Interpreting and Leveraging Browser Interaction

for Exploratory Search Tasks

Dominic Stange

, Michael Kotzyba

, Stefan Langer

and Andreas N

urnberger

DKE Group, Faculty of Computer Science, University of Magdeburg, Germany

Volkswagen AG, Germany

Keywords:

Web Usage Mining, Interaction Log Analysis, Exploratory Search, Recommender Systems, Interactive

Information Retrieval, Machine Learning.

Abstract:

In this paper we introduce a novel approach for modeling and interpreting search behavior for exploratory

search by using a so called exploration graph. We use an existing methodology of logging and analyzing user

interactions with a web browser and add an additional interpretation step that can be used, e. g. to integrate

sensemaking or browsing patterns into the log data. We conducted a user study and are able to show that:

(a) interaction logs can be interpreted semantically, (b) semantic interpretations lead to a more connected

exploration graph, and (c) multiple (even contradicting) interpretations of the same search behavior may exist

at the same time. We also show how our theoretical model can be applied in the area of professional search by

incorporating insights gained from the model into novel recommendation and machine learning approaches.

1 INTRODUCTION

In contrast to simple fact-ﬁnding search, in ex-

ploratory search techniques of learning and investi-

gating are used (Marchionini, 2006). Furthermore,

exploratory search is characterized as open-ended and

multifaceted with unclear goals (White and Roth,

2009; Wildemuth and Freund, 2012). These char-

acteristics make it difﬁcult to track and examine a

user’s learning process during search just by consider-

ing the interactions with a browser. Especially, in the

World Wide Web typical interactions of exploratory

search are likely to involve “many impasses, illstruc-

tured goals and tasks, navigation and exploration, and

substantial inﬂuences from the content that is encoun-

tered” (Card et al., 2001). Methods that enable efﬁ-

cient support during interactive (exploratory) search

could strongly improve the efﬁciency of the search

process, especially in the area of professional search

(e.g. technology scouting), which is a common task in

large enterprises (N

urnberger et al., 2015). However,

one major problem is that an appropriate data model

for exploratory search which also enables mecha-

nisms for effective storage and analysis is still miss-

ing. Therefore, we propose in this paper a data struc-

ture and preliminary results of a small user study to-

wards improved ranking and recommendation based

on semantic interpretations of exploration graphs.

This paper is structured as follows: The next sec-

tion describes related work. Sect. 3 provides a formal

model for exploratory search behavior. In Sect. 4 we

outline how the model allows us to interpret search

behavior and provide a problem deﬁnition. After-

wards, the conducted user study and preliminary re-

sults are described in Sect. 5. Then, Sect. 6 shows

how data about interpreted search behavior can be

used as input to a recommender system in a business

environment. In Sect. 7 a short discussion and con-

clusion is given.

2 RELATED WORK

The analysis of search behavior has a long history

in research and is rooted in the area of library and

information science with the goal to gain a holis-

tic view on the search process and to derive gener-

alized models (Bates, 1989; Marchionini, 1997). A

good overview on this macroscopic view is avail-

able in (Knight and Spink, 2008). Further research

in information retrieval focuses on the user’s con-

crete interactions to build models which represent the

user’s state or search success. For example, Hassan et

al. (Hassan et al., 2010) propose a Markov Model with

transition times and show that this approach performs

signiﬁcantly more accurate than traditional relevance-

Stange, D., Kotzyba, M., Langer, S. and Nürnberger, A.

Interpreting and Leveraging Browser Interaction for Exploratory Search Tasks.

DOI: 10.5220/0006372901910197

In Proceedings of the 19th International Conference on Enterprise Information Systems (ICEIS 2017) - Volume 3, pages 191-197

ISBN: 978-989-758-249-3

191

based models for predicting user search goal success.

Ageev et al. (Ageev et al., 2011) apply different model

approaches to predict user’s success for fact-ﬁnding

search tasks. Card et al. (Card et al., 2001) as well

as White and Drucker (White and Drucker, 2007)

use browser interaction analysis to create behavioral

graphs that model search behavior. Their studies rely

on browser interaction logs which store data about

how searchers interact with their Web browser. In par-

ticular, White and Drucker investigate browser trails,

i.e., the sequence of visited websites. For example,

if a user visits two websites, a and b, in a row and

returns to a to browse to a website, c, the correspond-

ing search trail can be written as a → b → a → c.

This approach allows to, e. g. analyze and investigate

searcher proﬁles.

Our approach facilitates browser interaction logs,

as well. We use them to model a user’s search be-

havior and keep more information about the actual

browsing activities in a search than with behavioral

graphs. With this model we are also able to create dif-

ferent interpretations of search behavior, based on ad-

ditional assumptions about a user’s sensemaking ac-

tivities during search. The problem of creating such

interpretations is formulated in the course of the next

sections.

3 A MODEL FOR

EXPLORATORY SEARCH

This section outlines a model to formally describe

search behavior based on browser interaction. The

model is particularly designed with an exploratory

search task in mind.

When a person explores the World Wide Web

using a Web browser there is always one website

visible in the browser. If the URL to this web-

site is, e. g. www.domain.com we say that the per-

son is in state ON(”www.domain.com”). If the per-

son moves to another website by interacting with the

Web browser, e. g. by clicking on a hyperlink within

the website and the URL to this second website is

www.anotherdomain.com, the action can be written

as GOT O(”www.anotherdomain.com”). Hence, the

corresponding search process of the person starting

on the home screen can be written as an alternating

sequence of states and actions:

State 1: ON (”home screen”)

Action 1: GOT O(”www.domain.com”)

State 2: ON (”www.domain.com”)

Action 2: GOT O(”www.anotherdomain.com”)

State 3: ON (”www.anotherdomain.com”)

Action 3: . . .

We call this sequence of states and actions the per-

son’s search history. The search history is the result

of observing the person’s search behavior. A search

history contains the websites a person visits during a

search process and the order in which they are vis-

ited over time. It also contains the actions through

which the websites are opened. Note that it is possible

for the person to visit a website multiple times during

search, e. g. when switching browser tabs, so different

states in the search history can have the same web-

site. One could say that through browser interaction

the person transitions between different states, where

each state is determined by the visited website. These

transitions can be written as

RESULT (ON(w),GOTO(w

)) = ON(w

i. e. the person moves from state ON(w), where web-

site w is visited, to another state, ON(w

), where web-

site w

is visited by performing the GOTO-action.

Web browsers typically process the data about

these websites and also the browser tabs in which they

are shown. Since we want to model a user’s interac-

tion with a Web browser our model has to consider

all browser tabs and the websites they display. To

simplify the model we assume that only one browser

window is used by the user. Therefore, only one web-

site is visible in the web browser at any point in time

while the rest of the websites are hidden. We call this

visible website the active website of the browser. The

browser state can then be deﬁned in terms of the ac-

tive and all opened websites.

Deﬁnition 1 (Browser State). Given a set W of web-

sites, a browser state s can be written as

s(W ) = ON(W ) = ON(w

,...,w

n−1

where each website w

∈ W is shown in a browser tab

and the underlined website w

is the active website of

browser state s, with 1 ≤ k ≤ n.

At the beginning the browser state is typically

empty, i. e. the set W is empty. We write this as

s = ON(∅). Similarly to the user’s search history the

interactions with the browser result in an alternating

sequence of browser states and user actions. We call

this sequence the browser history.

Deﬁnition 2 (Browser History). Given a set S of

browser states and a set A of user actions, a browser

history h is an alternating sequence

h(S ,A) = h s

, a

, s

, a

, ...,s

n−1

, a

, s

of browser states s

∈ S and user actions a

∈ A.

An example of a browser history in which the

websites w

, w

, and w

are visited is:

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

192

h s

: ON (∅), a

: GOTO(w

: ON (w

), a

: GOTO(w

: ON (w

), a

: GOTO(w

: ON (w

) i.

Initially, the browser history is empty. Then, web-

sites are opened in the course of the search process.

Notice that in the example the third website, w

, is

opened in a new browser tab. Thus, the previous web-

site, w

, is not removed from the set of websites W

in state s

. That is, if we only consider the GOT O-

action it is not possible to “predict” its exact outcome

for a browser state. In other words, the RESULT -

function is not determined for a browser state and the

GOTO-action. Therefore, we need to observe other

user actions to solve this ambiguity. Preliminary anal-

ysis shows that the eight actions of Table 1 are well-

suited for our search model to work as intended:

Table 1: User actions and their descriptions. The actions are

part of our exploratory search behavior model.

Action Description

URL open a website in the active browser tab

URL

open a website in a new browser tab

CLICK enter a URL in the address ﬁeld of the

active browser tab

CLICK

enter a URL in the address ﬁeld of a

new browser tab

CLOSE close an existing browser tab

SW ITCH switch from one browser tab to another

BW /FW move backward/forward within the lo-

cal history of a browser tab

QUERY enter a search query in the search box of

the active browser tab

The actions have their expected effect on a

browser state. The U RL-action opens a website for

a given URL which is entered in the address ﬁeld

of the active browser tab. The action removes the

active website and adds the new website to the next

browser state. For example, if the browser is in state

ON(w) and the user performs the URL-action the

transition is RESULT (ON(w),URL(w

)) = ON(w

The symbol “

” in the URL

-action indicates that

the website for the URL is opened in a new browser

tab. For example, if the browser is in state ON(∅)

and the user performs the URL

-action the corre-

sponding transition is RESULT (ON(∅),URL(w)) =

ON(w). The CLICK-action removes the active web-

site and adds the website of the clicked hyperlink to

the next browser state. Consistently, the CLICK

action opens the website of the hyperlink in a new

browser tab and, therefore, adds this website to the

next browser state without removing the active web-

site. The CLOSE-action removes the active website

from the next browser state. The SW ITCH-action

changes the active website from one website to an-

other website. The FW /BW -action uses the local his-

tory of a browser tab to navigate forward and back-

ward, replacing the active website with the one of

the local history. Lastly, the QUERY -action issues

a search query to the active website, removing it from

the next state and adding the corresponding search en-

gine results page (SERP). Using these user actions the

browser history for the previous example can be writ-

ten as:

h s

: ON (∅), a

: URL

: ON (w

), a

: CLICK (w

: ON (w

), a

: CLICK

: ON (w

) i.

As can be seen, this browser history describes the

person’s search behavior unambiguously. Note that

sometimes an action is not applicable for a browser

state. For example, the QUERY -action is only appli-

cable if there is a search box on the active website.

The actions are atomic in the sense that any of them

creates a new browser state. Some Web browsers pro-

vide functions that perform more than one atomic ac-

tion in one interaction, e. g. opening a hyperlink in a

background browser tab.

4 INTERPRETING BROWSER

HISTORIES

With the help of the eight user actions listed in Ta-

ble 1 of the previous section we can create a sufﬁ-

ciently rich model of any person’s exploratory search

behavior. Now, based on the model we want to intro-

duce the noval approach of interpreting this behavior

formally and also propose the concept of semantics in

this interpretation.

A structure that takes both the browser states and

transitions between them into account is a graph. We

call this the exploration graph.

Deﬁnition 3 (Exploration Graph). An exploration

graph x(V , E ) is a cyclic, directed graph with a set

V of vertices and a set E of edges, where elements

of E are ordered pairs (v

) of distinct vertices

∈ V .

An exploration graph describes how a user ex-

plores an information space based on the websites be-

ing visited during search. Every vertex in x represents

exactly one website. An example of an exploration

graph with ﬁve vertices and four edges is shown in

Interpreting and Leveraging Browser Interaction for Exploratory Search Tasks

193

→ s

START

Figure 1: An exploration graph for a random browser his-

tory. It consists of ﬁve nodes (w

to w

) that represent the

visited websites. The edges between the nodes indicate how

the information space is traversed. The edge labels indicate

between which states the connection is added.

Figure 1. The ﬁgure shows an extension to the exam-

ple from the previous section, where two additional

websites, w

and w

, are opened.

An exploration graph visualizes the different ex-

ploration paths of a user, e. g. w

→ w

or w

→

→ w

. Note the difference of the paths to the one-

dimensional strictly chronological sequence w

→

→ w

or a browser trail w

→ w

→

→ w

. The direction of an edge be-

tween two websites is determined by the direction

of the hyperlink between them, e. g. w

→ w

. For

ease of understanding the edges of the example ex-

ploration graph are labeled such that they show be-

tween which states they are added. In general, it is

possible that a user visits the same website multiple

times (e. g. following links on different websites to the

same target) and, hence, the graph can contain cycles.

New paths are created whenever a user enters a URL

manually or issues a search query to a search engine.

That is, in contrast to browser trails, for example, an

edge is not automatically appended to the last active

vertex, reducing possible randomness (e. g. the link

→ w

) and also avoiding heuristics to close a path,

c.f. (White and Drucker, 2007).

However, if a user issues a search query based on

something she has learned on a website, it is reason-

able to assume that this search query continues the

path that has led to the website. That is, search queries

may refer to the content of visited websites. This re-

lationship is typically not present in the browser his-

tory because there is no interaction to be logged. Yet,

an edge connecting the website with the SERP of the

query can be weighted by the degree/probability that

they are semantically related. Thus, edges in explo-

ration graphs can express semantics in search behav-

ior. For instance, the probability could indicate how

likely it is that the user has used insights gained from

a speciﬁc website to issue the search query. This is

likely to result in a more connected exploration graph,

because in addition to the actual browsing the seman-

tic relationships are taken into account. Following

this approach also means that there is not a single

“true” exploration graph but several possible instanti-

ations that are all different interpretations of the same

search behavior. In the example shown above not all

actions have been interpreted in a way that they cre-

ate an edge between two websites. For example, a

SW ITCH-action is not represented by an edge. This

is in contrast with the interpretation corresponding to

the browser trail, which - using our terminology - can

be thought of as being based on a search history rather

than a browser history. In some cases it may be rea-

sonable to use one interpretation over another depend-

ing on the research scenario. There are certainly more

types of interpretations possible. A detailed analysis

of possible interpretations is beyond the scope of this

paper but subject for future work.

The problem of creating an interpretation for (ex-

ploratory) search behavior can be stated as follows:

Problem Statement: Given a browser history

h(S ,A), ﬁnd an interpretation function

f : (h)

theory

−−−→ x

that best represents the search behavior modeled in h

as an exploration graph x(V ,E).

By adding the component theory to the problem

statement we emphasize that interpretations can make

assumptions about a user’s search behavior. These as-

sumptions can be used, e. g. to incorporate semantic

relationships into the exploration graph. An interpre-

tation can be understood as a projection of observa-

tional and/or sensemaking patterns on the data col-

lected via browser interactions of a user.

5 USER STUDY AND

PRELIMINARY RESULTS

One goal of the user study is to show how a semantic

interpretation of search behavior looks like and how

it compares to an interpretation without assumptions.

Using insights gained from a pre-study in which we

investigated the complexity of exploration graphs, we

designed the main study as follows.

Study Design. Participants are asked to search

for information about two predeﬁned topics. The

topics are selected to satisfy the attributes described

in (Wildemuth and Freund, 2012): i. e. learning

and investigation as goals; general, ill-structured and

open-ended problems; and an involved uncertainty.

One search topic is about the procurement of so-

lar energy panels for a private houshold as adapted

to (Gwizdka and Lopatovska, 2009). The other is

about body scanners and their affect on man’s health

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

194

Figure 2: An interpretation of browser interaction data with-

out assumptions about a user’s search behavior.

as adapted to (Wu et al., 2012). Participants are ad-

vised that each search task should take at least 15 min-

utes.

Data Preparation. For each search task we create

a single browser history by merging the interaction

data of all participants into a single log ﬁle. Combin-

ing the search activity of the participants allows us to

“simulate” a more extensive exploratory search with

multiple starting points, at least one for each partici-

pant. However, we lose the ability to examine individ-

ual search proﬁles, because the resulting interaction

log exhibits characteristics from multiple searchers.

We then create exploration graphs based on the log

ﬁles using two different interpretations, one with and

one without assumptions about the search behavior.

Preliminary Results. Nine participants aged be-

tween 26 and 33 participated the preliminary study.

Figure 2 shows the result of an interpretation for

the ﬁrst search topic without assumptions about the

search behavior, i. e. only the actual navigations be-

tween visited websites are considered in the interpre-

tation. As can be seen, several disconnected graph

components represent the search behavior. Most of

the starting points are created by visiting general-

purpose search engines, like Google. Note that some

participants visited the same websites (red rectan-

gles), so even based on this interpretation some of the

individual explorations are connected.

˜ A semantic interpretation of the same data is

shown in Figure 3. The graph is more connected, be-

cause some search engine results pages are connected

to relevant websites visited before. To calculate the

relevancy of a query we index every website at the

time it is visited in a local search engine. Whenever a

user issues a query we make a lookup to ﬁnd relevant

Figure 3: A semantic interpretation of browser interaction

data incorporating sensemaking patterns during search.

websites in this index for the same query terms. At the

beginning the index is empty. Only the most relevant

(already visited) website is selected to be continued.

In this interpretation it is necessary to add a param-

eter to the interpretation function to set a minimum

relevancy threshold. The example interpretation uses

no relevancy threshold. A sample of that exploration

graph for the ﬁrst user study is shown in Figure 4.

The graph shows an episode of search activity and it

outlines how pre-query browsing (as opposed to post-

query browsing) and search queries can contribute to

the same line of thought during a search process.

START

Figure 4: A sample exploration graph for the user study

using a semantic interpretation. It contains three Google re-

sults pages (G

to G

) and ﬁve content websites (W

to W

The query terms of G

are “solar panel advantages over con-

ventional methods”. The subsequently visited websites ex-

plain pros and cons of solar panels. Website W

, however,

deals with photovoltaics. The query terms of G

are “pho-

tovoltaic vs solar heat”. The websites W

and W

explain

the difference between solar and photovoltaic systems. Due

to the used interpretation function the path G

→ W

is appended to W

, indicated by the dotted arrow.

It is also possible to interpret the log data such

that search queries are appended to the top k relevant

websites at the same time, where k becomes another

parameter for the interpretation. This may express un-

certainty towards the user’s line of thought during ex-

ploration. By appending the query website to multiple

Interpreting and Leveraging Browser Interaction for Exploratory Search Tasks

195

paths it is possible to represent concurrent interpreta-

tions that may contradict each other. For example, a

query may be similarly relevant for websites in dif-

ferent paths. Keeping (some of) these interpretations

may support a later graph analysis or mining, because

more data about the search behavior is available.

6 RECOMMENDATIONS BASED

ON EXPLORATION GRAPHS

In the introductory section of this paper we brieﬂy

summarized exploratory search as open-ended and

multifaceted with unclear goals. Researches often ar-

gue that users in exploratory search often do not know

exactly what search queries to use in order to ﬁnd the

answers they desire. In this section we want to show

how exploration graphs can be leveraged in a recom-

mender system to suggest websites to be visited in

such situations. We believe that the analysis of explo-

ration graphs is an exciting approach to develop new

support mechanisms for enterprise search systems.

We have implemented a web-based search plat-

form which is currently rolled out in a large automo-

tive company as part of a long-term research project.

The overall goal of the project is to make complex

search tasks more efﬁcient and support collaboration

within teams. Our platform uses a client-side data

collection approach. With the help of a browser ad-

don we detect the search actions described in Sect. 3

and send them to a server where they are interpreted

as an exploration graph (see Sect. 4) and stored in a

graph database. The server processes the stream of in-

teraction data online. We create a single exploration

graph for all users. For the recommendations we per-

form graph mining to ﬁnd patterns in this exploration

graph. Domain experts can use a special-tailored

search user interface to query for relevant websites.

They are presented with relevant websites containing

the query terms and related websites (which might

not contain the query terms) based on the structure

and semantics of the exploration graph. This kind of

post-query recommendation is a novel approach and

particularly suited in business environments where

searchers are likely to address similar topics repeat-

edly over time, e. g. in patent search or technology

scouting. Especially, in a scenario where multiple do-

main experts are interested in the same (or similar)

topics, they can beneﬁt from each others’ search ef-

forts.

In order to further improve the quality of recom-

mendations we have added an additional action to

the list of user interactions: the LIKE-action marks

the active website as an interesting resource for the

current search. So, whenever a user issues a query

and browses to a website with relevant content the

user can “like” this website. The system then deter-

mines the path in the exploration graph from a pos-

sible query to this website (also taking into account

the length of this path) and recommends it in future

searches when similar queries are used. Examples for

patterns we examine for our recommendations are:

• Given a website w, ﬁnd all websites W

∗

in the

neighborhood of w that are “liked” by at least one

member of the group.

• Given a search query q, ﬁnd those search engine

results pages W

that are visited based on similar

search queries, and identify all websites W

∗

the neighborhoods of all websites w ∈ W

that

are “liked” by at least one member of the group.

Next, we want to brieﬂy outline two research direc-

tions we want to look into in the future.

Learning to Rank. It is possible to create machine

learning algorithms which modify the relevancy of

any visited website for a given search query based on

the exhibited search behavior using the data of an ex-

ploration graph. The LIKE-action allows us to use

both unsupervised as well as supervised and hybrid

approaches. It is even possible to extract additional

features from websites, e. g. website content, title, or

meta information, and incorporate this into the learn-

ing process. Then, suggestions are made based on

complex data, taking into account the structural and

semantical relationships of observed search behavior.

Learning to Automate. One of the main motiva-

tions to initiate the research project with the automo-

tive partner is the long-term goal to (partially) au-

tomate complex search tasks. Technologies like fo-

cussed crawlers have shown since two decades how

automated search algorithms can reduce the effort to

perform complex search manually. We believe that

using the data provided by exploration graphs we can

further improve the results of automated crawlers.

The advantage of training such a crawler based on

real exploration graphs is that such an approach is

ﬂexible with respect to the search topic and the indi-

vidual traits of the searchers. For example, a domain

expert can perform a complex search task manually

to provide an initial starting point for the learning al-

gorithm. Once, the algorithm learned to explore the

topic it can recommend relevant websites based on its

own explorations. These recommendations are then

visited (and further explored) by the domain expert,

again providing data for the algorithm to learn from.

Over time the topic will be explored increasingly ex-

haustive.

It is also possible to use methods from the ﬁeld of

transfer learning, e. g. to 1) apply a learned model

ICEIS 2017 - 19th International Conference on Enterprise Information Systems

196

of one topic to another topic for the same expert, or

2) apply a learned model of one domain expert to an-

other expert. With the help of exploration graphs we

believe it is possible to develop more elaborate algo-

rithms that are able to automate complex search tasks.

The result of this learning goes beyond “learning to

rank”-approaches described earlier, because it has to

take into account the utility of information. That is,

an automated search agent trained with data about

search behavior of human domain experts should sug-

gest only highly relevant and novel information.

7 CONCLUSION

This paper presents a hybrid log-based and observa-

tional approach for modeling search behavior. We

formulate the problem of interpreting search behav-

ior based on browser interaction logs and introduce

the idea of an exploration graph to model transitional

and semantic relationships during search. These in-

terpretations can help to keep valuable information on

how users explore an information space, if reasonable

assumptions about the search behavior can be made.

With the help of a user study we outline how seman-

tic interpretations compare to interpretations without

assumptions. Interpreting a user’s interactions during

search in an exploration graph may be key to various

new investigations, e. g. how users interact in groups

to fulﬁll a certain research task. Finding meaningful

interpretations becomes a new challenge in the analy-

sis of interaction logs. We also show how leveraging

the data inherent in exploration graphs can be used in

recommender systems to make search tasks in busi-

ness settings more efﬁcient. Since the quality of such

recommendations depends heavily on the quality of

the interpreted exploration graphs it is important to

put a lot of effort into creating meaningful interpre-

tations of search behavior in the ﬁrst place. We be-

lieve that leveraging data about exploration graphs is

a promising approach to tackle new research direc-

tions and produce highly innovative support systems,

especially for professional searchers.

REFERENCES

Ageev, M., Guo, Q., Lagun, D., and Agichtein, E. (2011).

Find it if you can: a game for modeling different types

of web search success using interaction data. In Proc.

of the 34th Int. ACM SIGIR Conf. on R&D in IR, pages

345–354. ACM.

Bates, M. J. (1989). The Design of Browsing and Berryp-

icking Techniques for the Online Search Interface.

Online Inform. Review, 13(5):407–424.

Card, S. K., Pirolli, P., Van Der Wege, M., Morrison, J. B.,

Reeder, R. W., Schraedley, P. K., and Boshart, J.

(2001). Information scent as a driver of web behav-

ior graphs: Results of a protocol analysis method for

web usability. In Proc. of the SIGCHI Conf. on Hu-

man Factors in Computing Systems, CHI ’01, pages

498–505.

Gwizdka, J. and Lopatovska, I. (2009). The role of sub-

jective factors in the information search process. J.

Assoc. Inf. Sci. Technol., 60(12):2452–2464.

Hassan, A., Jones, R., and Klinkner, K. L. (2010). Be-

yond dcg: User behavior as a predictor of a successful

search. In Proc. of the 3rd ACM Int. Conf. on Web

Search and Data Mining, pages 221–230. ACM.

Knight, S. and Spink, A. (2008). Toward a web search in-

formation behavior model. In Web Search, pages 209–

234. Springer.

Marchionini, G. (1997). Information Seeking in Electronic

Environments. Number 9. Cambridge university press.

Marchionini, G. (2006). Exploratory search: From ﬁnding

to understanding. Commun. ACM, 49(4):41–46.

urnberger, A., Stange, D., and Kotzyba, M. (2015).

Professional collaborative information seeking: On

traceability and creative sensemaking. In Semantic

Keyword-based Search on Structured Data Sources,

IKC 2015, pages 1–16, New York, NY, USA.

Springer-Verlag New York, Inc.

White, R. W. and Drucker, S. M. (2007). Investigating be-

havioral variability in web search. In Proc. of the 16th

Int. Conf. on WWW, pages 21–30.

White, R. W. and Roth, R. A. (2009). Exploratory Search:

Beyond the Query-Response Paradigm. Synthesis

Lect. on Information Concepts, Retrieval, and Ser-

vices.

Wildemuth, B. M. and Freund, L. (2012). Assigning search

tasks designed to elicit exploratory search behaviors.

In Proc. of the Symp. on HCI and IR, pages 4:1–4:10.

Wu, W.-C., Kelly, D., and Huang, K. (2012). User evalu-

ation of query quality. In Proc. of the 35th Int. ACM

SIGIR Conf. on R&D in IR, pages 215–224. ACM.

Interpreting and Leveraging Browser Interaction for Exploratory Search Tasks

197