DECK: Detecting Events from Web Click-through Data
Ling Chen1 Yiqun Hu2 Wolfgang Nejdl1
L3S Research Center1 School of Computer Engineering2
University of Hannover, Germany Nanyang Technological University, Singapore
lchen,nejdl@L3S.de yiqun.hu@gmail.com
Abstract
In the past few years there has been increased research
interest in detecting previously unidentified events from Web
resources. Our focus in this paper is to detect events from
the click-through data generated by Web search engines.
Existing event detection algorithms, which mainly study the
news archive data, cannot be employed directly because of
the following two unique features of click-through data: 1)
the information provided by click-through data is quite lim-
ited; 2) not every query issued to a Web search engine cor-
responds to an event in the real world. In this paper, we
address this problem by proposing an effective algorithm
which Detects Events from ClicK-through data (DECK).
We firstly transform click-through data to the 2D polar
space by considering the semantic dimension and temporal
dimension of queries. Robust subspace estimation is per-
formed to detect subspaces such that each subspace con-
sists of queries of similar semantics. Next, we prune un-
interesting subspaces which do not contain queries corre-
sponding to real events by simultaneously considering the
respective distribution of queries along the semantic dimen-
sion and the temporal dimension in each subspace. Finally,
events are detected from interesting subspaces using a non-
parametric clustering technique. Compared with an exist-
ing approach, our experimental results based on real-life
data have shown that the proposed approach is more accu-
rate and effective in detecting real events from click-through
data.
1. Introduction
With the prevailing publishing activities over the Inter-
net, the Web of nowadays covers almost every object and
event in the real world [18]. This phenomenon has moti-
vated the data mining community to discover knowledge
such as topics, events and stories from large volumes of
Web data [11, 3, 10, 15]. Although most of the existing
work focus on analyzing and extracting events from tra-
ditional Web data such as Web content and structures, re-
cent research direction is to detect events from user gener-
ated Web data. For example, the effort presented in [19]
aims to detect events from Weblogs, which have become
one of the dominant forms of self-publication on the Inter-
net. In [20], tags indicating real events are extracted from
a set of user-supplied tag data annotating pictures at the
Flickr site1 . Our focus in this paper is to detect events from
Web click-through data, which is generated by users (pas-
sively) when interacting with Web search engines. An ex-
cerpt of click-through data containing five entries is shown
in Table 1. Each entry of Web click-through data basically
records the following four types of information: an anony-
mous user identity, the query issued by the user, the time at
which the query was submitted for search, and the URL of
clicked search result [13].
The reason that click-through data serves as a promis-
ing data source for event detection is trifold. Firstly, sim-
ilar to the other resources on the Web, the click-through
data can be regarded as a sensor of the real world as well.
Click-through data containing query keywords and clicked
pages of users often reflect users’ response to recent real
world events. Secondly, the sharply accentuated role of Web
search engines as an entry point to the Web has given rise
to a huge volume of click-through data, which enables ef-
fective knowledge discovery. Thirdly, click-through data is
well formatted. Complicated data preprocessing on Web
content and structure information can be avoided.
However, detecting events from click-through data is a
nontrivial problem because of the following two observa-
tions: 1) The information provided is limited, so that the
semantic knowledge cannot be identified easily and clearly.
As shown in Table 1, the URLs which represent the search
results clicked by users are incomplete. Instead of referring
to the addresses of pages, the URLs refer to the addresses
of sites. Consequently, entries having the same URL may
have different semantics and correspond to different events.
Similarly, the same query keywords may have been issued
by users to inquire about different events. 2) A large amount
of click-through data do not represent any real event in the
world. Table 2 illustrates the top 10 most frequent entries
in the click-through data logged by AOL in March, 2006.
None of them are correlated with any real events.
1 http : //www.f lickr.com/
 User Query Time Page
337 jojo 3/19/2006 15:02 http://www.jojoonline.com
tion along the temporal dimension of query sessions in each
337 jojo 3/19/2006 15:12 http://www.azlyrics.com subspace, we prune uninteresting subspaces which do not
178 poker 3/14/2006 21:44 http://www.pokerroom.com contain query sessions corresponding to real events. Finally,
178 poker 3/14/2006 21:58 http://www.pokerstars.com
178 poker 3/14/2006 21:58 http://www.partypoker.com a non-parametric clustering method is used to cluster data
points in each interesting subspace. The generated clusters
Table 1. Example of click-through data. will be returned as detected events.
ID Query Page Freq. The main contributions of this paper are summarized as
1 google http://www.google.com 38608
2 mapquest http://www.mapquest.com 16518 follows.
3 yahoo http://www.yahoo.com 13860
4 ebay http://www.ebay.com 11669
• We develop a novel algorithm, KNN-GPCA, to seg-
5 google.com http://www.google.com 9284 ment Web click-through data into broad topics. KNN-
6 yahoo.com http://www.yahoo.com 8361 GPCA improves the robustness of the existing Gen-
7 myspace.com http://www.myspace.com 7169
8 myspace http://www.myspace.com 6678 eralized Principal Component Analysis (GPCA) [16]
9 www.google.com http://www.google.com 4739 by embedding the distribution of K nearest neighbors
10 bank of america http://www.bankofamerica.com 3957
constraint within the GPCA framework and estimates
Table 2. Frequent query-page pairs. subspaces with weighted least square approximation.
• We propose an interestingness measure, which simul-
Current research in this area is still in its infancy. To the taneously considers semantic certainty and temporal
best of our knowledge, the work done by Zhao et al. [18] certainty of click-through data, to justify the likelihood
is the first and the only one which detects events from that the subspace contains query sessions correspond-
Web click-through data. In their work, they presented a ing to events.
two-phase-clustering approach to detect events. In the first
phase, entries of click-through data, represented as query- • We evaluate the performance of DECK with extensive
page pairs, are clustered based on their semantics. In the experiments on real life click-through data collected
second phase, for each semantically similar cluster, query- by AOL. We compare the performance of DECK with
page pairs are clustered again based on their frequency evo- the existing two-phase-clustering approach [18].
lution over some predefined time intervals. The cluster- The rest of the paper is organized as follows. In Sec-
ing results will be returned as detected events. Neverthe- tion 2, we illustrate the polar transformation of Web click-
less, we observed that not every cluster returned by their through data. We describe our linear subspace estimation
approach can be regarded as the representation of one real algorithm, KNN-GPCA, in Section 3. In Section 4, we first
event, since a large portion of click-through data is not re- define the interestingness measure of subspaces and then
lated with events. Therefore, pruning irrelevant data is an discuss the detection of events from interesting subspaces.
important step involved in effective event detection from Experimental results based on real-life data sets are pre-
click-through data. sented in Section 5. We review related work in Section 6
In this paper, we propose a novel algorithm DECK, De- and conclude this paper in Section 7.
tecting Events from ClicK-through data, which proceeds in 2 Polar Space Representation
the following four steps, polar transformation, subspace Given a collection of Web click-through data, Zhao et
estimation, subspace pruning, and cluster generation, as al. [18] considered the data in unit of query-page pair which
shown in Figure 1. Given a collection of Web click-through consists of a user query and a page clicked by the user. For
data, we firstly transform it to the 2D polar space. Each example, the click-through data in Table 1 will be consid-
query session of Web click-through data (A query session, ered as five distinct query-page pairs. However, in DECK,
which will be formally defined in Section 2, contains a we study Web click-through data in unit of query session,
query and a set of corresponding pages clicked by a user.), which refers to an episode of interaction between a user and
is mapped to a point in polar space such that the angle θ a Web search engine. A query session consists of a query is-
and radius r of the point respectively reflect the semantics sued by the user and a set of pages clicked by the user on the
and the occurring time of the query session. Then, we pro- search result. The reason we adopt the unit of query session
pose a new method KNN-GPCA to estimate the subspaces instead of query-page pair will be given in the below.
of transformed click-through data (e.g., the dotted lines in
Definition 1 (Query Session) A query session S = (Q, P ),
Figure 1 (b)). Each estimated subspace contains query ses-
where Q = {q1 , q2 , · · · , qm } is a bag of keywords repre-
sions of some broad topic2 . Next, by considering both the
senting a user query issued to a search engine and P =
distribution along the semantic dimension and the distribu-
{p1 , p2 , · · · , pn } refers to the set of corresponding pages
2 TDT has its own topic definition. However, in this paper, we refer to clicked by the user on the search result. In addition, a query
a topic as some general theme, which may or may not be related with an session S is associated with a time point, T (S), when the
event. query is issued.
90o 90o
e
im
r: t
:semantics
0o
(a) Polar Transformation (b) Subspace Estimation
Figure 1. Overview of DECK.
For example, the first two entries in the Table 1 will
be considered as a query session because they indi-
cate that after issuing the query “jojo”, the user 337
clicked the two pages “http://www.jojoonline.com” and
“http://www.azlyrics.com”. Since entries of a query ses-
sion are temporally close to each other, in this paper, we
take the timestamp of the first entry as the occurring time
of the query session for simplicity. The extraction of query
sessions will be explained in Section 5.
It can be noticed that a query session usually contains
multiple query-page pairs which share the same user query.
The reason we adopt the unit of query session rather than
query-page pair is based on the following intuitive: if a user
inquires about some event, the multiple query-page pairs
corresponding to the query session usually represent the
same event. Thus, instead of clustering these query-page
pairs afterwards to discover events, we can group them to-
gether and study in unit of query session.
Before illustrating how to transform query sessions into
2D polar space, we first explain the reason that we choose
polar space instead of Cartesian Coordinate System. Given
a 2D space, 1D subspaces are defined to be lines passing
through the origin. If we transform each query session S
to a point (θ, r) in polar space such that the angle θ and
the radius r respectively reflect the semantics and the oc-
curring time of S, then query sessions of similar seman-
tics should be mapped to points which have similar angles
and locate along a line (subspace) passing through the ori-
gin. However, if we transform a query session S to a point
(x, y) in Cartesian Coordinate System such that the x and y
values respectively reflect the semantics and the occurring
time of S, then queries sessions of similar semantics will
be mapped to points which have similar x values and locate
along a line parallel to the y axis. Lines parallel to the y
axis (except the y axis itself) are not subspaces and cannot
be detected by subspace estimation algorithms.
Particularly, given two query sessions S1 and S2 , we map
them to two points (θ1 , r1 ) and (θ2 , r2 ) respectively such
that the more similar the two query sessions are in seman-
tics, the smaller the angle |θ1 − θ2 | is. Furthermore, if the
two query sessions are exactly same in semantics, then the
90o 90o
0o 0o 0o
(c) Subspace Pruning (d) Cluster Generation
closer their occurring time, the smaller the distance |r1 −r2 |.
We first consider mapping the semantics of a query ses-
sion S to an angle θ. Since the angle between two query
sessions reflects how similar they are in semantics, we need
to define the semantic similarity between two query ses-
sions. Recall that a query session contains a query and a
set of clicked pages. We define the semantic similarity be-
tween two query sessions by considering their similarities
in not only query keywords but also clicked pages. For both
queries and clicked pages, we use an adapted Jaccard coef-
ficient to measure their similarity.
Definition 2 (Semantic Similarity) Given two query ses-
sions S1 = (Q1 , P1 ) and S2 = (Q2 , P2 ), the semantic
similarity between S1 and S2 , denoted as Sim(S1 , S2 ), is
|Q1 ∩ Q2 | |P1 ∩ P2 |
Sim(S1 , S2 ) = α× +(1−α)×
max{|Q1 |, |Q2 |} max{|P1 |, |P2 |}
The weight coefficient α ∈ [0, 1] can be determined ex-
perimentally to assign different importance to the simi-
larities of queries and clicked pages. For example, con-
sider the query sessions in Figure 2 (a), where each en-
try of the table represents a query session. Let α = 0.5.
Sim(S1 , S2 ) = 1/2 ∗ 1/2 + 1/2 ∗ 2/3 =7/12.
Given a set of n query sessions {S1 , S2 , · · · , Sn }, a n×n
semantic similarity matrix M can be computed such that
each element mij = Sim(Si , Sj ). In other words, the rel-
ative semantics of a query session Si is represented by the
n-dimension row vector Ri = hmi1 , mi2 , · · · , min i of M .
In order to map the semantics of Si to an angle θi in polar
space, we need to reduce the dimension of Ri to 1. For di-
mension reduction, we perform Principle Component Anal-
ysis (PCA) on the semantic similarity matrix M . Then, the
first principal component is used to preserve the dominant
variance in semantic similarities. Let {f1 , f2 , · · · , fn } be
the first principal component which corresponds to the set
of query sessions {S1 , S2 , · · · , Sn }. A query session Si can
be mapped to a point (θi , ri ) where θi is computed as
fi − minj (fj ) π
θi = ×
maxj (fj ) − minj (fj ) 2
where minj (fj ) and maxj (fj ) are the minimum and max- ID Query Pages Time
imum values respectively in the first principal component. S1 q1q2 p1p2 2006-03-15 14:08:16
Obviously, θi is restricted to [0, π/2]. S2 q2 p1p2p3 2006-03-15 21:20:09
The mapping from the occurring time of a query session S3 q2 p2p3 2006-03-16 10:42:25
to the radius of a point can be handled directly. Given a set
S4 q3q4 p4p5 2006-03-09 19:22:18
of query sessions {S1 , S2 , · · · , Sn }, the radius of the point
(θi , ri ) corresponding to the query session Si is given by S5 q4 p4p6p7 2006-03-18 11:35:52

S6 q3q4 p4p8 2006-03-29 12:05:31

T (Si ) − minj (T (Sj ))
ri = S7 q5 p9p10 2006-03-22 20:41:11
maxj (T (Sj )) − minj (T (Sj )) (a) Example query sessions
where minj (T (Sj )) and maxj (T (Sj )) are respectively the
90o
earliest and latest occurring time of all query sessions. ri S6
S7
takes value in the range of [0, 1].
Consider the set of example query sessions in Figure 2
S5
(a) again. Query sessions S1 to S3 are similar in semantics
as well as in occurring time. Query sessions S4 through S6
are similar in semantics only. Query session S7 is dissim- S4
S3
ilar to any other query sessions in semantics. Figure 2 (b) S1

shows the polar transformation of the set of query sessions. S2

(In order to show the polar transformation clearly, we con- 0o

straint the angles of query sessions to [π/9, 4π/9] in this
example.) It can be observed from the figure that our polar
transformation has the following two features:
• Subspace consistency. The mapping from the seman-
tics of a query session to the angle of a point in polar
space causes query sessions of similar semantics to lie
on one and only one 1D subspace. For example, points
of S1 to S3 and points of S4 to S6 locate around the
two dotted lines in Figure 2 respectively. However, the
point of S7 appears as an outlier in the figure.
• Cluster consistency. The mapping from the occurring
time of a query session to the radius of a point in po-
lar space forces query sessions of similar semantics
and similar occurring time to appear as clusters in sub-
spaces. For example, points of S1 to S3 form a cluster
in the lower dotted line, while points of S4 to S6 dis-
tribute along the upper dotted line.
3 Subspace Estimation
According to the subspace consistency of polar transfor-
mation, our objective now is to estimate subspaces from
the transformed data such that each subspace contains
query sessions of similar semantics. For this purpose,
we propose a new subspace estimation algorithm KNN-
GPCA, based on Generalized Principal Component Anal-
ysis (GPCA) [16]. GPCA is an algebra-geometric approach
which simultaneously estimates subspace bases and assigns
data points to subspaces. The reasons why we develop our
algorithm based on GPCA are as follows: unlike prior work
on subspace analysis [5, 6], GPCA does not require initial-
ization, which usually results in local optimum, or restrict
the subspaces to be either orthogonal or trivially intersect-
ing. Furthermore, GPCA is capable of estimating subspaces
without prior knowledge on the number of subspaces.
(b) Example polar transformation
Figure 2. Example polar transformation.
However, as analyzed in the following subsection, the
performance of GPCA degrades in the presence of outliers.
In order to improve the robustness of GPCA, KNN-GPCA
takes into account the distribution of K nearest neighbors
of data points, which yields a weighted least square estima-
tion of subspaces. In this section, we first review the al-
gorithm of GPCA. Then, the embedding of the distribution
of K nearest neighbors is described. Finally, we estimate
subspaces using weighted least square technique.
3.1 Review of GPCA
Given a set of sample data points, GPCA estimates a
mixture of linear subspaces by the following three steps:
fitting polynomials to data, estimating the number of sub-
spaces and estimating the normal vectors of subspaces.
Based on the fact that each data point x ∈ RD satisfies
T
bi x = 0 where bi is the normal vector of the subspace to
which x belongs, a data point lying on one of the n sub-
spaces satisfies the following equation:
Yn
pn (x) = (bTi x) = 0 (1)
i=1
where {bi }ni=1 are normal vectors of the subspaces. It
can be converted to a linear expression by expanding the
product of all bTi x and viewing all monomials xn =
xn1 1 xn2 2 · · · xnDD (0 ≤ nj ≤ n, j = 1 · · · D, n1 + n2 +
· · · + nD = n) as system unknowns. By introducing the
Veronese map vn : [x1 · · · xD ]T 7→ [· · · , xn , · · ·] where
xn is a monomial of the form xn1 1 xn2 2 · · · xnDD , equation (1)
becomes the following linear X expression:
pn (x) = vn (x)T c = cn1 ···nD xn1 1 · · · xnDD = 0 (2)
where c ∈ R represents the coefficient of the monomial xn .
Since each polynomial pn (x) = vn (x)T c must be satisfied
by every data point, given a collection of N sample data
points {xj }N
j=1 , a linear system can be generated as
 
vn (x1 )T
 vn (x2 )T 
 
 · 
Ln c =   c = 0 (3)
· 
 
 · 
vn (xN )T
After fitting polynomials to data, the number of sub-
spaces is estimated based on the condition that there is one
unique solution for c. Thus, when there exists no noisy data,
the number of subspaces can be computed as the minimum
value i such that the rank of the polynomial
µ embedding ¶ ma-
i+D−1
trix Li equals to Mi − 1, where Mi = is
D−1
the number of different monomials in Li . In the presence of
noise, GPCA relies on some pre-defined threshold ² to com-
bine subspaces that are close to each other. For example, the
σr+1
rank of Li is determined as r if σ1 +···+σ r
< ², where σj is
the j-th singular value of Li .
Once the number of subspaces is determined, the vec-
tor of coefficient c can be computed accordingly. Then, the
normal vectors {bi }ni=1 can be solved by polynomial differ-
entiation in the absence of noise. Taking noise into account,
the estimation of normal vectors is cast as a constrained
nonlinear optimization problem which is initialized using
the normal vectors obtained by polynomial differentiation.
While GPCA provides an elegant solution to the prob-
lem of linear subspace estimation, there are some inherent
limitations of the algorithm. Particularly, we observed that
when outliers are present, accuracy of the following steps
of GPCA will be impaired:
• Estimation of the number of subspaces. In the pres-
ence of noise, GPCA determines the rank r of Li
based on some threshold ². When noise and outliers
are moderate, the difference between the r-th singular
value and the (r + 1)-th singular value of Li might be
slight. Thus, the appropriate threshold is hard to spec-
ify and the number of subspaces can be determined er-
roneously.
• Estimation of the normal vectors {bi }ni=1 . GPCA esti-
mates coefficients c before estimating the normal vec-
tors of subspaces. In the presence of noise, c is com-
puted as the singular vectors of Li associated with its
smallest singular values. However, when outliers are
included in the calculation, the computed singular val-
ues and vectors are prone to error. Consequently, the
coefficients and the normal vectors may not be deter-
mined appropriately.
Therefore, in order to improve the robustness of GPCA,
the impact of noise and outliers in the two estimation steps
90o
x3
x2
x1
nvar
svar
0o
Figure 3. Distribution of 3 nearest neighbors.
should be demoted. We achieve this purpose by assigning
weight coefficients to data points to distinguish true data
from noise and outliers. Particularly, we weigh data points
based on the distribution of its K nearest neighbors.
3.2 Weight Coefficient Assignment
The K th Nearest Neighborhood Distance (kN N D)
metric proposed in [7] detects and removes outliers based
on the following fact: in a cluster containing more than K
points, the kN N D for a data point is small; otherwise, it
is large. Recall that our polar transformation has the cluster
consistency property, which indicates that true data points
(e.g., query sessions corresponding to real events) lie in
clusters inside subspaces. The kN N D of true data points
should be small while the kN N D of noise and outliers
should be large.
However, instead of simply using the kN N D to differ-
entiate outliers and inliers, we consider the distribution of
the K nearest neighbors of data points. Given a data point
xi , let its K nearest neighbors be N Nk (xi ). Both the vari-
ance of the K nearest neighbors along the direction of the
subspace of xi (i.e., the direction from the origin to the point
xi ), denoted as svar(N Nk (xi )), and the variance of the
K nearest neighbors along the direction which is orthogo-
nal to the subspace direction, denoted as nvar(N Nk (xi )),
can be computed by PCA. For example, Figure 3 shows
the distribution of 3 nearest neighbors of three data points
x1 , x2 and x3 respectively. The two directions along which
svar(N N3 (x1 )) and nvar(N N3 (x1 )) are computed are in-
dicated by svar and nvar respectively in the figure.
If xi is a true data point, it forms some cluster together
with its K nearest neighbors. Thus, both svar(N Nk (xi ))
and nvar(N Nk (xi )) should be small. Consequently, the
sum S(N Nk (xi )) = svar(N Nk (xi )) + nvar(N Nk (xi ))
will be small as well. On the contrary, if xi is an outlier
(e.g., x2 in Figure 3), S(N Nk (xi )) will be large. How-
ever, even if a data point forms a cluster together with its
K nearest neighbors, it may not be a true data point if
its neighbors spread along the orthogonal direction of the
subspace (e.g., x3 in Figure 3). That is, the cluster does
     
not exist inside some subspace (Note that, directly using vn (x1 )(2..M )T c2 −W (x1 )vn (x1 )(1)
kN N D, this type of noisy data points cannot be identified).  vn (x2 )(2..M )T  c3   −W (x2 )vn (x2 )(1) 
    
In this case, svar(N Nk (xi )) is small but nvar(N Nk (xi )) W  · 
 · =
  · 

is large. Then, the ratio R(N Nk (xi )) =nvar( N Nk (xi ))/  ·  ·   · 
svar(N Nk (xi )) is large. Hence, we assign a weight, vn (xN )(2..M )T cM −W (xN )vn (xN )(1)
W (xi ), to a data point xi based on S(N Nk (xi )) and (7)
R(N Nk (xi )) as follows, The above equation can be succinctly written as W Ac =
1 d, where A is the matrix whose rows are vn (xi )(2..M )T ,
W (xi ) = (4) i = 1, 2..N and d is the right side of the above equation.
1 + S(N Nk (xi )) × R(N Nk (xi ))
By minimizing the objective function kd − AckW , we can
The value of W (xi ) ranges from 0 to 1. When the data obtain the weighted least square approximation of ci , i =
point xi lies in a cluster where data points spread along 1, 2..N as
the direction of the subspace, the value of R(N Nk (xi )) c1 = 1 and [c2 , · · · , cM ]T = (AT W T W A)−1 (AT W T W d)
is 0 in the absence of noise and is very small in the pres- (8)
ence of noise. The value of S(N Nk (xi )) is small as well. Note that, since we use the diagonal matrix of weight co-
Hence, the weight W (xi ) is close to 1. In other words, efficient W to demote the impact of noise and outliers, the
we assign a large weight to a true data point. Otherwise, estimation error of coefficient vector c is reduced.
R(N Nk (xi )) and/or S(N Nk (xi )) are large, which results To estimate the normal vectors {bi }ni=1 , we calculate
in a small W (xi ). Thus, the impact of noise and outliers them in the absence of noise using polynomial differen-
will be reduced by small weight values. The decision of the tiation as in original GPCA. The computed vectors serve
parameter K will be discussed in the below. to initialize the following constraint nonlinear optimization
3.3 KNN-GPCA which differs from GPCA in the introduction of the weight
Taking into account the weight of each data point, the coefficients: N
X
linear system of equation (3) is modified as min W (xj )kx̃j − xj k2
 W Ln c =   j=1
W (x1 ) vn (x1 )T Yn
 W (x2 )   vn (x2 )T  subject to (bTi x̃j ) = 0 j = 1, · · · , N (9)
  
 ·  · c = 0 i=1
   where x̃j is the projection of xj onto its closet subspace.
 ·   · 
W (xN ) vn (xN ) T By using Lagrange multipliers λj for each constraint, the
(5) above optimization problem is equivalent to minimizing the
where W (xi ) is the weight of data point xi . We then per- following function
XN Yn
form a Singular Value Decomposition (SVD) on W Ln to
(W (xj )kx̃j − xj k2 + λj (bTi )) (10)
estimate the number of subspaces. Note that, if xi is an out-
j=1 i=1
lier, its small weight will pull it back to the origin. Since Taking partial derivatives with respect to x̃j and equating it
origin exists in all null space, xi will not increase singular to 0, we can solve for λj /2 and W (xj )kx̃j − xj k2 . By re-
values. Consequently, the selection of the threshold ² here placing them into the objective function (9), the simplified
will not be as crucial as before. objective function on the normal vectors can be derived as
After estimating the number of subspaces, the coefficient XN Qn
W (xj )(n i=1 bTi xj )2
vector c can be computed using the weighted least square En (b1 , · · · , bn ) = Pn Q (11)
technique. Particularly, we express the left side of equation j=1
k i=1 bi l6=i (bTl xj )k2
(5) 
as After obtaining an initial estimate of normal vectors using
   
vn (x1 )(2..M )T c2 W (x1 )vn (x1 )(1)c1 polynomial differentiation, one can use standard nonlinear
 vn (x2 )(2..M )T   c3   W (x2 )vn (x2 )(1)c1  optimization techniques to optimize equation (11).
    
W  ·   · +
   ·  We now discuss some implementation and parameter is-

 ·  ·   ·  sues in this subspace estimation step. In order to compute
vn (xN )(2..M )T cM W (xN )vn (xN )(1)c1 the weight of each data point, we need find its K nearest
(6) neighbors. Due to the high complexity of directly discov-
where vn (xi )(2..M ) is a vector containing all elements ering K nearest neighbors, in our implementation, we op-
of vn (xi ) except
µ ¶ for the first element vn (xi )(1). M = timize to perform a bisecting k-means clustering first and
r+D−1 discover K nearest neighbors in each cluster only. In esti-
where r is the estimated subspace num-
D−1 mating the set of normal vectors {bi }ni=1 , we observed that
ber. In order to calculate a basis of coefficient vector, let the convergence of equation (11) is slow. Hence, in our
c1 = 1. Then, the equation (5) can be expressed as implementation, we employed a weighted k-means iteration
method to estimate the optimal vectors. We assign weighted
data points to their nearest subspaces and update the normal
vectors of subspaces. This process is performed iteratively
till there is no change to the subspaces. According to our
experimental results, this method achieves the same perfor-
mance of equation (11) but with a faster convergence rate.
There are two parameters involved in this step: K, which
is the number of nearest neighbors, and ², which is the
threshold used to decide the rank of the polynomial embed-
ding matrix. The former can be decided easily. For exam-
ple, since noise and outliers are usually far away from true
data points, any small value of K can be used. The selec-
tion of K can be fixed for the computation of weights for
all data points. The decision of the parameter ² depends on
the data. As shown in our experiments (Section 5.2), by
demoting the impact of outliers and noises, the difference
between the rth and (r + 1)th singular values of the polyno-
mial embedding matrix is enlarged and observable, which
makes the decision of ² straightforward.
4 Event Detection
After estimating subspaces from the transformed polar
space, each subspace contains query sessions of similar
topic. However, not every subspace is interesting such that
it contains clusters corresponding to real events. For ex-
ample, consider the query sessions with queries on popular
public portals (e.g., google). Due to their intense frequency
during the whole observed time period, such query sessions
will form a subspace although they do not represent any real
event. For example, the subspace s1 in Figure 4 represents
such an uninteresting subspace. In this section, we first dis-
cuss pruning uninteresting subspaces. Then, various events
are detected from interesting subspaces.
4.1 Subspace pruning
We distinguish interesting and uninteresting subspaces
based on the intuitive premise used by [8]: the appear-
ance of an event is signaled by a “burst of activity”, with
certain features rising sharply in frequency when the event
emerges. In our work, we simultaneously consider two fea-
tures: the occurring time of query sessions as well as the
semantics of query sessions. Particularly, if a subspace is
interesting such that it contains query sessions correspond-
ing to events, both the occurring time and the semantics
of query sessions in the subspace should exhibit certain
“bursts”. Recall that our polar transformation respectively
maps the occurring time and semantics of a query session
to the radius and angle of a point in polar space. The tem-
poral “burst” and semantical “burst” should be reflected by
the certain distribution of data points along the subspace
direction and the orthogonal direction of the subspace re-
spectively. For example, as shown in Figure 4, the subspace
s2 might be an interesting subspace since its data points
distribute certainly along the subspace direction and the or-
thogonal direction of the subspace.
90o
s1
vn
s2
v1
hm
h1
0o
Figure 4. Interestingness of subspaces.
Note that, we cannot simply use the variance measure to
define the interestingness of a subspace. The reason is as
follows. A subspace may contain multiple events (e.g., pe-
riodical events). If there are multiple events contained in a
subspace, then the variance of data points in the orthogo-
nal direction (e.g., the “semantic” dimension) will be small,
while the variance of data points in the subspace direction
(e.g., the “temporal” dimension) will be large. Hence, in
our approach, we employ the entropy measure to define the
interestingness of a subspace. Particularly, we project data
points to the two directions respectively and calculate the
respective histograms of the distributions. Let hh1 , h2 , · · · ,
hm i and hv1 , v2 , · · · , vn i, where hi and vi are individual
bins, be the two corresponding histograms. The interesting-
ness of a subspace can be computed as follows.
Xm n
X
I(si ) = 1 − [−p hi log hi − (1 − p) vi log vi ] (12)
i=1 i=1
where p ∈ [0, 1] is a weight which can be determined in
experiments to assign different importance to the entropy
values in the two directions. For example, if p = 1, then
only the temporal “burst” is considered. The interesting-
ness measure takes values from 0 to 1. The more certain
the distributions in two directions, the smaller the entropies
in the brackets of Equation (12), the greater the value of
interestingness. Given some threshold ζ, subspace si will
be pruned as an uninteresting subspace if I(si ) < ζ. We
observed from our experimental results (Section 5.2) that
I(si ) of an interesting subspace si is usually much greater
than I(sj ) of an uninteresting subspace sj , which makes it
easy to decide the value of ζ.
Note that, the calculation of the histograms may be bi-
ased by noisy data points. In our implementation, we re-
move noise by an inlier growing method before projecting
data points. The details are given in a full version of this
paper (the reference is omitted temporarily to comply with
the double blind reviewing guidelines).
Algorithm 1 DECK.
Input: A set of query sessions
Output: A set of query session clusters corresponding to real events
Description:
1: Transform query sessions to polar space
2: Estimate subspaces of query sessions using KNN-GPCA
3: for Each estimated subspace do
4: Project inliers to the subspace direction and orthogonal direction
respectively
5: Compute the entropy of the distribution histogram in the two di-
rections
6: if Interestingness of the subspace is less than some threshold then
7: Prune the subspace
8: end if
9: end for
10: for Each interesting subspace do
11: Perform mean shift clustering
12: end for
13: Return the clustering results
4.2 Query session clustering
After pruning uninteresting subspaces, events can be
detected from the remaining subspaces by clustering data
points based on the cluster consistency property of polar
transformation. Although there are a plethora of published
clustering techniques, methods relying upon a priori knowl-
edge of the number of clusters are not applicable here be-
cause the number of events in each subspace cannot be de-
cided easily. Hence, in our approach, we employed a non-
parametric clustering technique called Mean Shift Cluster-
ing [4]. Mean shift clustering is an application of the mean
shift procedure, which successively computes the mean shift
vector which always points toward the direction of the max-
imum increase in the density and converges to a point where
the gradient of density function is zero.
Based on the mean shift procedure, we perform mean
shift clustering on data points in each subspace as follows.
Firstly, mean shift procedure is run with all the data points
to find the stationary points of the density estimate. Sec-
ondly, discovered stationary points are pruned such that
only local maxima are retained. The set of all points that
converge to the same mode defined the same cluster. The
returned clusters are expected to represent real events. The
complete algorithm of DECK is shown in Algorithm 1.
5 Performance Study
In this section, we study the performance of DECK. We
first describe the data set used in our experiments. Then, we
present and analyze the results of conducted experiments.
5.1 Data Set
The real-life Web click-through data collected by
AOL [13] from March 2006 through May 2006 are used.
According to [13], if a user clicked on more than one page
from the list returned from a single query, these pages ap-
pear as successive entries in the data. Hence, in our exper-
iments, we simply extract successive pages corresponding
to the same query and the same user as a query session.
The timestamp of the first clicked page of a query session
is taken as the occurring time of the session. We manu-
ally identified a set of events from the data set, including
both predictable events (e.g., the Memorial Day on May
29, 2006) and unpredictable events (e.g., the death of Dana
Reeve, an American actress, on Mar 6, 2006). For each
event, we identified a set of query keywords and selected
all query sessions which contain the query keywords and
happen close to the date of the event. After filtering events
which are represented by less than 50 query sessions, a total
of 35 events are used in our experiments. The complete list
of events is given in the full version of this paper. We then
randomly select query sessions which do not represent any
real events, together with the query sessions corresponding
to real events, to generate five data sets, which respectively
contain 5K, 10K, 20K, 50K and 100K query sessions.
5.2 Result Analysis
Performance of DECK. We first compare the perfor-
mance of DECK with the existing two-phase-clustering al-
gorithm [18]. Given the set of clusters returned as detected
events, the existing algorithm [18] finds a best match be-
tween discovered clusters and true events. A best match of a
true event is defined as a cluster that has the maximum over-
lap with the true event in terms of the number of common
query-page pairs (query sessions). We further constrain that
the number of common query-page pairs (query sessions)
should be no less than some specified threshold (in our ex-
periments, we set the threshold as 50% of the query-page
pairs or query sessions representing the true event). Then,
the evaluation metrics, precision and recall, can be com-
puted as follows. Precision is the ratio of the number of
correctly detected events to the overall discovered clusters.
Recall is the ratio of the number of correctly detected events
to the total number of events.
The experimental results are shown in Figures 5 (a) and
(b) respectively. The existing algorithm is referred to as
2PClustering in the figures (please ignore the other two al-
gorithms, DECK-GPCA and DECK-NP, temporarily). We
observe that DECK outperforms 2PClustering in both pre-
cision and recall. Since the algorithm 2PClustering did not
discuss how to decide the number of clusters, we use the
number of events generated by DECK. Then, since many
clusters generated by 2PClustering do not represent any
events, both the precision and recall values are low.
Since the best match approach used by the existing al-
gorithm may not be objective enough, we further evaluate
the performance using an entropy measure. For each gen-
erated cluster i, we compute pij as the fraction of query-
page pairs (query sessions) representing thePtrue event j.
Then, the entropy the of cluster i is Ei = − j pij log pij .
The total entropy can be calculated as the sum of the en-
tropiesPof each cluster weighted by the size of each cluster:
m
E = i ni ×E n , where m is the number of clusters, n is
i
 0.9 DECK 2P Clustering 1 DECK 2P Clustering 0.3

0.8 DECK-GP CA DECK-NP 0.9 DECK-GP CA DECK-NP DECK

0.25 2PClust ering
0.7 0.8
DECK-GPCA
0.7
0.6 0.2 DECK-NP
Precision

0.6

Entropy
Recall
0.5
0.5 0.15
0.4
0.4
0.3 0.1
0.3
0.2 0.2
0.05
0.1 0.1
0 0 0
5K 10K 20K 50K 100K 5K 10K 20K 50K 100K 5K 10K 20K 50K 100K

No. of Query Sessions No. of Query Sessions No. of Query Sessions

(a) (b) (c)

Figure 5. Precision, recall and entropy of DECK, 2PClustering, DECK-GPCA and DECK-NP.
total number of query-page pairs (query sessions) and ni is
the size of cluster i. The experimental results are shown in
Figure 5 (c). Again, DECK works better than 2PCluster-
ing. The reason is similar as before. Given the number of
clusters generated by DECK, 2PClustering generates clus-
ters with larger size. Since it does not prune any data, the
entropy of 2PClustering is higher than DECK.
3
10event
Singular Value Gap Ratio
σr+1
σ1 +···+σr < ². They recursively increase the number of
subspaces until the condition is satisfied. Hence, in our ex-
periments, the value r is automatically decided by fixing the
value of ² as 1.0E-3 for both GPCA and KNN-GPCA. Then,
the ratios of the gap computed by KNN-GPCA to the gap
computed by GPCA are shown in Figure 6. We notice that
KNN-GPCA does enlarge the gap and improves the possi-
bility of selecting appropriate ².
20event 35event

2.5
6
10event 20event 35event
2
5
Interestingness Ratio

1.5
4

1
3

0.5
2

0
1
5K 10K 20K 50K 100K

No. of Query Sessions 0

5K 10K 20K 50K 100K

No. of Query Sessions

Figure 6. Performance of subspace estima-
tion.
Figure 7. Performance of subspace pruning.
Performance of Subspace Estimation. We also study Performance of Subspace Pruning. In order to exam-
the performance of the subspace estimation step of DECK. ine the subspace pruning step of DECK, we implemented
Particularly, we examine the effectiveness of KNN-GPCA. another alternative version of DECK, referred to as DECK-
Hence, we implemented an alternative version of DECK, NP (stands for DECK with No Pruning), which skips the
referred to as DECK-GPCA, which employs the original subspace pruning step. The performance of DECK-NP is
GPCA algorithm to estimate subspaces. As shown in Fig- shown in Figures 5 (a), (b) and (c) also. Since DECK-
ures 5 (a), (b) and (c), DECK-GPCA performs worse than NP does not prune any uninteresting subspaces, more clus-
DECK. It indicates that KNN-GPCA effectively demotes ters will be generated by this algorithm. Hence, although it
the impact of outliers and improves the robustness of sub- achieves similar recall as DECK, its precision is even lower
space estimation. than 2PClustering. Since it employs KNN-GPCA to esti-
As analyzed in Section 3, after assigning lower weights mate subspaces, the correctly discovered clusters are of high
to outliers, the selection of the parameter ², which deter- quality. Hence, its entropy is better than DECK-GPCA.
mines the number of subspaces, can be decided more easily. We further conduct experiments to examine whether the
Hence, we further investigate how KNN-GPCA enlarges threshold ζ, which is used to prune uninteresting subspaces,
the gap between the rth and (r + 1)th singular values of can be selected easily. The experiments are conducted on
the polynomial embedding matrix. We conduct the exper- the five datasets by varying the number of events. For each
iments on the five datasets by varying the number of true dataset, we order the estimated subspaces according to their
events. Note that, in the presence of noise, both GPCA and interestingness values. We then compute the interesting-
KNN-GPCA decide the number of subspaces by assuming ness ratio between the pair of successive subspaces with the
there are two subspaces first. Then, they justify whether greatest difference in their interestingness values. The ex-
perimental results are shown in Figure 7, where the high
ratios indicate that appropriate ζ values can be decided with
ease.
6 Related Work
The problem of event detection is part of a broader ini-
tiative called Topic Detection and Tracking [2]. Particu-
larly, event detection can be divided into two categories:
retrospective detection and on-line detection [17]. The for-
mer refers to the detection of previously unidentified events
from accumulated historical collection, while the latter en-
tails the discovery of the onset of new events from live feeds
in real-time. Our approach detects events from a collection
of Web click-through data and belongs to the first category.
We thus review some related retrospective event detection
algorithms. In [17], Yang et al. proposed to use an ag-
glomerative clustering algorithm, augmented Group Aver-
age Clustering, to discover events from the corpus. As a
recent example, Li et al. [11] proposed a multi-model ap-
proach which models both content and time information of
documents explicitly. The particular feature of their algo-
rithm is the usage of an auto-adaptive sliding windows on
time line, which overcomes the inflexible usages of times-
tamps in traditional retrospective detection algorithms. The
most significant difference between our work and the exist-
ing work is caused by the different data source, Web click-
through data, we considered. Direct clustering is not ap-
plicable for our problem because Web click-through data is
not necessarily correlated with events. Time window is no
longer a critical issue here as temporal burst can be observed
obviously based on a window size of a day.
Recently, there has been significant research interest in
detecting events from text stream using feature-pivot ap-
proaches. This line of research is inspired by Kleinberg’s
seminal work that describes extracting bursty features using
an infinite automaton model [8]. Fung et al. [21] proposed
to identify bursty features and then cluster features to gen-
erate bursty events. Their approach has the restriction that
each feature exclusively belongs to one event. However,
in our problem, a bursty query keyword (or clicked page)
probably belongs to different events. The work presented
by He et al. [22] also detects events by examining features
first. They analyzed every feature using Discrete Fourier
Transformation and classified features to categories which
correspond to different types of events (e.g., periodic and
aperiodic events). Their approach cannot be used directly
in our problem as the semantic information of click-through
data cannot be captured and the irrelevant data cannot be
pruned.
7 Conclusions
Web click-through data was recently identified as a po-
tential source for event detection. In this paper, we propose
a novel approach, called DECK, for detecting events from
click-through data. The main features of DECK are sum-
marized as follows. Firstly, DECK simultaneously consid-
ers the semantic information and temporal information of
click-through data. By transforming click-through data to
points in 2D polar space such that the angles and radiuses
of points respectively reflect the semantics and timestamps
of queries, information of the two dimensions are consid-
ered at the same time in both the subspace estimation step
(e.g., each data point is weighted based on the distributions
of its K nearest neighbors in both the subspace direction
and the orthogonal direction) and the subspace pruning step
(e.g., entropy in both the subspace direction and the orthog-
onal direction are used to measure the interestingness of a
subspace). Secondly, in the subspace estimation step, data
points are weighed based on the distribution of its K near-
est neighbors instead of kNND, so that a particular type of
outliers can be identified. Thirdly, in the subspace pruning
step, an interestingness measure based on entropy instead of
variance was proposed, so that subspaces containing multi-
ple events can be identified as interesting. The experimen-
tal results on real-life Web click-through data show that our
approach is accurate and effective in detecting events from
Web click-through data.
References
[1] J. Allan, J. G. Carbonell, G. Doddington, J. Yamron, and Y. Yang. Topic
detection and tracking pilot study: Final report. In DARPA Broadcast News
Transcription and Understanding Workshop, 1998.
[2] J. Allan, R. Papka, and V. Lavrenko. On-line new event detection and track-
ing. In SIGIR, 1998.
[3] J. Allan, C. Wade, and A. Bolivar. Retrieval and novelty detection at the
sentence level. In SIGIR, 2003.
[4] D. Comaniciu and P. Meer. Mean shift: A robust approach toward feature
space analysis. In IEEE Transaction On Pattern Analysis And Machine Intel-
ligence, volume 24, 2002.
[5] J. Ho, M.-H. Yang, J. Lim, K.-C. Lee, and D. J. Kriegman. Clustering appear-
ances of objects under varying illumination conditions. In CVPR (1), 2003.
[6] K. ichi Kanatani. Motion segmentation by subspace separation and model
selection. In ICCV, 2001.
[7] Q. Ke and T. Kanade. Robust subspace clustering by combined use of knnd
metric and svd algorithm. In CVPR, 2004.
[8] J. M. Kleinberg. Bursty and hierarchical structure in streams. In KDD, 2002.
[9] R. Kumar, U. Mahadevan, and D. Sivakumar. A graph-theoretic approach to
extract storylines from search results. In KDD, 2004.
[10] W.-S. Li, K. S. Candan, Q. Vu, and D. Agrawal. Retrieving and organizing
web pages by “information unit”. In WWW, 2001.
[11] Z. Li, B. Wang, M. Li, and W.-Y. Ma. A probabilistic model for retrospective
news event detection. In SIGIR, 2005.
[12] Q. Mei and C. Zhai. Discovering evolutionary theme patterns from text: an
exploration of temporal text mining. In KDD, 2005.
[13] G. Pass, A. Chowdhury, and C. Torgeson. A picture of search. In The First
International Conference on Scalable Information Systems, 2006.
[14] J. Shi and J. Malik. Normalized cuts and image segmentation. IEEE Trans.
Pattern Anal. Mach. Intell., 22(8), 2000.
[15] A. Sun and E.-P. Lim. Web unit mining: finding and classifying subgraphs of
web pages. In CIKM, 2003.
[16] R. Vidal, Y. Ma, and S. Sastry. Generalized principal component analysis. In
CVPR, 2003.
[17] Y. Yang, T. Pierce, and J. G. Carbonell. A study of retrospective and on-line
event detection. In SIGIR, 1998.
[18] Q. Zhao, T.-Y. Liu, S. S. Bhowmick, and W.-Y. Ma. Event detection from
evolution of click-through data. In KDD, 2006.
[19] Q. Zhao, P. Mitra, B. Chen Temporal and Information Flow Based Event
Detection from Social Text Streams. In AAAI, 2007.
[20] T. Rattenbury, N. Good, M. Naaman Towards automatic extraction of event
and place semantics from flickr tags. In SIGIR, 2007.
[21] G.P. Fung, J.X. Yu, P. S. Yu, H. Lu Parameter Free Bursty Events Detection
in Text Streams. In VLDB, 2005.
[22] Q. He, K. Chang, E.-P. Lim Analyzing feature trajectories for event detection.
In SIGIR, 2007.