VizClick

Visualizing Clickstream Data

Rajat Kateja

1,3,∗

, Amerineni Rohith

1,3,∗

, Piyush Kumar

2,3∗

and Ritwik Sinha

Indian Institute of Technology Guwahati, Guwahati, India

Indian Institute of Technology Bombay, Mumbai, India

Adobe Research, Bangalore, India

Keywords:

Clickstream, Visualization, Association, Geo-spatial, Site Topography, Big Data.

Abstract:

Clickstream data is ubiquitous in today’s web-connected world. Such data displays the salient features of

big data, that is, volume, velocity and variety. As with any big data, visualizations can play a central role in

making sense and generating hypotheses from such data. In this paper, we present a systematic approach of

visualizing clickstream data. There are three basic questions we aim to address. First, we explore the inter-

dependence between the large number of dimensions that are measured in clickstream data. Next, we analyze

spatial aspects of data collected in web-analytics. Finally, the web designers might be interested in getting a

deeper understanding of the website’s topography and how browsers are interacting with it. Our approach is

designed for business analysts, web designers and marketers; and helps them draw actionable insights in the

management and reﬁnement of large websites.

1 INTRODUCTION

A clickstream is deﬁned as the series of mouse clicks

made by the user of a website. For websites, click-

streams serve as a source of highly valuable informa-

tion. In particular, it tells the website owners about

their users, and what they are interacting with. This

information helps in aiding marketing decisions and

help in providing a personalized experience. While

the information contained in clickstream data is be-

yond dispute, it comes with many challenges. These

challenges stem mostly from the fact that clickstream

data displays all aspects of big data (McAfee et al.,

2012). Namely, clickstream data is high in vol-

ume and velocity, with one day’s worth of such data

amount to many tens or hundreds of gigabytes for any

major website. Further, clickstream data is large in

variety and is inherently heterogeneous (Wei et al.,

2012).

Information visualization may be deﬁned as a

visual representation of data that helps in generating

hypotheses and making inferences. The history

of data visualization goes a long way back (Tufte

∗

First three authors have equal contribution. We would

also like to acknowledge Sandeep Zechariah George K (sge-

orge@adobe.com) for his continuous support during the

project and especially for his guidance with D3.

and Graves-Morris, 1983). The second half of the

century saw the formal development of the

grammar of graphics (Cleveland, 1993). The last two

decades have seen an explosion in the development of

platforms for the development of professional quality

information visualizations, R (R Core Team, 2013)

being one of the more popular platforms. While R

has signiﬁcant capabilities to generate high quality

graphics (Sarkar, 2008; Wickham, 2009); there are

limitations in creating highly interactive visualiza-

tions. Interactive visualizations are important for

big data because they give the user the ability to

hierarchically explore the data. There have been some

effort to create platforms for interactive visualizations

(Swayne et al., 2003), but, much of these have been

overshadowed lately by JavaScript based graphing

libraries

. We see two advantages of these, ﬁrst,

one can have customizable levels of interactivity, and

second, they can be viewed on a browser, making

them easier to share. Among these libraries, D3

(Bostock et al., 2011) has received considerable

interest and as a result is a mature framework for

interactive visualizations. Much of the visualiza-

tions created in this paper use and build on D3.

http://socialcompare.com/en/comparison/

javascript-graphs-and-charts-libraries

247

Kateja R., Rohith A., Kumar P. and Sinha R..

VizClick - Visualizing Clickstream Data.

DOI: 10.5220/0004687102470255

In Proceedings of the 5th International Conference on Information Visualization Theory and Applications (IVAPP-2014), pages 247-255

ISBN: 978-989-758-005-5

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

Data

Select features

and apply

filters

Sort data into

sessions

Session level

summarization

Create

custom

parameters

Topographic Visualization

Calculate

associations

Aggregate to

state level data

Selection

Summarization

Association Visualization

Geographic Visualization

Visualizations

Figure 1: A schematic representation of the work ﬂow. We

start with ﬁltering the raw data, which is then sorted and

summarized into sessions. This sorted data is used to create

transition matrices for visualizing the site topography. The

session level data is further processed to create visualiza-

tions of association and at the geographic level. Finally, the

entire solution was packaged as a website.

In this paper, we present a framework for answer-

ing important questions from clickstream data using

visualizations. Adobe



Analytics

provides mar-

keters the most comprehensive set of tools to mea-

sure and track all aspects of usage of an organiza-

tion’s website. We built and tested our framework on

the clickstream data of www.adobe.com collected us-

ing Adobe



Analytics. We deal with three aspects of

clickstream data.

First, as is evident, many parameters of interest

to the marketer are collected in clickstream data. For

example, referring domain, time spent on the website,

number of clicks in a single session, visits to different

site sections, and events on the website (like purchase

or downloads). Some of these large number of param-

eters may be associated with each other, the presence

or absence of such associations is informative to the

marketer. Given the large size of data and the number

of features, these associations are not apparent. In this

paper, we evaluated a number of approaches to mea-

sure association between two features in the context

of clickstream data. Given a measure of association,

an equally challenging aspect is to present these pair-

wise associations among a large set of parameters in

a way that enables ﬁnding patterns.

Next, many website owners (particularly e-

commerce websites) are interested in analyzing geo-

spatial aspects of the clickstream data. To address

questions like, “In which regions are our products in

demand?”; “Where are we doing poorly?”; “If we had

some marketing budget, what regions should we con-

centrate on?” We analyzed clickstreams and summa-

http://www.adobe.com/in/solutions/digital-

analytics/marketing-reports-analytics.html

rized it at some geographical level. We used a con-

tinuous area Cartogram (Dougenik et al., 1985) as a

means to present this data and enable information ex-

traction. We next presented geographic data in a man-

ner that helps detect deviations from expectation, thus

aiding anomaly detection.

Finally, we contrast the site design with how users

interact with it. The web designer has a notion of

relationships between different content on the web-

site, however, users may have a very different view of

these relationships. The clickstream data tell us how

users see the structure of the website. Such compar-

isons provide meaningful insights to web designers

on how to perform minor or major reorganizations of

their website. There have been several studies on site

topography (Lee et al., 2001; Brainerd and Becker,

2001; Ferreira de Oliveira and Levkowitz, 2003), but

they do not propose interactive visualizations and do

not allow for analysis at higher levels of granularity

(up to the level of individual pages). We created vi-

sualizations, which incorporate custom clustering, so

that the website owner can focus on relevant areas and

make deductions.

The rest of the paper is organized as follows. The

second section explains the data used as well as the

pre-processing necessary. It also describes the meth-

ods devised to address the three broad goals of the

paper. The third section is dedicated to results which

show how making inferences is made easier with the

help of such visualizations. The last section deals

with conclusions and future work.

2 DATA AND METHODS

The process of going from the raw clickstreams to

the ﬁnal visualizations from which a user can ex-

tract actionable insights is a long one (Figure 1).

We had the entire clickstream data of all trafﬁc to

www.adobe.com for three days. Each click had mea-

surements on 250 features. The ﬁrst step in our work

involved ﬁltering columns and rows. Many of the 250

features were missing (either partially or completely).

Using the description of the data, we restricted our

analysis to 70 ﬁelds which had relevant information.

Some of the ﬁelds are: visitor IDs, visit number, hit

time, referrer, browser, resolution, geographic region,

IP, page section. Also, given our interest is spatial vi-

sualizations, we restricted our data to only those click-

streams that originated in the USA, to ensure homo-

geneous data. This ﬁltering led to each day having

about 10 million clicks and about a third as many ses-

sions.

For the ease of computation at a later stage, we

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

248

0.00

0.25

0.50

0.75

1.00

−1.0 −0.5 0.0 0.5 1.0

True Correlation

Measure of Association

Measure Entropy Measure Cramer's V

Figure 2: Comparison between Cramer’s V and Entropy

Measure.

sorted this data by session. This sorted data was used

to compute the transition matrices needed for the site

topography visualizations. For visualizing associa-

tions we summarized data at the session level. The

data was further aggregatedon the basis of geographic

location (States of the USA) for the geo-spatial visu-

alizations. Finally the entire solution was packaged

as a webpage with sections dedicated to each type of

analysis, where one could select parameters (like time

of analysis or features to consider). We next describe

the visualizations in more detail.

2.1 Association

A number of parameters of interest are generated in

clickstream data. The ﬁrst question to address in any

data analysis is to identify the inherent structure in

the data. In other words, is there a suggestion of

relationship between different measured dimensions.

The best measure of association between continuous

variables is the correlation coefﬁcient (Kutner et al.,

2004). However, in clickstream data, a number of

parameters of much interest are categorical in nature

(Agresti, 2002), that is, there is no inherent order in

the values (for example, operating system, referring

domain and purchase ﬂag). Hence, we want a mea-

sure of the strength of association between two vari-

ables when either of them can be a continuous or cat-

egorical variable.

As a solution, we binned the continuous variables

to convert them to categorical variables. Such an ap-

proach was necessary because categorical variables

cannot be converted to continuous variables. Now

that we had decided that all the variables will be cat-

egorical in nature, we still needed to decide upon the

best measure of association between two categorical

variables. While there is considerable work on deﬁn-

ing measures of association between two categorical

variables (Press, 1992), we concentrated on the fol-

lowing as they are better accepted in practical usage

Please note here that the p-value of χ

test of indepen-

: C coefﬁcient, Cramer’s V and Entropy based asso-

ciation measure.

2.1.1 Choice of Measure of Association

Of the three measures considered, the C Coefﬁcient

does not take into account the number of bins of the

categorical variable, and hence could not be used to

compare the association between pairs of variables

unless they have equal number of bins. Hence this

measure was not appropriate for our situation. So we

were left with two option, Cramer’s V and Entropy

based measures. We conducted experiments with

these to see which one suits our needs the best. We

generated bivariate normal data with a known correla-

tion structure (referred to as true correlation in Figure

2). Then we binned the random numbers generated

to convert them into categorical variables. Next, we

calculated their association using these two measures.

The results have been plotted in Figure 2. First, notice

that both measures generally increase with increasing

correlation coefﬁcient, which is a desirable property.

Second, both measures are bounded between [0, 1],

which is important for comparability. Finally, both

measures, though increasing, do not have an linear re-

lationship. Although both measures behave similarly,

Cramer’s V lacks the smoothness that one would de-

sire in a measure of association. Hence, we decided

to use the entropy based association measure in our

work.

2.1.2 Entropy based Association

Here is a brief description of the entropy based associ-

ation measure. Given two categorical variables X and

Y, with m and n possible values, respectively, let N

be the number of observation which have the i

value

of X and j

value of Y. Similarly, let N

be the total

number of observations having the i

value of X, N

. j

be the total number of observations having j

value

of Y and, let N be the total number of observations.

Mathematically,

∑

, N

. j

∑

, N =

∑

i, j

Next, deﬁne

, p

. j

dence of the contingency table is a measure of the deviation

from the null hypothesis (independence) and will converge

to 0 as the size of data increases. Hence, this measure is not

usable for our problem.

VizClick-VisualizingClickstreamData

249

The bivariate and univariate entropy measures are de-

ﬁned as follows,

H(X, Y) = −

∑

i, j

ln p

H(X) = −

∑

ln p

H(Y) = −

∑

. j

ln p

. j

Finally, the entropy based association measure is,

U(X, Y) = 2[

H(X) + H(Y) − H(X, Y)

H(X) + H(Y)

2.1.3 Practical Concerns and Visualization

Techniques

Having decided on a measure of association, we had

to decide on the optimal number of bins when con-

verting a continuous variable into a categorical one. A

number of approaches have been proposed to address

this (Sturges, 1926; Freedman and Diaconis, 1981).

We used the Sturges method, as the other approaches

were suggesting too many bins, leading to high cardi-

nality of the resulting categorical variable.

Once all the variables were converted to categori-

cal variables, we still needed to ensure that the cardi-

nality of these variables was not too large. This was

important for the validity of the association measures,

which behave badly when some marginals of the con-

tingency table are sparse. To reduce the number of

possible values of a categorical variable, we pooled

file:// C:/Users/risinhaa Google Drivve/A obee/A obe - Internship 20133 V/ izClickkpaper_vv .1 v/v ctor-images/heatmap.svv

time_website

click_length

new_vis

purchase

shopping_cart_open

shopping_cart_add

shopping_cart_checkout

browser

resolution

connection

help

blog

forum

flash

reader

creative

acrobat

photoshop

dreamweaver

shop

adobe_max

author

account

workspace

store

time_website

click_length

new_vis

purchase

shopping_cart_open

shopping_cart_add

shopping_cart_checkout

browser

resolution

connection

help

blog

forum

flash

reader

creative

acrobat

photoshop

dreamweaver

shop

adobe_max

author

account

workspace

store

Figure 3: A heatmap representation of the association ma-

trix. Each row and column represents some parameter of

interest. The color intensity at some row and column denote

the strength of this association. Each different color denotes

the set of parameters that form a homogeneous cluster of

highly dependent parameters.

all less frequent (≤ 0.1% of the data) values into a

group called “other”.

After computing the associations based on the en-

tropy measure, we considered a number of ways in

which they can be visualized. Our ﬁrst approach was

to use an interactive heat map of the association ma-

trix. Next, we represented the associations as a graph,

with nodes being the parameters and edges represent-

ing the strength of association. We used the forced

layout to display this graph. In addition to these, we

used a chord layout to represent the association val-

ues between each pair of parameters. The thickness

of the chord corresponding to the association value.

We further added an animation to the chord diagram

to represent associations changing over time. All the

mentioned visualizations and their interactivity are

explained in the Results section.

2.2 Geographic Visualizations

The most granular representation of a geographic lo-

cation is the latitude and longitude. Unfortunately,

this information is not available in our clickstream

data. In our experiments, deriving this from the IP

address is error prone. Hence, we decided to restrict

our analysis to some higher level geographicentity. In

our data, when restricting to clicks from the USA, the

State of the visitor was consistently available. Hence,

all our analysis in this section is restricted to analyz-

ing data at the state level.

The ﬁrst step to making geographicvisualizations,

is to summarize data at some geographic unit (in our

case, the State). We used Cartograms (Dougenik

et al., 1985) to visualize features by State. A Car-

togram is powerful as it provides multiple encodings

for representing information. While color is used as

one encoding, it gives the additional ability to use

rescaled area as a visual representation. This dual

encoding is particularly useful for users with achro-

matic vision as they can infer the proportions based on

the region’s scaled area. However, this does require

the user to have ready access to the original unscaled

map, to observe differences. We start our visualiza-

tion with a representation of the unaltered map and

upon feature selection by the user, the map smoothly

transitions into the Cartogram, representing a metric

of interest.

Most features considered for association were

considered here as well. The visualizations were per-

formed for each of the three days and across all fea-

tures. The transitions over features or days allows for

contrasting differences. The results section presents

some samples of this visualization and its interactiv-

ity.

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

250

time_website

click_length

new_vis

purchase

shopping_cart_open

shopping_cart_add

shopping_cart_checkout

browser

resolution

connection

help

blog

forum

flash

reader

creative

acrobat

photoshop

dreamweaver

shop

adobe_max

author

account

workspace

store

time_website

click_length

new_vis

purchase

shopping_cart_open

shopping_cart_add

shopping_cart_checkout

browser

resolution

connection

help

blog

forum

flash

reader

creative

acrobat

photoshop

dreamweaver

shop

adobe_max

author

account

workspace

store

time_website

click_length

new_vis

purchase

shopping_cart_open

shopping_cart_add

shopping_cart_checkout

browser

resolution

connection

help

blog

forum

flash

reader

creative

acrobat

photoshop

dreamweaver

shop

adobe_max

author

account

workspace

store

Figure 4: Figure A shows the default view of the chord diagram. Figure B shows the “browser” feature highlighted, it is clear

that it is highly associated with “OS” and “resolution”. Figure C has the “new visitor” highlighted and it is strongly associated

with “time spent on website” and “average click length of session”. This piqued our interest and upon further exploration, we

found that new visitors tend to spend lesser time on the website and have fewer clicks.

While the Cartogram provides a succinct visual-

ization of a particular metric, there are some situations

where it might be misleading. One such situation is

when the user wants to contrast the observations with

some expectation. For example, one may expect that

visits to www.adobe.com follow the distribution of in-

ternet population across States. Deviations from this

expectation may be of concern. With this motivation,

we created a visualization that highlights such devi-

ations, using color as a visual encoding to represent

deviations. Expectations other than internet popula-

tion may be based on State Gross Domestic Product

(GDP) or State Disposable Income, which may be of

more interest to a marketer. The marketer may be

concerned that a state with signiﬁcant GDP is con-

tributing little to sales. We refer to such infographics,

which even out the metric based on some underlying

factor as “normalized geographic visualization” in the

rest of the paper.

2.3 Website Topography

Next we propose a series of visualizations that help

the website owner understand how the users interact

with the website to identify possible areas of improve-

ment. The ﬁrst step in doing this is the creation of

a transition matrix based on the clickstreams; which

stores how many times users have gone from one page

to another. Alternatively, this information may be

stored as a directed graph, where each node is a page,

and the edge is the volume of trafﬁc between them (in

the given direction). In addition to the pages as nodes,

we added two artiﬁcial nodes to represent entry to the

website and exit from the website.

The ﬁrst question to address for a website owner is

to understand which pages lead to the highest bounce

rate, that is, a high proportion of users arriving at the

page are leaving the site altogether. To represent this,

we used a bar graph, where the length of the bar rep-

resents the bounce rate. Additionally, we used color

to represent the number of users arriving at the partic-

ular page, thus identifying the most important pages

where the bounce rate is a worry.

Next, we delve more into how users interact with

the website. Our initial attempt was at displaying the

website as a graph with the edges representing transi-

tions. Unfortunately, in our data, we had of the order

of 10, 000 distinct pages. This meant that visualiz-

ing the raw graph was not feasible. The structure of

this network is lost if we attempt to visualize only the

more active nodes. To overcome this challenge, we

had to cluster nodes in some manner. We explored

two different approaches to do this.

First, we used the modularity based approach to

detecting communities in large networks (Blondel

et al., 2008). This algorithm, when given a graph,

produces successive clustering which is based on the

connectivity of the nodes in the graph. Successive

clustering leads to a hierarchical tree, leaf nodes rep-

resenting individual webpages and internal nodes rep-

resenting clusters. This tree is then visualized using

an interactive bubble plot. The bubble plot creates a

bubble for each leaf node, and combines these to form

larger bubbles representing a cluster, which may fur-

ther group to form a higher level cluster. The size of a

bubble is proportional to the total trafﬁc on that page.

We also name the cluster by identifying the major

keywords in URLs for each cluster, where keywords

are deﬁned as sections of website to which webpages

belongs. Another level of interactivity is added by

providing links to the actual pages which upon click-

ing leads the user to the page.

The second way to cluster the webpages was

based on the hierarchy in the site structure. Here we

use the URL’s of the webpage to identify parents to

webpages and cluster based on these. This cluster-

ing is then visualized as a collapsible tree, where the

user can expand only the web-sections she wants to

VizClick-VisualizingClickstreamData

251

Figure 5: Figure A shows the original map of US with no distortion, the default view. Figure B shows the Cartogram for

number of new visitors for each State. Figure C shows the number of purchases from each State. The inﬂation in size of

Washington DC is clearly apparent from Figure B and C. Also note that California is blown up in both of them, possibly

indicating a pattern that needs further exploration.

concentrate on. As the next step, to make the trafﬁc

patter apparent, we added interactivity. When a user

hovers over any node in the tree, all its highest de-

gree neighbors (either in-degree or out-degree) in the

graph are highlighted (by increasing the node’s radius

proportional to the edge weight). This view provides

an interactive view of the user behavior keeping the

web designer’s site structure intact.

In addition to the above, we considered another in-

teresting question which can be answered from click-

stream data. If the data suggests that a lot of users are

navigating from page A to page C, via a set of pages

, then it may be interest to see if a link to page C

can be provided on page A. We identiﬁed such pairs

based on the transition matrix and visualized them as

a bar chart.

3 RESULTS

In this section, we show the visualizations we have

created and highlight some inferences that can be eas-

ily drawn from them.

3.1 Association

Figure 3 shows the heatmap of the association ma-

trix. The association matrix was used as a distance

matrix to create clusters of parameters, based on hi-

erarchical clustering. For example, the largest cluster

in the center has the sections of the website that are

frequently visited together, such as, “shopping cart

open” and “shopping cart add”. Each cluster is shown

with a color and cross cluster cells are shown in black.

The color intensity represents the strength of associ-

ation, higher associations having darker shades. As

expected, each variable is perfectly associated with it-

self. An interesting aspect that can be easily identiﬁed

from this view is the small cluster of three parameter

“OS”, “browser”and “resolution”, which points to the

high association between the three.

A graph is a natural representation of a distance

matrix. Hence, we next created a graph with a

forced layout. The nodes are the parameters and

edge weight is the association between each pair of

nodes. The graph (not shown in the paper) is, unfor-

tunately, too cluttered to draw any signiﬁcant infer-

ences. This motivated us to explore some other visu-

alization which presents the same information, with

less clutter. Hence, we used a Chord diagram to rep-

resent the same information (Figure 4).

In the ﬁrst view of this visualization, the arc length

provided to each of the parameters is proportional to

the sum of its association values with all other param-

eters. For example, “OS” has a higher arc length than

“account”, representing that “OS” is more associated

with other parameters, whereas “account” is not re-

lated to any other parameter. Also clearly noticeable

is the thick chord between “shop” and “photoshop”,

perhaps indicating that a lot of people looking to shop

on Adobe’s website are interested in Photoshop



. In

part B (which can be achieved by hovering over a pa-

rameter), the focus shifts to “browser” and helps the

user concentrate on this particular aspect. As noted

from the heatmap as well, “browser” is highly asso-

ciated with “resolution” and “OS”. Finally in part C,

the “new visitor” parameter is highlighted (which is

a binary ﬂag indicating whether the visitor is new or

has visited before). Notice the high association be-

tween “new visitors” and the “time spent on the web-

Figure 6: Geographic visualization of purchases normalized

for the internet population in each state. California, which

was highlighted in Figure 5, is now seen to be following

expectations.

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

252

get help flash

shop photoshopshop photoshop

flash helpflash help

help account downloadhelp account download

help account forum

helphelp

creativecreative

acrobatacrobatacrobat

shop photoshopshop photoshop

shop photoshop

download

shop photoshop

download

flash getflash getflash getflash get

flash accountflash accountflash accountflash account

accountaccountaccountaccount

flash help blogflash help blogflash help blog

flash

blog help

download help

accountaccountaccount

acrobat help readeracrobat help reader

acrobat help reader

download acrobat helpdownload acrobat helpdownload acrobat help

accountaccountaccount

flash forum helpflash forum helpflash forum help

flash forum help

flash help forum

account creativeaccount creativeaccount creativeaccount creative

shop photoshopshop photoshop

shop photoshop

download

help photoshop shop

shop photoshop

creative

download

shop photoshop

download

flash help forum

IUUQXXXBEPCFDPNQSPEVDUTQIPUPTIPQTIPXDBTFIUNM

Figure 8: Figure A shows the entire hierarchy of the website based on modularity as successive bubbles. Next, Figure B

shows one of the bubbles focused on, and the names of various other sub-bubbles are also visible which gives an idea of the

prominent tags in those bubbles. Figure C shows the URLs that are visible upon hovering over an individual webpage. These

bubbles also serve as links to move directly to those webpages.

site”. This raises a hypothesis that may be tested to

get more insights from the data. In such an explo-

ration, we found that new visitors spend lesser time

on the website than others. Since all the parameters

are categorical in nature with no inherent ordering,

the actual trend cannot be predicted based just on the

association values.

3.2 Geographic Visualizations

Figure 5 provides views of the Cartogram. In part A,

we see the original map of the USA with no scaling.

In part B, the Cartogram is drawn according to the

metric “new visit” (number of new visitors from each

State). Part C shows the Cartogram “purchases”. An

interesting insight is that the size of Washington DC

inﬂates considerably when purchases are considered.

We providethe option to visualize various metricslike

these and cross metric comparisons are facilitated by

smooth transitions between Cartograms. Such infor-

mation and comparisons provide actionable insights

for marketers.

While the Cartogram provides an effective tool to

use size as a visual encoding for a metric, there are

important patterns that are not highlighted. For ex-

ample, in all the cartograms we created, we see con-

sistently that California is signiﬁcantly larger than its

original size, could this be because California has a

http:// blogs.adob e.com/typ blograp hy/2013/0 3/introd ucing-ad obe-blan k.html

http:// www.macromedi a.com/suppo rt/docume ntation /en/fla shplayer/hel p/hel p06.htm l

http:// www.macromedi a.com/suppo rt/docume ntation /en/fla shplayer/hel p/setting s_manager.ht ml

http:// www.macromedi a.com/suppo rt/docume ntation /en/fla shplayer/hel p/setting s_manager07 .html

ad_us_di scount-software-coupons_di scount-software-coupon s.html

HTML Trial Download : Download : Adobe Photoshop CS6

en-US: Expo rtPDF.Acrobat.com : pricin g.html

AIR App Trial Download : Badge Cl ickthrough: Adobe Ph otoshop CS6

en-US: Create PDF.Acrobat.com : features.htm l

HTML Trial Download : Download : Adobe Acrobat XI Pro

http:// www.macromedi a.com/suppo rt/docume ntation /en/fla shplayer/hel p/hel p01.htm l

http:// helpx.a dobe.com /en/acroba t.html

http:// helpx.a dobe.com /acrobat/rel ease-note/rel ease-notes-acrobat-read er.html

AIR App Trial Download : Badge Cl ickthrough: Adobe Acro bat XI Pro

en-US: Create PDF.Acrobat.com : pricin g.html

Adobe Commun ities: Ado be Ill ustrator: Illu strator: Discussi ons: View T hread

http:// helpx.a dobe.com /acrobat/kb/pri nting-ti ps-acrobat-reader.ht ml

Adobe Commun ities: InDesig n: InDesign: Discussi ons: View T hread

http:// helpx.a dobe.com /acrobat/u sing/flash-pla yer-needed -acrobat-reader.h tml

Adobe Commun ities: Fla sh Player: Fla sh Player FAQs: Discussions: View T hread

Adobe Commun ities: Ado be Photo shop: Photo shop General Discussi on: Discussions: View T hread

muni ties: Adob e Premie re Forums: Premi ere Pro CS5, CS 5.5, &am p; CS6: Discussions: Vie w Thre

http:// www.adobe.com /products/eul as/#flash_pl ayer

http:// www.adobe.com /misc/terms.htm l

http:// helpx.a dobe.com /reader/using /print-pdfs.htm l

Adobe Commun ities: Acroba t: Acrobat Wi ndows (read-only): Di scussions: View T hread

http:// help.a dobe.com /en_US/Rea der/9.0/con tent.htm l

ad_us_di gital -editio ns_digit al-edit ions.html

en-US: wel come: wel come.ht ml

Adobe Commun ities: Ado be Reader Fo rums: Adobe Re ader: Discussions: View T hread

ad_us_read er_reader.ht ml

Adobe Commun ities: Fla sh Player: Using Fl ash Player: Di scussio ns: View Th read

ad_us_Fla shPlayer_fl ashplayer.htm l

http:// helpx.a dobe.com /reader/using /reader-x-topics.htm l

www.adobe. com : Pag e Not Found

https://kuler.a dobe.com /

http:// helpx.a dobe.com /acrobat.h tml

http:// www.adobe.com : noscript

http:// www.adobe.com /misc/unsuppo rted_devi ce.html

cc_en-us_Home _PageVi ew

ad_us_Acrob atPro_acroba tpro.html

ad_us_do wnload_ downloa d.html

Photoshop .com: PSco m Home: Home

ad_us_Fla shPlayer_tech -specs.html

http:// helpx.a dobe.com /flash-playe r/kb/installa tion-probl ems-flash-playe r-mac.html

http:// www.adobe.com /m/fla shplayer/app le.htm l

http:// labs.adobe .com/do wnloads/fla shplayer.htm l

http:// helpx.a dobe.com /flash-playe r/kb/flash-player-go ogle-chrom e.html

Adobe Worldwide Licensing : logi n.htm

ad_us_Acrob atFamil y_acrobat. html

http:// www.adobe.com /products/eul as/

http:// www.adobe.com /

http:// helpx.a dobe.com /photoshop /topics.html

http:// helpx.a dobe.com /reader.htm l

0 5 10 15 2 0 25 30 35

count

Figure 7: Pages with the highest bounce rates. The length

denotes the proportion of bounce, and the color denotes the

number of visitors. Red signiﬁes the highest trafﬁc pages,

yellow is next, followed by blue and then green.

large share of the internet users of the USA? The Car-

togram fails to answer this question. This led to a

need for visualizations based on normalizations, that

help detect deviations from what is expected.

In the normalized geographic visualization pre-

sented in Figure 6, red implies negative deviation and

green represents positive deviation from expected be-

havior, with light grey being the neutral color. As can

be seen, California is very close to grey implying that

the number of purchases from there are as expected

according to population of internet users in that state.

Alaska on the other hand is bright green, generating

a ﬁnding that may warrant further exploration. Simi-

larly, Michigan is deep red, and draws the marketer’s

attention as expected from such a visualization.

3.3 Website Topography

As mentioned earlier, before delving deep into web-

site topography, we ﬁrst started with some primary

questions that a website owner may have. Figure 7

displays the pages with the highest bounce rates, the

color coding further denoting which of these are the

highest trafﬁc pages.

We next clustered the webpages based on modu-

larity and visualized them using the bubble plot. Fig-

ure 8 shows this visualization. Part A shows the de-

fault view of the visualization, with successive clus-

ters represented as successive bubbles. The size of a

node is proportional to the sum of the in-degree and

out-degree. Part B displays the interactive capabilities

of the visualization, here a particular bubble has been

brought into focus by clicking on it. Bubbles have

been named depending on the major sections of the

website that are a part of the bubble’s nodes. In part

C, a particular webpage node has been highlighted

which displays the URL of the webpage on the bot-

tom of the visualization. The visualization interface

also supports clicks on the node to directly move to

VizClick-VisualizingClickstreamData

253

the webpage representing the node.

Next, we created a visualization based on the web-

site’s structure (Figure 9). As explained earlier, in

this visualization, hovering on particular node leads

to its neighboring nodes being highlighted (increases

in size). This increase is proportional to the edge

weights between the two nodes. By doing so, we

avoid unnecessary cluttering in representing the edges

https://www.adobe.com/support/downloads/product.jsp

https://www.adobe.com/cfusion/support/index.cfm

https://www.adobe.com/cfusion/exchange/index.cfm

https://www.adobe.com/cfusion/entitlement/index.cfm

https://www.adobe.com/account/account-information.html

https://www.adobe.com

https://www.acrobat.com/exportpdf/en/home.html

https://www.acrobat.com/createpdf/en/print-to-pdf.html

https://licensing.adobe.com/sap(bD1lbiZjPTAwMw==)/bc/bsp/sap/zavlcommon/error.htm

http://www.photoshop.com/tools/get_flash

http://www.photoshop.com/products/mobile/express/ios

http://www.photoshop.com/home

http://www.photoshop.com

http://www.macromedia.com/support/documentation/en/flashplayer/help/help09.html

http://www.adobe.com/support/security/bulletins/apsb13-01.html

http://www.adobe.com/support/downloads/detail.jsp

http://www.adobe.com/products/premiere.html

http://www.adobe.com/products/kuler.html

http://www.adobe.com/products/illustrator.html

http://www.adobe.com/products/flash-media-encoder.html

http://www.adobe.com/products/discount-software-coupons.edu.html

http://www.adobe.com/products/creativesuite/faq.html

http://www.adobe.com/products/creativesuite.html

http://www.adobe.com/products/creativecloud/tech-specs.edu.html

http://www.adobe.com/products/creativecloud.html

http://www.adobe.com/products/creativecloud.edu.html

http://www.adobe.com/products/air/faq.html

http://www.adobe.com/products/acrobat/convert-pdf-converter.html

http://www.adobe.com/privacy/opt-out.html

http://www.adobe.com/education/products/creativesuite/mastercollection.edu.html

http://www.adobe.com/downloads/cs2_downloads/index.html

http://www.adobe.com/cart.html

http://www.adobe.com

http://helpx.adobe.com/support.html

http://helpx.adobe.com/reader/using/reader-x-topics.html

http://helpx.adobe.com/photoshop/topics.html

http://helpx.adobe.com/photoshop-elements/topics.html

http://helpx.adobe.com/media-encoder/topics.html

http://helpx.adobe.com/flash-player.html

http://helpx.adobe.com/content/help/en/premiere-pro/using/default-keyboard-shortcuts.html

http://helpx.adobe.com/acrobat/kb/troubleshoot-reader-installation-windows.html

http://help.adobe.com/en_US/premierepro/cs/using/WS1c9bc5c2e465a58a91cf0b1038518aef7-7f35a.html

http://help.adobe.com/en_US/photoshop/cs/using/WSfd1234e1c4b69f30ea53e41001031ab64-76dba.html

http://help.adobe.com/en_US/photoshop/cs/using/WS0DECAB01-C01F-4498-A12F-18B94EC9DACF.html

http://help.adobe.com/en_US/Reader/9.0/content.html

http://help.adobe.com/en_US/FlashPlayer/LSM/WS6aa5ec234ff3f285139dc56112e3786b68c-7ff0.html

http://get2.adobe.com/flashplayer/download

http://get.adobe.com/reader/direct

http://feedback.photoshop.com/photoshop_family/products/photoshop_family_adobe_revel

http://blogs.adobe.com/lightroomjournal/2012/10/camera-raw-7-2now-available.html

EXIT

ENTRY

root

Figure 9: Site Topography. Hovering over a node increases

the size of its neighbors (based on in-degree or out-degree).

but still provide access to the information. Also by

examining the inter-cluster trafﬁc, analysts can get

an idea of cross cluster movements. If inter-cluster

movements are highly signiﬁcant, the web designer

can consider redesigning their site structure, to pro-

vide a better user experience. Another use of this vi-

sualization is as follows: Suppose, there is a sudden

drop in the number of users moving from page A to

page B. From the visualization presented, such sud-

den changes can be easily observed, and may point to

a broken link or other anomalies.

Lastly, we identiﬁed suggested links for certain

pages based on the trafﬁc pattern. We visualized this

as a histogram (not shown in the paper), with each

row corresponding to a pair of webpages (A, C). The

pair is such that it might improve user experience if a

link to page C is added on page A.

4 CONCLUSIONS

We have compiled a collection of visualizations that

can be used to understand three important aspects of

clickstream data, namely, association between param-

eters, spatial aspects of clickstream data and site to-

pography. We have shown that a number of inter-

esting and useful insights can be generated based on

the visualizations created. In addition, these visual-

izations serve as a way of generating hypotheses that

can then be tested on the data.

The proposed visualizations can serve as the ﬁrst

step in making sense out of large web analytics data.

Additionally, we have focused extensively on creat-

ing visualizations that are interactive. Given the size

of clickstream data, it is important for the user to have

the ability to dig down further into areas that might be

of interest. Also, interactivity provides a way of visu-

alizing large amounts of data without over powering

the visual senses of the viewer. Our visualizations of

the website topography are designed to achieve this

goal. However, in the context of visualizing large

graphs, there is scope for building in even higher lev-

els of interactivity without losing out on the ability to

communicate information.

5 ADDITIONAL RESOURCES

Please visit this link for a short video description

of VizClick, highlighting some of its interactivity:

http://youtu.be/fv9qrU0EZJE.

IVAPP2014-InternationalConferenceonInformationVisualizationTheoryandApplications

254

REFERENCES

Agresti, A. (2002). Categorical data analysis, volume 359.

John Wiley & Sons.

Blondel, V. D., Guillaume, J.-L., Lambiotte, R., and Lefeb-

vre, E. (2008). Fast unfolding of communities in large

networks. Journal of Statistical Mechanics: Theory

and Experiment, 2008(10):P10008.

Bostock, M., Ogievetsky, V., and Heer, J. (2011). D

data-driven documents. Visualization and Computer

Graphics, IEEE Transactions on, 17(12):2301–2309.

Brainerd, J. and Becker, B. G. (2001). Case study: E-

commerce clickstream visualization. In infovis, pages

153–156.

Cleveland, W. S. (1993). Visualizing data. Hobart Press.

Dougenik, J. A., Chrisman, N. R., and Niemeyer, D. R.

(1985). An algorithm to construct continuous area car-

tograms. The Professional Geographer, 37(1):75–81.

Ferreira de Oliveira, M. C. and Levkowitz, H. (2003). From

visual data exploration to visual data mining: A sur-

vey. Visualization and Computer Graphics, IEEE

Transactions on, 9(3):378–394.

Freedman, D. and Diaconis, P. (1981). On the histogram

as a density estimator: L-2 theory. Probability theory

and related ﬁelds, 57(4):453–476.

Kutner, M. H., Nachtsheim, C., Neter, J., et al. (2004). Ap-

plied linear regression models. McGraw-Hill New

York.

Lee, J., Podlaseck, M., Schonberg, E., and Hoch, R. (2001).

Visualization and analysis of clickstream data of on-

line stores for understanding web merchandising. In

Applications of Data Mining to Electronic Commerce,

pages 59–84. Springer.

McAfee, A., Brynjolfsson, E., et al. (2012). Big data:

the management revolution. Harvard business review,

90(10):60–66.

Press, W. H. (1992). Numerical recipes in Fortran 77: the

art of scientiﬁc computing, volume 1. Cambridge uni-

versity press.

R Core Team (2013). R: A Language and Environment for

Statistical Computing. R Foundation for Statistical

Computing, Vienna, Austria.

Sarkar, D. (2008). Lattice: multivariate data visualization

with R. Springer.

Sturges, H. A. (1926). The choice of a class inter-

val. Journal of the American Statistical Association,

21(153):65–66.

Swayne, D. F., Lang, D. T., Buja, A., and Cook, D. (2003).

Ggobi: Evolving from xgobi into an extensible frame-

work for interactive data visualization. Computational

Statistics & Data Analysis, 43(4):423–444.

Tufte, E. R. and Graves-Morris, P. (1983). The visual dis-

play of quantitative information, volume 2. Graphics

press Cheshire, CT.

Wei, J., Shen, Z., Sundaresan, N., and Ma, K.-L. (2012). Vi-

sual cluster exploration of web clickstream data. In Vi-

sual Analytics Science and Technology (VAST), 2012

IEEE Conference on, pages 3–12. IEEE.

Wickham, H. (2009). Ggplot2: elegant graphics for data

analysis. Springer Publishing Company, Incorporated.

VizClick-VisualizingClickstreamData

255