimilarity measures. Pattern Recognition,
clustering is an established area of research with a bulk of
37(5):943–952, May 2004.
applications and an increased interest nowadays due to the
webandthedigitizationofinformation. Ontheotherhand, [11] J. H. Friedman and J. J. Meulman. Clustering objects
thesubspaceclusteringareagrowsfastashighdimensional- onsubsetsofattributes.JournaloftheRoyalStatistical
itycomprisesoneofthebasicfeaturesofnowadaysdata. A Society: Series B (Statistical Methodology), 66(4):825–
few works combine concepts from both domains and could 849, 2004.
serve as a starting point for further exchange of ideas and
bestpracticesbetweenthesedomainsthatwouldbefruitful [12] S. Gu¨nnemann, E. Mu¨ller, S. Raubach, and T. Seidl.
for their further development. Flexiblefaulttolerantsubspaceclusteringfordatawith
missing values. In D. J. Cook, J. Pei, W. Wang, O. R.
Za¨ıane, and X. Wu, editors, ICDM, pages 231–240.
7. ACKNOWLEDGMENTS
IEEE, 2011.
The authors would like to thank Arthur Zimek for useful
commentsandsuggestionstoimprovethequalityofthepa- [13] J. Han and M. Kamber. Data Mining: Concepts and
per. Techniques. Academic Press, 2nd edition, 2006.
SIGKDD Explorations Volume 15, Issue 2 Page 7
[14] J. Z. Huang, M. K. Ng, H. Rong, and Z. Li. Auto- [28] M. Shafiei, S. Wang, R. Zhang, E. Milios, B. Tang,
mated variable weighting in k-means type clustering. J. Tougas, and R. Spiteri. A systematic study of doc-
IEEE Transactions on Pattern Analysis and Machine umentrepresentationanddimensionreductionfortext
Intelligence, 27(5):657–668, 2005. clustering. Technical Report CS-2006-05, Faculty of
Computer Science, Dalhousie University, Canada, July
[15] A. K. Jain, M. N. Murty, and P. J. Flynn. Data clus- 2006.
tering: a review. ACM Computer Surveys, 31:264–323,
1999. [29] M. Steinbach, G. Karypis, and V. Kumar. A compar-
ison of document clustering techniques. In In KDD
[16] I. T. Jolliffe. Principal Component Analysis. Springer, Workshop on Text Mining, 2000.
2nd edition, 2002.
[30] Y. Yang and J. O. Pedersen. A comparative study on
[17] K. Kailing, H.-P. Kriegel, and P. Kro¨ger. Density- feature selection in text categorization. In Proceedings
connected subspace clustering for high-dimensional oftheFourteenthInternationalConferenceonMachine
data. In Proceedings of the 4th SIAM International Learning, pages 412–420, 1997.
ConferenceonDataMining(SDM),LakeBuenaVista,
[31] M.Zimmermann,E.Ntoutsi,andM.Spiliopoulou.Ex-
FL, 2004.
tracting opinionated (sub)features from a stream of
product reviews. In Discovery Science - 16th Interna-
[18] H.-P. Kriegel, P. Kro¨ger, and A. Zimek. Clustering
tional Conference (DS), Singapore, 2013.
high dimensional data: A survey on subspace cluster-
ing, pattern-based clustering, and correlation cluster-
ing. ACM Transactions on Knowledge Discovery from
Data (TKDD), 3(1):1–58, 2009.
[19] H.-P. Kriegel, E. Ntoutsi, M. Spiliopoulou,
G. Tsoumakas, and A. Zimek. Mining complex
dynamic data. Tutorial at the European Conference
on Machine Learning and Knowledge Discovery in
Databases (ECML PKDD), Athens, Greece., 2011.
[20] T. K. Landauer, D. S. McNamara, S. Dennis, and
W. Kintsch. Handbook of Latent Semantic Analysis.
Lawrence Erlbaum Associates, 2007.
[21] E.Mu¨ller,S.Gu¨nnemann,I.Assent,andT.Seidl.Eval-
uatingclusteringinsubspaceprojectionsofhighdimen-
sional data. In Proceedings of the 35nd International
Conference on Very Large Data Bases (VLDB), Lyon,
France, 2009.
[22] H. Nagesh, S. Goil, and A. Choudhary. Adaptive grids
for clustering massive data sets. In Proceedings of the
1st SIAM International Conference on Data Mining
(SDM), Chicago, IL, 2001.
[23] L. Parsons, E. Haque, and H. Liu. Subspace clustering
for high dimensional data: A review. SIGKDD Explo-
rations, 6(1):90–105, 2004.
[24] M. F. Porter. An algorithm for suffix stripping. Pro-
gram, 14(3):130–137, 1980.
[25] C. M. Procopiuc, M. Jones, P. K. Agarwal, and T. M.
Murali. A Monte Carlo algorithm for fast projective
clustering. In Proceedings of the ACM International
ConferenceonManagementofData(SIGMOD),Madi-
son, WI, 2002.
[26] M. Radovanovic, A. Nanopoulos, and M. Ivanovic. On
theexistenceofobstinateresultsinvectorspacemodels.
In SIGIR, pages 186–193, 2010.
[27] G. Salton. Automatic Text Processing: The Trans-
formation, Analysis, and Retrieval of Information by
Computer. Addison-Wesley, 1989.
SIGKDD Explorations Volume 15, Issue 2 Page 8