Computer Science and Information Systems ?(?):??–?? DOI: N/A
ResXplorer: Revealing Relations between Resources for
Researchers in the Web of Data
Laurens De Vocht1, Selver Softic2,
Ruben Verborgh1, Erik Mannens1, and Martin Ebner2
1iMinds - Data Science Lab, Ghent University,
Sint-Pietersnieuwstraat 41 B1 9000 Ghent, Belgium.
2Institute for Information Systems and Computer Media, Graz University of Technology,
Inffeldgasse 16c 8010 Graz, Austria.
Abstract. Recent developments on sharing research results and ideas on the Web,
such as research collaboration platforms like Mendeley or ResearchGate, enable
novel ways to explore research information. Current search interfaces in this ﬁeld
focus mostly on narrowing down the search scope through faceted search, keyword
matching, or ﬁltering. The interactive visual aspect and the focus on exploring
relationships between items in the results has not sufﬁciently been addressed before.
To facilitate this exploration, we developed ResXplorer, a search interface that inter-
actively visualizes linked data of research-related sources. By visualizing resources
such as conferences, publications and proceedings, we reveal relationships between
researchers and those resources. We evaluate our search interface by measuring how
it affects the search productivity of targeted lean users. Furthermore, expert users
reviewed its information retrieval potential and compared it against both popular
academic search engines and highly specialized academic search interfaces. The
results indicate how well lean users perceive the system and expert users rate it for
its main goal: revealing relationships between resources for researchers.
Keywords: Exploratory Search; Social Media; Digital Libraries; Research Collab-
oration Tools; Science 2.0; Web of Data; Linked Data
Peer-reviewed research publications as well as related metadata from bibliography archives
are widely available on the web. They offer a vast amount of information on related pub-
lications and can facilitate suggesting new contacts, collaborators, and interesting custom
events. Usually the platforms supporting this information exchange expose a Web API
that allows access to the structured content, or the information is present as Linked Data.
Social media, such as Twitter and Facebook, are often used during scientiﬁc events.
Researchers use them to comment and discuss about each other’s work, and to exchange
research related materials . They also use Web collaboration tools like Mendeley
or ResearchGate to present their scientiﬁc work. Such academic social networks have
become wide-spread and can have millions of regular users . Resources for research
are not always easy to explore: they rarely come with strong support for identifying,
linking, and selecting those that can be of interest for further investigation.
2 Laurens De Vocht et al.
Personalized adaptation of the Web to the needs of researchers is the main vision
of Research 2.0 . Research 2.0 depicts using such Web 2.0 tools and principles in
scientiﬁc research and learning. It is an application ﬁeld of “Technology Enhanced Learn-
ing” which covers the entirety of learning and research with use of new media. It is an
approach to science that maximally leverages information-sharing and collaboration tools
and emphasizes the advantages of increased online collaboration between researchers.
Therefore, we developed a personalized interactive Semantic Search environment based
on our search infrastructure and data from diverse open Linked Data repositories including
scientiﬁc publication archives and social media. The purpose of this research is to offer
a set of tools and services which researchers can use to discover resources as well as to
facilitate collaboration via the web. These tools and services are considered as mash-ups,
APIs, publishing feeds, and specially designed interfaces based on social proﬁles [32,45].
These tools and services are in line with the principles of Research 2.0.
The ﬁrst use cases, data architectures, mash-up concept studies and prototypes on
aligning the social web with semantics in the context of research, we introduced in 2011 .
The data modeling concepts were discussed in [38,39] while we investigated the back-end
used for ResXplorer, a framework for discovery of chains of links between resources .
The aligning and matching of research related semantic resources was the main scope of
our work on dynamic alignment of scientiﬁc resources such as web collaboration tools
and digital archives [15,40]. We introduced the ﬁrst prototypes of the ResXplorer search
interface at conferences in late 2013  and 2014 . One of the ﬁrst live versions
participated at the Semantic Web Challenge 2013 . The goal of all these publications
was to evolve the concept, demonstrate the interface and visualization, trigger discussion
and gain insight on the exploration workﬂow.
We investigate how researchers can ﬁnd the information they need to gain insight
about the people, conferences or publications they want to explore, which can be formu-
lated under the form of following research questions:
–How effective can the execution of context-sensitive search tasks in semantically
annotated digital archives be facilitated by revealing relations between resources (i.e.
adequately addressing the researcher’s intent) and leveraging the links to social media
and web collaboration tools?
–What is the contribution of exposing resources part of a revealed relation that the
researcher was not aware of beforehand?
–For which kind of tasks does an increasing amount of user actions by the researcher
positively inﬂuence the relevance and precision of the search results?
–Does the underlying approach excel in revealing relations in a certain context com-
pared to state of the art? At what cost does it come in terms of required user actions
and precision for more traditional search tasks?
We will test the following hypotheses:
–Interacting with the search results reﬁnes and improves the result set because inter-
action with the result set makes the information contained in the initial search query
–Each iteration enables the researcher to deﬁne more and more targeted queries, lead-
ing to more precise results.
We propose ResXplorer3to address these research questions by focusing exactly on
exploring chains of links between these resources. ResXplorer is one of the ﬁrst practical
solutions combining the social web and the semantic web in an interactive search environ-
ment that visually emphasizes and represents the search context and results. We focus on
showing how it is an example for the use of interactive visualizations to enable knowledge
discovery in Linked Data, which can be invaluable to researchers. One of the use cases
we support, focuses on the end-user usability of semantically enriched researcher proﬁles.
In this use case, our prototype “ResXplorer” shows relations between researchers based
upon the semantic analysis of researcher’s tweets and aligned with information about
conferences and proceedings, users are presented how they are (indirectly) related based
on their institutions, visited locations, and conferences they contributed to. As a measure
of usability we investigated the ability of our search interface to support the construction
of a good cognitive model of the underlying data and the relations within the data. Finally,
we measured the effectiveness and productivity of the interface by checking to which
extent end-users carry out knowledge-intensive and analytical tasks.
2. Related Work
The focus of this section is to discuss and pinpoint the efforts done in the ﬁeld of interfaces
for search and in the ﬁeld of discovery of relevant content for researchers.
2.1. Alternatives without semantic graph in the back-end
Approaches not relying on semantic graphs typically recommend multi-faceted infor-
mation to discover for example research, popular citations. One of these methods is
behavioral pattern discovery in research (e.g., publishing papers, citing papers): the focus
there lies in to understand user behavior and discover patterns to apply to for example
web search, recommender systems and advertising.
Traditional methods usually consider the behaviors as simple user and item connec-
tions, or represent them with a static model. However, in practice, user behavior is dy-
namic: it includes correlations between the user and multiple types of objects and evolve
over time. An example for the latter is the Flexible Evolutionary Multi-faceted Analysis
(FEMA) framework for both behavior prediction and pattern mining . FEMA utilizes
a ﬂexible and dynamic factorization scheme for analyzing human behavioral data se-
quences, which can incorporate various knowledge embedded in different object domains.
Related to this is the question on how possible author-keyword associations evolve
over time. Rather than relying on graphs, some approaches manage to implement and
apply this using tensors . Working with graphs does not necessarily mean that they
always require the entire knowledge base to be semantically annotated. Graph algorithms
such as for clustering have been found to be a good way to identify experts in large
knowledge bases where the semantics are being introduced only in the last stage of the
analysis for uniquely identifying resources with Uniform Resource Identiﬁers (URIs). For
example, related author proﬁles, which are initially parsed as unstructured, tree- or tab-
ular structured data, can be dynamically clustered together via an author disambiguation
4 Laurens De Vocht et al.
2.2. Interfaces for Research based on Linked Data
Our approach focuses on ﬁnding relationships between resources. It is a distinct example
solution and implementation of an adaptive and intelligent web-based system  yet does
not currently aim to compete in terms of feature completeness. A review on related search
interfaces for science leads us to a few working solutions worth mentioning, despite the
huge amount of published Linked Data especially publications meta data. A number of
working solutions worth mentioning are: RKB Explorer, Faceted FBLP, and Bibbase.
RKB Explorer4, a visual browser which originated from the ReSIST5network
of excellence, which unites within its realm many sources of scientiﬁc data. This visual
browsing interface is based on categorised pre-selection and focuses on people, organ-
isations, publications, and courses and materials. The search always centers around the
selected category which makes the context based browsing less ﬂexible but focused.
Within the visualisation RKB Explorer evaluates relations of the ﬁrst degree. In compar-
ison to RKB Explorer our approach is more user and search centric rather than concept
and context centric. In our interface, a user proﬁle affects the pre-selection of search
results. Users can conﬁgure the search context by executing searches for resources or by
expanding one or more resources.
Another advanced research related effort is Faceted DBLP search6. The search ap-
proach in this case resides on DBLP++ data which enhances DBLP with additional key-
words and abstracts as available on public web pages. It integrates facets on Time, Venues,
Publications Years and Authors and delivers the results in various formats. These formats
include: BibTeX, regular web pages, DOI identiﬁers, or RDF. Faceted DBLP offers a
good ﬂexibility in ﬁltering and narrowing down the results as well as implementing basic
syntactic query expansion based upon single word and whole phrase in an anonymous
way. Retrieval is done by classic search engines and result selection is done by ranking
without any possible relation to the user proﬁle.
BibBase7 has an interface to leverage the personal publications into the Web
of Data and integrates the retrieval of author publications with a small sample from
Mendeley8, DBLP9and Zotero10.
2.3. Systems for Semantic Search
In recent years signiﬁcant efforts have been made in semantic technologies and regarding
the Semantic Search. An overview on related work in this area can be found in [46,47]. Se-
mantic Web search engines like Hermes  have been developed which regarding initial
intention and application is very closely related to the use case of our work. Whereas
Hermes tries to translate the keywords into structured queries, our approach expands
results using paths within the connected Linked Data graphs. ResXplorer uses the user’s
social proﬁle as additional context information for these graphs.
ResXplorer shares the goal of search, data about research publications, and intended
audience with Google Scholar (GScholar)11, Microsoft Academic (MA) Search12 , ARnet
Miner 13 and Falcons 14. Besides Hermes also Falcons [7,6], Watson , SWSE
 and Sindice  were developed for retrieving semantic data from the Web. These en-
gines primarily rely on keywords as starting positions for deﬁning and specifying queries.
Hermes, Falcons and SWSE, also support more advanced querying capabilities, including
basic SPARQL graph patterns. In general, the semantic matching frameworks within these
semantic search engines reside on the approach of matching graph patterns against RDF
data with, if applicable, inference run against the RDFS/OWL ontologies. This kind of
semantic matching mechanism is also widely implemented by a wide range of RDF stores.
2.4. Vocabularies to annotate Social Media
Common vocabularies to annotate social media as Linked Data are: Friend of A Friend
(FOAF)15, Semantically Interlinked Online Communities (SIOC)16 [3,2] and Dublin Core17
. FOAF describes the user proﬁles, their social relations and resources. SIOC is mostly
combined with FOAF and Dublin Core for creating instances of web entries like blogs,
microblogs, mailing list entries, forum posts, along with other entries from Web 2.0
platforms [43,38,14]. Passant et al. improved mapping social proﬁles with related content,
such as via interlinking tags [33,34,35].
3. Interacting with Research and Social Media Data
Our approach collects and uses information from resources already explored by other,
more experienced, researchers. This is especially interesting for cases when looking for
the next practical piece of information or when trying to ﬁnd a solution for a problem
that requires out-of-the-box thinking (e.g., when forming the exact search query requires
background knowledge of a domain unfamiliar to the researcher). The interaction and
models in Figure 1 show how the researchers use this method.
Researchers can deﬁne and select their intended search goal over several iterations.
When users are looking for new leads, they get an overview of possible objects of interest
(like points of interest on a street map) by having their activities and contributions linked
on social media and other platforms such as their own research publications proﬁle.
3.1. Data Model
The Data Model has two spaces. It has a Linked Data space and an Entity space. The
former is the representation of the data loaded into the model and the latter are the entities,
each having a URI, a label, a type and a description consisting of one or more Linked Data
triples. In this section we describe the two types of data that we model: Research Data and
Linked Data extracted from social media.
6 Laurens De Vocht et al.
Fig. 1: The information researchers share via the Web services of research collaboration
tools and social media is structured and transformed to RDF and interlinked with Linked
Open Data. The resulting entities in the data model form the base for the semantic model.
This process is outlined in Section 3.1. When researchers search, they interact RV: I
presume “indirectly”? with the semantic model which we detail in Section 3.2.
Research Data Research data is described as Linked Data using state-of-the-art vocab-
ularies. We model research data with respect to their usage and wide popularity within
the Semantic Web community, as well as to their applicability for the proposed use case.
One of the modeling domains of interest are scientiﬁc events and their relatedness to
bibliographical archives. The “Digital Bibliography and Library Project” (DBLP)18
provides bibliographic information on major computer science journals and proceedings
and indexes more than 2.3 million articles. Besides it also has many links to home pages
of computer scientists. The Conference Linked Data (COLINDA) data set resolves this
connection. COLINDA describes conferences using the Semantic Web for Research Com-
munities (SWRC)19 ontology . Especially important for this decision was that DBLP
Linked Data also applies this ontology to describe its resources.
Linked Data from social media We created an annotated set of extracted conference
hash tags mentioned in tweets of researchers which would be associated with correspond-
ing tweets and which can be used for further mining tasks like label based matching of
scientiﬁc events in Linked Data sets e.g., COLINDA, DBLP. The motivation for linking
data from social media: ‘social data’ as such, is manifold:
Link discovery It allows detecting and creating links between the users and the data they
Timely context It enforces a timely and personalized context to the search.
Relationships It adds additional relationships between users and resources that are con-
tained in the more static data and potentially introduce additional references to other
Linked Open Data.
In our case, besides persons, locations, conferences and scientiﬁc publications, the
researcher oneself is an important resource for the context of our search scenario. To
triplify the personal proﬁle of the researcher, we have used the FOAF vocabulary accord-
ing to the common method within the community to describe the proﬁle of a person or
agent [18,34]. We obtained Twitter-related content from our tweets cache Grabeeter20,
which contains very useful set of more than a million tweets from couple of thousands
of users in academia and the Twitter API21 via a self developed “Semantic Proﬁling
Framework”. The necessity of creating our own framework emerged from the lack of
solutions which turn microblog posts into RDF which met our needs. There have been
some attemps like Semantic Tweet22, Smesher23 or SMOB24 but they only partially met
our needs. The other reason why we created own solution built on prior insights of
Semantic Web community is that in this way we were also able to integrate the tweets and
proﬁles from Grabeeter. The platform should be capable to integrate some other research
related social platforms like e.g., Mendeley over the user accounts information. In this
way, for instance, tweets can be aligned to publications at Mendeley in cases where users
own both Twitter and Mendeley accounts.
To model micro blog content we used SIOC combined with the Dublin Core vocabu-
lary to model the tweet text and information about it, such as post date and link to creator
as recommended in earlier works on this topic [3,2]. For creator proﬁles we used FOAF
to model and preserve the information. We used the Mendeley API25 to enhance creator
proﬁles, using the data from corresponding Mendeley proﬁles, with links to publications,
scientiﬁc events, and persons found there.
3.2. Semantic Model
We introduce Research Objects , which center and group reﬁned entities of extracted
and integrated data in the Data Model and represent:
–Events: scientiﬁc conferences, seminars and/or lectures
–Publications: articles, reports, tutorials and/or posts
–Locations: both real-world and online (web pages, webinars)
–Concepts: topics, categories and/or classiﬁcations
Research objects enable and facilitate the use of research related information. The
metadata that describes research objects facilitates searching and retrieving them.
23 http://bnode.org/blog/2009/02/16/devx-article- about-semantifying- and-sparqling-
8 Laurens De Vocht et al.
Deﬁning Research Objects A single Research Object can contain links to and infor-
mation about an online tutorial, details about a seminar, links to fragments of related
papers and tutors or people who are known to have contributed to the entities of this
speciﬁc object. Researchers deﬁne a search query for their research and have it parsed
by our system for identiﬁcation in terms of the Semantic Model. Based on the research
goal, we select the found research objects that are most closely related, after they have
been reﬁned from their representation in the Data Model. Publishing Linked Data into
the cloud, however, does not ensure the required reusability, but the use of research
objects in a semantic model should provide the reproducibility that enables validation
of results . We align the entities present in the Data Model with the registered activities
of researchers by providing their proﬁles and feeds of social media. Researchers generate
those by sharing and monitoring online activities such as blogs, (micro)posts, tags, shares
and other resources.
Searching Research Objects We center searches around several research targets that
a researcher wants to relate with another. We also combine related resources based on
common links they share, such as being related to and containing more information about
a Research Object. The users generate their own views by exploring and searching among
the Research Objects in the model and can share or compare those with other researchers
or earlier searches. All those views together lead to a personalized environment. This will
boost interaction with and grouping of similar views and objects to bigger packages that
ultimately lead to the discovery of even more relations. We customary map all objects
for users based on their “researcher proﬁle”. We extract the research proﬁle based on the
content researchers monitor on social media or the resources they have shared over it.
Most of the researchers today own a proﬁle in a scientiﬁc or common social network like
Twitter and Facebook26 or on research related platforms like ResearchGate27, Mendeley,
or Google Scholar.
4. Interactive Search
In each search session users can combine keyword-based disambiguation combined with
visual reﬁnements through expansion queries on the semantic entities we recognize as
facets. The idea behind their usage as facets is to offer always and at each step a complete
understanding of why certain results are showed. We do not want to let the algorithm
assume things about researchers preferences. Since it is meant to be an exploratory search,
the point is to involve the researchers on the base of input-output principle in a guided
approach through facets and three dimensions: shapes, colors and size. They choose what
they want to see and get that result delivered. As a parallel process in the back-end,
the engine discovers additional relations between the search results and presents them
as alternatives to the already acquired information. In this way we automatically expand
or narrow down the facet range available to the researcher. This leads the researchers
through the data by offering them at each point in time exploration and involvement of
new and already found items into the search. This section ﬁrstly describes the front-end
and back-end, and illustrates the dynamics of the interactive search from a users point of
view using an example.
4.1. Remarks on signiﬁcance of social media in interactive search
Social media Linked Data primarily contributes strongly to result and search personal-
ization. In back-end it acts as nexus between conference and digital archive repositories
(here COLINDA and DBLP) through the links pointing to this sources from social me-
dia proﬁle data of the user. Further, it enhances the results with additional event and
publication information from users’s Twitter or Mendeley account. Since not every user
has both accounts or any of them the involvement of this data happens through explicit
user’s authorization to ResXplorer to include his Twitter or Mendeley account into search
process. The data is then extracted and brought into semantic form as well interlinked
in background to offer a starting search conﬁguration adapted to user. Missing the social
media links primary results in shortages respectively insights on relations between the
authors who share same scientiﬁc events and publications which are identiﬁable only via
social content their accounts which are not necessary included in COLINDA or DBLP
The decision process during the search is supported by a real-time keyword disambigua-
tion to guide the researchers in expressing their research needs. We do this by allowing
users to select the correct meaning from a drop down menu that appears below the search
box. Presenting candidate query expansion terms in real-time, as users type their queries,
can be useful during the early stages of the search .
Researchers can deﬁne and select their ‘intended’ search goal over several iterations.
A combination various resources is then presented to the researchers. In case they have
no idea which object or topic to investigate next, they get an overview of possible objects
of interest (like points of interest on a street map). Researchers deﬁne a search query for
their research and have it parsed by our system.
Based on the ability of humans to rapidly scan, recognize, recall images and detect changes
in size, color, and shape, we aim to enhance the guidance of users during their search by
using several visual aids of which the three most visible are:
1. Shape: We group sets of types in large groups and represent them using a shape.
Types that cannot be assigned a group are grouped in a category ’Miscellaneous’.
The shapes help the user to distinguish between the types of offered results
2. Color: Every entity has a type and associated unique color. For a certain result set
the user gets an immediate impression of the nature of the found resources. Figure 2
depicts two different objects related to other objects and therefore have a different
shape and size. On the left of the search interface there is a legend explaining the
researcher the meaning of shapes and colors.
3. Size: Each entity is ranked according to novelty and relation to the context and enlarge
those that should attract attention from the researcher ﬁrst, this is shown in Figure 3.
The novelty quantiﬁes the degree of being new, original or unusual. Particularly in this
context it entitles resources that are remarkable and differ from the others because
of their direct relations with neighbours or their semantics (in terms of occurring
predicates). A goal of the search is to explore information not seen before which
makes it difﬁcult to deﬁne an accurate search goal. Besides allowing to search speciﬁc
entities, our visualization facilitates exploratory browsing. This is particularly useful
when information seeking with unclear deﬁned search targets .
10 Laurens De Vocht et al.
Fig. 2: Different shape and color to dis-
Fig. 3: Different sizes to guide the user’s
Figure 4 shows how researchers can track the history of their search: the explored relations
are marked red and clearly highlight the context of a search. This is a good example of how
our system adapts to the users and their environments. It shows one of the ways how to
build a model of the goals and knowledge of an individual user, and the model is used
throughout the interaction with the user. Researchers can click on a list of resources they
have searched to focus the visualization. A screencast of the search interface is available
online28. In this screencast, we show how researchers interact with the search interface
and the above described visualization.
Fig. 4: A red line marks the explored relations in the visualized search context.
The back-end supports the search process for tools like ResXplorer with two main tasks:
discovering links between two resources and the ranking of found links and resources as
shown in Figure 5. We ﬁnd paths between resources for discovering links to match the
seed input given by the researcher. With the delivery of ﬁrst results, our engine expands
the query and enhances the context by pathﬁnding and neighbours resolution within the
“Everything is Connected engine” (EiCE) . It uses the ‘distance’ to the ﬁrst query as a
measure for ranking the result. The EiCE is used here to compute heuristically optimized
minimum cost paths between pairs of researchers, publications, conferences or mixed
pairs. The heuristics take into account the rarity of resources to avoid common resources
(that have many in- or outgoing links) and the semantic relatedness between resources.
Each time a user adds another resource to the results, the visualized path between the
resources takes these factors into account.
Indexed Linked Data
Everything is Connected
Online Research Publications
(available as Linked Data)
Fig. 5: Applications like ResXplorer use the Everything is Connected engine for ﬁnding
relations between resources.
4.4. Example Illustrating the Dynamics
Each search starts within the search interface where a user can either login or query
anonymously the Semantic Search Engine. Our search interface distinguishes between
two types of queries: a query which consists of several keywords as seeds given as
input and a proﬁle-driven query, used as preset for further search, driven initially by user
Except for the ﬁrst step, the querying paradigm applies to the personalized search as
well. The query in the ﬁgure illustrates the common case where a researcher enters the
search process by entering simple keywords and tries to resolve the context of “ﬁnding
useful resources from a certain conference”.
1. One searches for a speciﬁc conference “Linked Data” and articles related to “WWW2012”.
Firstly, as Figure 6 shows, the visualization focuses ﬁrst on the logged in user upon
which the user can choose to expand on of the neighbouring resources.
We note that the user changes the focus of the result view by clicking on a resource:
the resource encircled with a mixed line. Within the simpliﬁed query progression
process, entered keywords are ﬁrst mapped towards the entities and properties in the
12 Laurens De Vocht et al.
Fig. 6: User expands a direct neighbor after ResXplorer focuses on logged in user (encir-
cled with dashes).
2. As a search result the engine delivers ﬁrst set of links for each keyword entered, such
as in Figure 7 for “Linked Data”.
Fig. 7: User searches for “Linked Data” and ResXplorer reveals the chain of links between
the selected document (encircled with dots) and the user.
3. If available, the system also delivers the types of entities discovered in index. When
the user searches for the next keyword “WWW2012”, relations to other already visu-
alized resources are exposed as indicated in Figure 8.
4. By entering the location, for example “Germany”, one could narrow further the focus
of the context by location. Each time a combination of various resources is visualized,
the application suggests new queries: they are generally most useful for reﬁning the
system’s representation of the researcher’s need.
In case they have no idea which entity to focus on or what topic to investigate next
they get an overview of possible entities of interest, like points of interest on a
street map. By proﬁling their activities and contributions on social media and other
platforms such as their own research publications, the afﬁnity with the proposed
resources is enhanced.
5. With each further iteration the user can choose either one of two actions:
– Query Expansion: The user expands the query space by clicking the results
retrieved by initial keyword based search.The resolution of results is based upon
the properties of Linked Data instances like rdf:label,owl:sameAs,rdf:seeAlso,
Fig. 8: After the user focused on a common resource of both and searched for
“WWW2012”, ResXplorer reveals relations to the selected conference “WWW2012”
(encircled with dot-dashes).
dc:title,dc:spatial or dc:description. Those properties have been used in genera-
tion of Linked Data instances to preserve conference shortcuts (e.g. WWW2012),
point to link of proceedings of a conference or ,to connect alternative link about
it, as well to literally describe the venues of scientiﬁc events.
– Additional Query Formulation: Additional query expansion happens either
through adding further keywords as well as through keyword combinations al-
ready entered where the back-end tries to deliver additional results based upon
connection paths between the resources. What happens in return is that the engine
tries to identify the terms that have been searched in the result space. In cases
when they can be resolved by a Linked Data instance, the algorithm continues
step by step looking via links to the neighbors of the instance to ﬁnd a path to
other terms identiﬁed by the engine as well. After a certain number of steps (here,
seven) it terminates if it is unsuccessful.
We consider non-Linked Data researchers to be typical lean users of our system for the
purposes of the Research 2.0 case and Linked Data specialists as expert users, especially
because of their familiarity with the underlying graph structure. The primary target group
of users of our system are non-Linked Data researchers, and in this particular research
2.0 use case, scientiﬁc researchers. They interact with research objects and not the more
complex underlying Linked Data model. Another group of users which we must bear in
mind are domain experts, as they are likely to have a very good understanding of data
structure and content in their domain, and bring this knowledge to guide both browsing
14 Laurens De Vocht et al.
research and targeted searches. Based on the experience of our previous work ,
we selected experts and researchers in computer science and digital media as test group
representatives. This group of people is also our target audience.
Exploratory search represents a cognitively very intensive activity. Therefore conduc-
tion of searches should be possible with minimal interruptions. According to White et
al. [52,51] : “Techniques such as questionnaires and interview techniques can be valuable
tools, but one must be careful to include them in the experiment in such a way as to
not interfere with their exploration”. We evaluated the tools in two ways: lean user tests
and expert user reviews. These methods give us insight in how the users perceive the
tools and show us quickly potential bottlenecks . They also deliver us insight on how
precise our solution performs in comparison to the existent state of the art solutions of
industry as well as academia. All of the compared solutions target the same audience.
They differ in implementation and interface design, but more importantly, they have more
or less valuable stock of users. The choice of applied evaluation methodology was made
by applying relevant aspects out of already existing achievements in this ﬁeld introduced
in [27,50,20] and adapting them to our speciﬁc use case. Since we want to offer a solution
for research and learning purposes but also for wider community of users, a user centered
methodology plays a decisive role in our evaluation process.
We have developed the interface for search in a research linked data knowledge base
combining the latest Linked Data technologies with an advanced indexing and path ﬁnd-
ing system. We build our implementation upon our earlier work using the “Everything
is Connected” engine (EiCE)  EiCE and Web 2.0 technologies (such as JQuery and
5.3. Lean User Study
We have evaluated the effectiveness and productivity of our environment in this user study.
The conducted user study focused on ResXplorer aimed at measuring the effectiveness,
productivity and impact of the features of our environment. Sixteen test users were se-
lected to participate to the observation and they received no information about the tool in
advance. According to Faulkner  this number of test users is enough to reach nearly
high level of certainty for ﬁnding the most of the existing usability problems.We asked
the test-users to participate in a controlled experiment - to ﬁnd a relevant person to contact
or a conference to attend. They were asked to execute speciﬁc assignments and afterward
to ﬁll in a questionnaire with qualitative questions about their experience during the test.
Furthermore we conducted a A/B testing survey among the users to measure the impact
of the personalization and the EiCE features.
Controlled Experiment. During the controlled experiment, users were asked to think
aloud and their actions were recorded while an evaluator observed the comments and
took notes. Each test took about 30 to 45 minutes. Their assignment was as follows:
Assignment The users had to mark all found resources relevant to them. Then, users
could choose between three actions: searching, adding top related resources; this is
done through disambiguated keyword based search on topics knowingly related to the
initial search term e.g. choosing Tim Berners-Lee as initial keyword and WWW 2013
next related keyword in search, or expanding neighbors of found resources. In the last
case they could chose between direct or indirect neighbors of the centrally focused
node in the visualization. A ‘top related’ resource is the resource directly linked to
the node in focus (centered) that shares the most common links with it.
Effectiveness measures how often a displayed result (R) related to a resource was
marked relevant by the user (M).
Each action that delivered new resources to the result set resulted in an increase of quality
of the result set.
Productivity measures this increase. The quality of a result set is the number of marked
relevant resources compared to the total number of visualized resources. Productivity P
measures the increase of effectiveness Ekafter each test-user set of search actions in
where Ekis measured effectiveness after the action ak.
rRatio indicates the impact of the newly visualized nodes on the existing
visualized resources. We get this ratio by dividing productivity by effectiveness. We learn
from this ratio how beneﬁcial the last action was to the search as it indicates the percentage
of the relevant nodes contributed by the last search action.
The data in Figure 9 shows that adding a top related resource was not done often by
the users and added only a couple of resources to the result set. However, it proved to
be the most effective action as the users marked 13/26 (50%) of the visualized resources
relevant. The data in Figure 9 also shows an increase of 12% in productivity in based on
an average over all test users. Adding top related resources resulted in a result set that
contained 12% more relevant nodes as before adding top related nodes. The E/Pr ratio is
As Figure 10 indicates, searching for a resource was the most productive of all type
of actions (+25%) and has an E/Pr ratio of 81% (0.25/0.31 ). This is remarkable as the
user action effectiveness of searching is much lower than adding a top related resource
on average over all test users. Adding top related resources resulted in a result set that
contained +12% more relevant nodes as before adding top related nodes, even though it
has higher effectiveness (50%). This means that the impact of each added resource when
16 Laurens De Vocht et al.
Fig. 9: Overview of user actions used to visualize new resources to measure the effective-
ness of the actions
searching is much bigger, because the quality of the result set was not relatively high at
the moment users decided searching. On average less than 31% of the resources, which
would result in an increase in productivity if of the newly added resources at least 31%
was marked relevant according to users.
Fig. 10: User Action Effectiveness and Productivity results
The effectiveness of expanding resources 53/166 (32%) is about the same as searching
for a resource 54/174 (31%). As the user actions resulted in about as many new resources
in the case of searching and expanding, this is a very reliable comparison. Expanding
the direct neighbors is the most productive (+6%) expansion. Expanding further related
neighbours retains the quality of the result set and barely impacts it, but the productivity
is still positive (+1%). Users mostly expanded direct neighbours, which leaded to 94 new
resources compared to 72 expansions of indirect neighbours.
Feature Impact Survey. We conducted a survey among the users to measure the impact
of the two most important features of ResXplorer: personalization (using social media
data) and pathﬁnding (with EiCE). We presented the users screenshots of result sets
in ResXplorer in A/B pairs, achieved with one of the features enabled or with both
disabled/enabled. They were asked to rate on a Likert scale from −3 to +3, thus from
more towards A to more towards B, which result set they preferred without knowing
which one had which feature(s) enabled.
Fig. 11: Impact on the result set relevancy of ResXplorer features according to users.
Figure 11 shows disagreement or no clear positive impact for simple queries when
EiCE is enabled and a rather negative impact when personalization is enabled for simple
queries. The results are more positive where more than 60% of the users agrees that for
complex queries the results when using the EiCE are preferred. For personalization the
ratio is 45% positive against 36% negative, the bias is less positive here, but clearly better
than the case for personalization with simple queries. When looking at enabling both fea-
tures vs. disabling both features, nearly 66% prefers the results with both personalization
en EiCE enabled and 56% in case of the simple queries.
The results are in line with previous studies we did on: (i) the dynamic alignment
of social data with conference publication data ; and (ii) the usability study of the
“Researcher Afﬁnity Browser” . All these ﬁndings back the emphasis at several places
in the paper on the signiﬁcance of pathﬁnding and social media (personalization) in
interactive exploratory search.
5.4. Expert User Reviews
As already previously applied  for exploratory search we used a task based approach
to obtain expert user reviews. The goal of the reviews is to compare ResXplorer against
industry reference academic search interfaces and related academic projects, the state-of-
the-art (SOTA). Two researchers – search interface experts – independently reviewed the
performance of each of these search interfaces. They were familiar with all of the tools
beforehand. We selected a set of six representative tasks supported by these systems for
the reviews in Table 1.
We designed the search tasks optimized for the SOTA search engines and for ResX-
plorer and they are either simple (e.g. single fact or source) or complex (combinations of
facts and sources). We outlined the a priori, thus before presenting it to the expert users,
expected suitability of these tasks in Table 2.
In each of these tasks the experts had to indicate after each interaction by either a click
or text input, how many relevant results they found. Their actions were recorded so that
we could count the total number of actions for each task and the number of results after
For each of the tasks we measured the average precision (between 0 and 1) and the
efﬁciency (expressed as number of actions needed).
18 Laurens De Vocht et al.
Table 1: List of tasks executed by the expert users.
T1Find proof that Chris(tian) Bizer is an author.
T2Find out three different people that know or are known by the person in T1 (e.g. co-authors).
T3Find out three different kinds of relations between the person in T1 and Chris(tian) Bizer.
T4Find three different conferences on the subject Artiﬁcial Intelligence.
T5Find at least two people that have a paper included in the proceedings in two consequent editions of the
WWW (World Wide Web) Conference.
T6Find: (i) at least one publication that was presented in 2011 in a WWW workshop (co-)organized by
Tim Berners-Lee (e.g. LDOW - Linked Data on the Web); and (ii) at least one publication with an
author that relates this publication to both the ‘2011 publication and the ISWC Conference 2010.
Table 2: A priori optimal suitability of the search tasks.
Average Precision measures the average of the search precision over all the required
actions in certain task. Thereby the precision  corresponds in this case to the effec-
tiveness of the kth search action as deﬁned for the user evaluation:
and the average precision over all actions Ain certain task:
average precision =AP =∑
However, the actions are different so a direct comparison for ResXplorer between the user
action effectiveness and the precision measured here is not possible. It also would make
no sense as the user tests focused on lean users while the experts are specialized in search
Efﬁciency, expressed as the number of actions (Nx) when users perform a certain task
(Tx). The lower the score, the less actions the experts needed to successfully complete the
To verify that the expert reviews are similar enough to be considered, we measured
the inter-rater agreement among them. We selected therefore the chance corrected agree-
ment (κ) measure  (−1<κ<1). The inter-rater agreement of the results between
the experts is substantial (κ=0.61 and F-measure 0.83) according to the Landis et al.
scale . The visualization in Figure 12 shows the mean average results for each of the
tested search interfaces and indicates how well the expert reviews match.
Tables 3 and 4 display the results of the expert evaluations of ResXplorer in compari-
son to two industry references and three research projects in the same domain. In ARNet
Miner and ResXplorer the autocomplete facilitated instant and precise matches. In Mi-
crosoft Academic Search, Google Scholar and Falcons the ﬁrst page of results contained
the necessary results and Google Scholar and Microsoft Academic Search promoted the
matching result as a suggestion on top of the list.
0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1
Mean Average Search Precision E1 E2
Fig. 12: The agreement between the experts on the ratings over all search interfaces combined is substantial.
Table 3: The search precision for getting the ﬁrst search results returns all true positive
matches except ArnetMiner returned 4 out of 5 false positives in T1. ResXplorer is not as
precise as the other interfaces for T2but excels in T3.(brighter = better)
Google Scholar 1.00 0.90 0.35 1.00 0.43 0.62 0.72
MA Search 1.00 1.00 0.63 1.00 0.90 0.64 0.86
Falcons 1.00 0.95 0.63 0.78 0.60 0.68 0.77
ResXplorer 1.00 0.84 0.84 0.70 0.39 0.80 0.76
ARNetMiner 0.60 1.00 0.81 0.74 0.20 0.49 0.64
Faceted DBLP 1.00 1.00 0.83 0.95 0.52 0.45 0.79
Table 4: An increased number of user actions does not always guarantee more precise
(intermediate) results, but it does for ResXplorer, except in T5.(brighter = better)
Google Scholar 1 1 2 2 5 3 15
MA Search 1 1 3 1 2 4 12
Falcons 1 2 3 3 3 6 18
ResXplorer 1 2 4 3 4 4 21
ARNetMiner 1 1 3 1 2 3 11
Faceted DBLP 1 1 3 1 2 3 10
T3 is a non-direct relation ﬁnding task and that is the main goals of ResXplorer while
T2 requires zooming in depth around a speciﬁc property of a person. ResXplorer intends
to maintain the broad overview at all times during the search which induces some noise
for task like T2.
In T4 the industry references beat the research engines. T4 requires skimming or
ﬁltering a list of conferences which is not supported in ResXplorer and in Falcons and
ArnetMiner not to the same degree as the industry references. Faceted DBLP also scores
well for T4 thanks to the faceted search interface and tight DBLP link. For T4 required the
20 Laurens De Vocht et al.
Google Scholar interface scrolling through two pages to ﬁnd three different conferences,
many results of the same conference. Microsoft Academic Research allowed searching
speciﬁcally for items of the type conference. That explains the highest rating here, all
results were on the ﬁrst page in contrast to Google Scholar. in Falcons the results were a
little less accurate and did not allow searching speciﬁcally for conferences either. ResX-
plorer did, but as it did not provide a list but a limited set of entry points for exploration.
This meant the search was repeated to ﬁnd different entry points leading to a conference,
in fact three times, each time to ﬁnd a new conference. ARNet Miner provided a result
view, containing distracting of widgets, not all material was relevant for the search. It
included relatively many false positives to interpret but all results were found after one
search action. The expert users judge the results presented in the a priori deﬁned complex
tasks having the most irrelevant results and they needed at least 2 actions in T5 and even
3 actions in T6 to resolve the search task. The highest effectiveness was found for MA
Search in T5 and for ResXplorer in T6. In terms of efﬁciency Google Scholar required
the most actions and in T5 and Falcons in T6.
We observed that searching by keywords for resources increases the result set with the
most new relevant resources, while it is on average as effective as expanding existing
resources in the result set. The most effective user action was adding top-related nodes to
ResXplorer is situated in the mid-range in terms of mean average search precision
and requires relatively lots of action from the user. However, ResXplorer is best when the
task consisted of relating resources that are not directly related or when at least the user
is not aware of how they are related. That is precisely the goal we wanted to show with
ResXplorer and the methods and techniques that drive it. Furthermore, this pinpoints,
once again, to the importance and the need of user-centered evaluation concept within
conducted measurements. The main concept of ResXplorer resides on the idea of an
interactive search interface which leads the researcher through the process of expansion
and exploration of results to the hidden implicit valuable information discoveries which
are uncovered in such a process. The balanced choice of comparable solutions: two of
them from industry (MA Search and Google Scholar) and three of them from research
domain (ARNet Miner, Falcons and Faceted DBLP); this allows good positioning and
qualitative reviewing of our solution.
Having visually more advanced solutions like MA Search and ARNet Miner and those
with less search interface interactivity possibilities like Google Scholar, Faceted DBLP
and Falcons we also want to cover the essential aspect in evaluation for user driven search
applications which considers the visual representation and analysis of search results and
interaction possibilities on search interface . In order to outline the differences between
conventional search interfaces for scientiﬁc resources and our approach, we used a set of
“Visual representation and analytics” based on guidelines identiﬁed by .
We compared the features of the search interfaces used in the expert evaluation as
listed in table 5. We notice that industry references as MA Search and Google Scholar
lack the interactivity with a visual representation, although MA Search for instance of-
fers visual interfaces to the search results. On the other hand, ARNet Miner supports
various visualizations based on data mining algorithms, like e.g. clustering, executed
on the retrieved data in combination with the search results. Faceted DBLP features an
interactive, all-round faceted search interface. Falcons Object Search  is considered as a
keyword-based search engine for linked objects with extensive virtual documents indexed.
Those documents consist from associated literals but also from the textual descriptions of
associated links and linked objects. The results are ranked according to a combination
of their relevance to the query and their popularity. Besides a classical list representation
Falcons allows enhanced text based browsing of Linked Data as well ﬁltering on concepts
Table 5: Comparision of functionality of different search interfaces for research.
Usability Criterion ResXplorer MA Search GScholar ARnet Miner Falcons Faceted DBLP
Query (forms / keyword)
Query (formal syntax) G# G#
View results as ordered list #
Visual presentation # # #
Interactively reﬁne search # # G# G#
Combine and relate searches # # # # #
Data overview #
Detail on demand
Generic / Engine Reusable ? ? G# ? ?
Support for scalability
History G# G# # # #
View original source G#
Feature coverage = full G#= partially #= none ? = uncertain
With our search engine users can combine any searches and interact the results that
exposes relationships between them. This is a feature not found in conventional search
interfaces. It offers search for publications, as well as supports relation visualization
on author level. We visually emphasize discovered types of entities and relations. In
comparison to the current existing solutions we can use the snapshot of social content
published by researchers on social media and collaborative platforms like Twitter and
Mendeley to make a pre-set for exploratory search. This feature is unique to our solution.
Furthermore, the method by which we generate context-based results differs from ARnet
Miner because we do not rely on data mining and machine learning techniques to resolve
the research related information. Our approach uses afﬁnity based ranking derived from
the social context and search process itself. We use graph based algorithms which perform
independently of underlying Linked Data. In comparison to the existing search solutions,
our interface is designed to visually explore the research space, rather than to support
classical keyword based search. This exploration is based on personal preference and
serendipity of information in the data set (publications, persons, events). This data is
enhanced by additional information (e.g. venues of events) related to the search. Unlike
Microsoft Academic Search and ARnet Miner our graph visualization is expandable and
includes entities from Linked Data and description of relations between them. Since pre-
sets of the search reside on actualized social media content of the user our solution adapts
better on changes of information and trends from social media. This aspect differs strongly
from the conventional approaches mentioned here.
22 Laurens De Vocht et al.
7. Conclusions and Future Work
We have presented a semantic model for searching resources in the Web of Data developed
for scientiﬁc research. We have demonstrated and implemented the model with current
state-of-the art technologies. The model uses research objects to represent semantically
modeled research data. We explained how applying this model contributes to implemen-
tations for research use cases.
The result is a semantic search application providing both a technical demonstration
and a visualization that could be applied in many other disciplines beyond Research
2.0. The main contribution of our work is, besides retrieving resources from Linked
Data repositories, allowing researches to interactively explore relationships between the
resources and entities like events or persons related to their work.
Compared to existing well established search interfaces with similar goals and the
same target audience, ResXplorer is situated in the mid-range when compared to both
industry and academic projects targeting the same use case, but requires more interaction
of the users. However when a task consists of ﬁnding resources that connect a given state-
ment, such as ﬁnding common items between two authors of an article, ResXplorer has
relatively the highest search precision. Considering that the implementation of ResXplorer
is still in the prototype phase, the potential of a visual and interactive search interface is
well demonstrated and understood by the target users.
We will clearly distinguish between proposing new afﬁnities between certain resources
versus exploring the proposed resources in detail and explaining the motivation behind the
afﬁnities, where we characterize each afﬁnity, between researchers and resources, by the
amount of shared interests and other commonalities. To make ResXplorer more precise in
classical search and retrieve scenarios, more accurate ﬁlters on the search keywords and
results are crucial. We will analyze the efﬁciency further as a smaller number of actions
does not always lead to the most efﬁcient interface, certainly if it requires more thinking
and judging from the users: more straightforward steps might be more efﬁcient than less
but more complicated steps. In this work we have focused on proposing novel afﬁnities
related to a personalized search context, but it is important to allow researchers to explore
the proposed afﬁnities more in depth. Further improvements on the ranking criteria should
improve the precision of proposed afﬁnities and the results even further.
The research activities that have been described in this paper were funded by Ghent Uni-
versity, iMinds (Interdisciplinary institute for Technology) a research institute founded by
the Flemish Government, Graz University of Technology, the Institute for the Promotion
of Innovation by Science and Technology in Flanders (IWT), the Fund for Scientiﬁc
Research-Flanders (FWO-Flanders), and the European Union.
1. Bechhofer, S., Buchan, I., De Roure, D., Missier, P., Ainsworth, J., Bhagat, J., Couch, P.,
Cruickshank, D., Delderﬁeld, M., Dunlop, I., Gamble, M., Michaelides, D., Owen, S., New-
man, D., Suﬁ, S., Goble, C.: Why linked data is not enough for scientists. Future Generation
Computer Systems 29(2), 599 – 611 (2013)
2. Breslin, J.G., Decker, S., Harth, A., Bojars, U.: SIOC: an approach to connect web-based
communities. International Journal of Web Based Communities (IJWBC) 2(2), 133–142 (2006)
3. Breslin, J.G., Harth, A., Bojars, U., Decker, S.: Towards semantically-interlinked online
communities. In: Gomez-Perez, A., Euzenat, J. (eds.) European Semantic Web Conference
(ESWC). Lecture Notes on Computer Science, vol. 3532, pp. 500–514. Springer (2005)
4. Brusilovsky, P.: Methods and techniques of adaptive hypermedia. In: Adaptive hypertext and
hypermedia, pp. 1–43. Springer (1998)
5. Brusilovsky, P., Peylo, C.: Adaptive and intelligent web-based educational systems. Interna-
tional Journal of Artiﬁcial Intelligence in Education 13(2), 159–172 (2003)
6. Cheng, G., Qu, Y.: Searching linked objects with falcons: Approach, implementation and
evaluation. International Journal on Semantic Web and Information Systems (IJSWIS) 5(3),
7. Cheng, G., Qu, Y.: Searching linked objects with falcons: Approach, implementation and
evaluation. Int. J. Semantic Web Inf. Syst. 5(3), 49–70 (2009)
8. Dadzie, A.S., Rowe, M.: Approaches to visualising Linked Data: A survey. Semant. web 2(2),
89–124 (Apr 2011)
9. d’Aquin, M., Baldassarre, C., Gridinoc, L., Angeletou, S., Sabou, M., Motta, E.: Characterizing
knowledge on the semantic web with Watson. In: Garcia-Castro, R., Vrandecic, D., Gmez-Prez,
A., Sure, Y., Huang, Z. (eds.) EON. CEUR Workshop Proceedings, vol. 329, pp. 1–10. CEUR-
10. De Vocht, L., Coppens, S., Verborgh, R., Vander Sande, M., Mannens, E., Van de Walle, R.:
Discovering meaningful connections between resources in the web of data. In: Proceedings of
the 6th Workshop on Linked Data on the Web (LDOW2013) (2013)
11. De Vocht, L., Deursen, D.V., Mannens, E., de Walle, R.V.: A semantic approach to cross-
disciplinary research collaboration. Internation Journal of Emerging Technologies in Learning
(iJET) 7(S2), 22–30 (2012)
12. De Vocht, L., Mannens, E., Van de Walle, R., Softic, S., Ebner, M.: A search interface for
researchers to explore afﬁnities in a Linked Data knowledge base. In: Proceedings of the 12th
International Semantic Web Conference Posters & Demonstrations Track. pp. 21–24. CEUR-
13. De Vocht, L., Softic, S., Dimou, A., Verborgh, R., Mannens, E., Ebner, M., Van de Walle, R.:
Visualizing collaborations and online social interactions at scientiﬁc conferences for scholarly
networking. In: Proceedings of the Workshop on Semantics, Analytics, Visualisation: Enhanc-
ing Scholarly Data (SAVE-SD 15); 24th International World Wide Web Conference (2015)
14. De Vocht, L., Softic, S., Ebner, M., M¨
uhlburger, H.: Semantically driven social data aggregation
interfaces for research 2.0. In: Proceedings of the 11th International Conference on Knowledge
Management and Knowledge Technologies. pp. 43:1–43:9. i-KNOW ’11, ACM (2011)
15. De Vocht, L., Softic, S., Mannens, E., Ebner, M., Van de Walle, R.: Aligning web collaboration
tools with research data for scholars. In: Proceedings of the Companion Publication of the 23rd
International Conference on World Wide Web Companion. pp. 1203–1208. WWW Companion
’14, International World Wide Web Conferences Steering Committee (2014)
16. De Vocht, L., Softic, S., Mannens, E., Van de Walle, R., Ebner, M.: Resxplorer : interactive
search for relationships in research repositories. In: International Semantic Web Conference :
Semantic Web Challenge, Abstracts. p. 8. Trentino, Italy (2013)
17. Dimou, A., De Vocht, L., Van Compernolle, M., Mannens, E., Mechant, P., Van de Walle, R.:
A visual workﬂow to explore the web of data for scholars. In: Proceedings of the Companion
24 Laurens De Vocht et al.
Publication of the 23rd International Conference on World Wide Web Companion. pp. 1171–
1176. WWW Companion ’14, ACM (2014)
18. Ding, L., Kleb, J., Mueller, W.: How the semantic web is being used: An analysis of FOAF
documents. In: Proceedings of the 38th International Conference on System Sciences (2005)
19. Ebner, M., Reinhardt, W.: Social networking in scientiﬁc conferences - Twitter as tool for
strengthen a scientiﬁc community. In: Cress, U., Dimitrova, V., Specht, M. (eds.) Learning
in the Synergy of Multiple Disciplines, Proceedings of the EC-TEL 2009. Lecture Notes in
Computer Science, vol. 5794. Springer, Berlin/Heidelberg (October 2009)
20. Faulkner, L.: Beyond the Five-User assumption: Beneﬁts of increased sample sizes in usability
testing. Behavior Research Methods, Instruments, & Computers 35(3) (2003)
21. Glaser, H., Millard, I.C., Jaffri, A.: RKBExplorer.com: a knowledge driven infrastructure for
linked data providers. In: Proceedings of the 5th European semantic web conference on The
semantic web: research and applications. pp. 797–801. ESWC’08, Springer-Verlag, Berlin,
22. Graves, A.: Creation of visualizations based on Linked Data. In: Proceedings of the 3rd Inter-
national Conference on Web Intelligence, Mining and Semantics. p. 41. ACM (2013)
23. Hogan, A., Harth, A., Umbrich, J., Kinsella, S., Polleres, A., Decker, S.: Searching and brows-
ing linked data with swse: The semantic web search engine. Web Semantics: Science, Services
and Agents on the World Wide Web 9(4), 365 – 401 (2011)
24. Hripcsak, G., Rothschild, A.S.: Agreement, the F-measure, and reliability in information re-
trieval. Journal of the American Medical Informatics Association 12(3), 296–298 (2005)
25. Hristoskova, A., Tsiporkova, E., Tourw, T., Buelens, S., Putman, M., De Turck, F.: Identifying
experts through a framework for knowledge extraction from public online sources. In: 12th
Dutch-Belgian Information Retrieval Workshop (DIR 2012), Ghent, Belgium. pp. 19–22 (2012)
26. Jiang, M., Cui, P., Wang, F., Xu, X., Zhu, W., Yang, S.: Fema: Flexible evolutionary multi-
faceted analysis for dynamic behavioral pattern discovery. In: Proceedings of the 20th ACM
SIGKDD International Conference on Knowledge Discovery and Data Mining. pp. 1186–1195.
KDD ’14, ACM, New York, NY, USA (2014)
27. Kraaij, W., Post, W.: Task based evaluation of exploratory search systems. In: SIGIR 2006
workshop, Evaluating Exploratory Search Systems (2006)
28. Landis, J.R., Koch, G.G.: The measurement of observer agreement for categorical data. bio-
metrics pp. 159–174 (1977)
29. Ley, M.: The DBLP computer science bibliography: Evolution, research issues, perspectives.
In: String Processing and Information Retrieval. pp. 1–10. Springer (2002)
30. Oren, E., Delbru, R., Catasta, M., Cyganiak, R., Stenzhorn, H., Tummarello, G.: Sindice.com:
a document-oriented lookup index for open linked data. Int. J. of Metadata and Semantics and
Ontologies 3, 37–52 (2008)
31. Pace, S.: A grounded theory of the ﬂow experiences of web users. International journal of
human-computer studies 60(3), 327–363 (2004)
32. Parra Chico, G., Duval, E.: Filling the gaps to know More! about a researcher. In: Proceedings
of the 2nd International Workshop on Research 2.0. At the 5th European Conference on
Technology Enhanced Learning: Sustaining TEL,. pp. 18–22. CEUR-WS (Sep 2010)
33. Passant, A., Bojars, U., Breslin, J.G., Decker, S.: The SIOC project: Semantically-interlinked
online communities, from humans to machines. In: Padget, J.A., Artikis, A., Vasconcelos,
W., Stathis, K., da Silva, V.T., Matson, E.T., Polleres, A. (eds.) Coordination, Organizations,
Institutions, and Norms in Agent Systems V. Lecture Notes in Computer Science, vol. 6069,
pp. 179–194. Springer (2009)
34. Passant, A., Bojars, U., Breslin, J.G., Hastrup, T., Stankovic, M., Laublet, P.: An overview of
SMOB 2: Open, semantic and distributed microblogging. In: Cohen, W.W., Gosling, S. (eds.)
ICWSM. pp. 303–306. The AAAI Press (2010)
35. Passant, A., Breslin, J.G., Decker, S.: Open, distributed and semantic microblogging with
SMOB. In: Benatallah, B., Casati, F., Kappel, G., Rossi, G. (eds.) ICWE. Lecture Notes in
Computer Science, vol. 6189, pp. 494–497. Springer (2010)
36. Powers, D.: Evaluation: From precision, recall and F-measure to ROC., informedness, marked-
ness & correlation. Journal of Machine Learning Technologies 2(1), 37–63 (2011)
37. Softic, S., De Vocht, L., Mannens, E., Van de Walle, R., Ebner, M.: Finding and exploring
commonalities between researchers using the resxplorer. In: Learning and Collaboration Tech-
nologies. Technology-Rich Environments for Learning and Collaboration - First International
Conference, LCT 2014, Held as Part of HCI International 2014, Heraklion, Crete, Greece, June
22-27, 2014, Proceedings, Part II. pp. 486–494 (2014)
38. Softic, S., Ebner, M., M¨
uhlburger, H., Altmann, T., Taraghi, B.: Twitter mining #microblogs
using #semantic technologies. 6th Workshop on Semantic Web Applications and Perspectives
pp. 1–9 (2010)
39. Softic, S., Vocht, L.D., Mannens, E., Ebner, M., de Walle, R.V.: COLINDA: modeling, repre-
senting and using scientiﬁc events in the web of data. In: Proceedings of the 4th International
Workshop on Detection, Representation, and Exploitation of Events in the Semantic Web
(DeRiVE 2015) Co-located with the 12th Extended Semantic Web Conference (ESWC 2015),
Protoroz, Slovenia, May 31, 2015. pp. 12–23 (2015)
40. Softic, S., Vocht, L.D., Mannens, E., de Walle, R.V., Ebner, M.: Finding and exploring com-
monalities between researchers using the resxplorer. In: Learning and Collaboration Tech-
nologies. Technology-Rich Environments for Learning and Collaboration - First International
Conference, LCT 2014, Held as Part of HCI International 2014, Heraklion, Crete, Greece, June
22-27, 2014, Proceedings, Part II. pp. 486–494 (2014)
41. Sure, Y., Bloehdorn, S., Haase, P., Hartmann, J., Oberle, D.: The SWRC ontology - semantic
web for research communities. Progress in Artiﬁcial Intelligence pp. 218–231 (2005)
42. Tang, J., Zhang, J., Yao, L., Li, J., Zhang, L., Su, Z.: Arnetminer: extraction and mining of
academic social networks. In: Proceedings of the 14th ACM SIGKDD international conference
on Knowledge discovery and data mining. pp. 990–998. ACM (2008)
43. Tao, K., Abel, F., Gao, Q., Houben, G.J.: Tums: Twitter-based user modeling service. In:
Garcia-Castro, R., Fensel, D., Antoniou, G. (eds.) ESWC Workshops. Lecture Notes in Com-
puter Science, vol. 7117, pp. 269–283. Springer (2011)
44. Tran, T., Herzig, D.M., Ladwig, G.: Semsearchpro - using semantics throughout the search
process. Web Semantics: Science, Services and Agents on the World Wide Web. 9(4), 349–364
45. Ullmann, T.D., Wild, F., Scott, P., Duval, E., Vandeputte, B., Parra Chico, G.A., Reinhardt, W.,
Heinze, N., Kraker, P., Fessl, A., Lindstaedt, S., Nagel, T., Gillet, D.: Components of a research
2.0 infrastructure. In: Lecture Notes in Computer Science,. pp. 590–595. Springer (2010)
46. Uren, V., Lei, Y., Lopez, V., Liu, H., Motta, E., Giordanino, M.: The usability of semantic
search tools: A review. The Knowledge Engineering Review. 22(4), 361–377 (Dec 2007)
47. Uren, V., Sabou, M., Motta, E., Fernandez, M., Lopez, V., Lei, Y.: Reﬂections on ﬁve years of
evaluating semantic search systems. International Journal of Metadata, Semantics and Ontolo-
gies (IJMSO) 5(2), 87–98 (2010)
48. Van Noorden, R.: Online collaboration: Scientists and the social network. Nature News
512(7513), 126–130 (Aug 2014)
49. Weibel, S.: The Dublin Core: a simple content description model for electronic resources.
Bulletin of the American Society for Information Science and Technology 24(1), 9–11 (1997)
50. White, R.W., Marchionini, G.: Examining the effectiveness of real-time query expansion.
Information Processing & Management 43(3), 685–704 (2007)
51. White, R.W., Marchionini, G., Muresan, G.: Evaluating exploratory search systems: Introduc-
tion to special topic issue of information processing and management. Inf. Process. Manage.
44(2), 433–436 (Nov 2008)
52. White, R.W., Muresan, G., Marchionini, G.: Evaluating exploratory search systems. EESS 2006
p. 1 (2006)
53. Xin, R.S., Hassanzadeh, O., Fritz, C., Sohrabi, S., Miller, R.J.: Publishing bibliographic data
on the semantic web using BibBase. Semantic Web 4(1), 15–22 (2013)