This paper presents a Linked-Data-driven Web portal for the eld of learning analytics. The portal allows users to browse the linked datasets and explore data about researchers, conferences, and publications. Additionally, users can interact with various dynamic visualization applications and perform analysis, e.g., study temporal change of research trends. Based on the provided datasets on Learning Analytics and Knowledge (LAK) and Educational Data Mining (EDM), we enriched the data with geospatial locations of research institutes, topics extracted from papers, and the expertise of researchers. The interactive modules of the Web portal are then designed and implemented using the enriched RDF data. The implemented modules can be divided into two groups. The rst group is concerned with providing dynamic and interactive visualization of the data, such as the modules of Conference Participants and Reference Map. The modules in the second group are designed for more advanced analysis and discovery of new knowledge, such as the modules of Scholar Similarity and Reviewer Recommendation. The modules have been designed following a loosely coupled, modular infrastructure, and can be easily migrated and reused in other projects. H.2.8 [Database Applications]: Scienti c Databases| bibliography database; H.4 [Information Systems Applications]: Miscellaneous
sectionIntroduction and Motivation The Learning Analytics
and Knowledge (LAK) dataset provides a rich collection of
data extracted from publications in the eld of learning
analytics and structured in a machine-readable manner [
The presented Web portal (called DEKDIV4) can be accessed at: http://stko-exp.geog.ucsb.edu/lak/.
3This is the only conference in the dataset that o ers an extensive history worth analyzing. In the future, more conferences can be added. 4Short for Data Enrichment, Knowledge Discovery and Interactive Visualization.
The LAK dataset contains information about publications and researchers in the Learning Analytics and Knowledge (LAK) conference series (from 2011 to 2013) and the Educational Data Mining (EDM) conference series (from 2008 to 2013). This dataset also contains a special issue from the Journal of Educational Technology and Society on learning analytics and knowledge. In contrast to other publication data hubs (e.g., DBLP5 and CiteSeer6) on the Linked Open Data (LOD) cloud, the LAK dataset provides full texts and full references in addition to the typical bibliographic data. As a result, it is not only possible to explore the collaboration relations (e.g., via co-authorships), but also to mine key concepts, research topics, and citation networks. This facilitate a more holistic understanding of the learning analytics and educational data mining research elds.
In this work, we further enrich the provided LAK dataset
with key concepts and topics extracted from papers,
geographic locations of authors' a liations, as well as the
expertise of researchers and paper citations imported from
Arnetminer and Microsoft Academic Search (MAS). While
these additional data have been imported into our own triple
store7, they can also be merged with the existing LAK
dataset. In addition, a customized daemon can be developed
to dynamically enrich the data whenever new publications
(e.g., papers from the LAK 2014 conference) are added. In
fact, this is the strategy that has already been adopted in our
previous Linked Data portal for the Semantic Web Journal
to synchronize the information about new publications [
Based on the full text of each publication in the LAK
dataset, we extracted important concepts using the Alchemy
API8. Alchemy is a Web service that provides automatic
natural language processing functions. Each of the extracted
concepts (key phrases) is associated with a relevance value
which indicates the term's importance in relation to the
entire paper. The extracted concepts were used in the Key
Concepts module to give users an overview of a paper's
content. These concepts were also used to display the trend of
research topics in the Topic Trending module. Additionally,
we use a Latent Dirichlet allocation (LDA) model to extract
latent topics from the full text data. LDA is an
unsupervised, generative probabilistic model used to infer the latent
topics in a textual corpus [
6http://thedatahub.org/dataset/ rkb-explorer-citeseer 7http://stko-exp.geog.ucsb.edu/pubby 8http://www.alchemyapi.com/ LDA were used for the functions in Scholar Similarity and Reviewer Recommendation.
Scienti c activities (e.g., co-publications, citations, and
references) generally show geospatial patterns. However, such
patterns have rarely been considered in traditional
scientometric analysis which often focuses on numeric values (e.g.,
the H-index) [
In addition to the key phrases, topics, and geospatial locations, we also integrated the scientometric data from two major Web services: Arnetminer and Microsoft Academic Search. Among the 853 authors in the LAK dataset, Arnetminer contains data for 595 of them, and such data includes publications by these authors not only from the LAK dataset, but also from other journals and conferences. We imported the researcher expertise data (which were extracted based on all publications of a given author) from Arnetminer into our local triple store, and used these expertise data for the Potential Collaborators module. Microsoft Academic Search is another source whose data have been integrated into the presented Web portal. MAS contains a large amount of information about a liations and citations. Such data have been used to nd the a liation of authors in the reference data. The citation data from MAS have also been integrated to show the external citations which are not stored in the existing LAK dataset.
This section presents the functionality modules that serve the purpose of interactive visualization and animation. Visualization plays an important role in data analysis, since humans can often easily recognize patterns visually (in contrast to machines). When designing this Web portal, we paid special attentions to each aspect of the visualizations created. For instance, we tried to ensure that each visualization ts the data being presented, and that the user interface enables users to interact and understand the data. The subsections below describe some of these modules, and the role that visualization plays in their conceptualization.
The Collaborative Institutes module is designed to help users
explore the spatial distribution of co-authorship. In order
to realize this, the number of co-authored conference papers
between two institutes are displayed on a global map. As
shown in Figure 1, a link between two institutes indicates
research collaborations between authors from these institutes.
Interactions provided by this module include moving the
mouse over links and nodes to show additional details about
each of the institutes and publication information. Using
this geo-visualization, one can see spatial patterns of
research collaborations. For example, some researchers prefer
to collaborate domestically for certain conferences, as
opposed to international collaborations. Previous studies [
The Conference Participants module was developed to provide an interactive visualization and animation of the geographic distributions of EDM & LAK conference authors; see Figure 2. The size of the yellow circles represents the number of participants that attended the conference; the larger the circle, the more attendees from the mapped institute. Hovering the mouse over a circle on the map displays the total number of authors who participated in the conference. From a purely visual perspective, it appears that the geographic distribution of participants is strongly in uenced by the region where the event took place. For instance, LAK 2013 was held in Leuven, Belgium, and it attracted many European researchers, while many participants of EDM 2013, which was held in Memphis, USA, are from the United States.
The Academic Network module is constructed to show an interactive visualization of an author's academic network based on co-author links. A graph-node approach (often found in social network analysis) has been implemented to connect authors with each other through their LAK/EDM publications. This technique views the academic network as a set of relationships composed of nodes and links. In this module, each author is presented as a node in the network, and a link connects two nodes if co-authorship exists between these two authors. The total number of coauthorships between two authors is recorded as an attribute of the link and is visually represented though the stroke width of each link (Figure 3). In addition to the visualization of the academic network, measurements of the network are also provided. For example, the Centrality of a node represents its relative importance within a network, and four types of centrality are visually presented in the Academic Network module when a node is selected.
Based on the enriched and geocoded data, we have designed a module called Reference Map. This module presents an animated and interactive visualization demonstrating the geographic distribution of a paper's references. Using the reference data provided by the LAK dataset, we extracted the names of the rst authors in the reference records, and use MAS to nd the institutions of these authors. We then visualized these institutions as smaller bubbles on a global map, and created animated links from these bubbles to the location of the published paper, where a larger bubble is created (see Figure 4). The message that we are trying to deliver through this visualization is: scienti c publications (i.e., the smaller bubbles) emerged at di erent locations of the world, and new papers (i.e., the bigger red bubble) was created and published when other researcher adopted these ideas (i.e., the animated linked). In this module, users can see additional information about authors, institutions, and each reference when hovering over the bubbles and links.
The richness of the LAK/EDM dataset allows the development of a wide assortment of knowledge discovery tools. In this section, we present some of the modules which are developed based on data mining techniques. These modules utilize the enriched dataset, and o er informative metrics to facilitate the understanding of not only the dataset itself but also the LAK community. In addition, some of these modules provide useful services which, until now, may have not been fully realized.
The module of Active Scholar is designed to discover the most productive researchers based on the provided dataset. We developed a SPARQL query which ranks authors in each conference according to the number of their publications. The top 30 authors were then selected, and displayed in the center of the visualization canvas (see Figure 5). The number after the name of each author denotes the number of papers that this speci c author published in this conference. While authors who have two or more publications are displayed at the top, we have to randomly select authors who have a single publication in the conference due to the limited visualization space. In addition to the most active scholars, we also identi ed and visualized the most popular topics based on the number of publications in this topic. By hovering mouse on one of the authors, users can see the topics that are related to this author. Similarly, moving mouse on one topic will show all the authors who have publications on this topic. Based on this module, we have discovered some very active scholars (e.g., Ryan S.J.d Baker) who often publish more than 2 papers in the LAK conferences.
The Scholar Similarity module measures the similarity
among authors in the LAK dataset using a Multidimensional
Scaling (MDS) approach. The full text publications of each
author have been selected as input for a Latent Dirichlet
allocation (LDA) model consisting of 20 topics. LDA is an
unsupervised, generative probabilistic model used to infer
the latent topics in a textual corpus [
The Reviewer Recommendation module (Figure 6) is built on the concept that researchers who have already published on a topic have the potential to become good reviewers for new papers on the same topic. Similar to the Scholar Similarity module, this tool uses LDA to generate a set of topics from a corpus of full text publications. The authors of these publications are then de ned as a distribution across the topic space. Given a new submission (in the form of an abstract), the topic distribution of this material can be inferred from the existing LDA topics. Again, using the Jensen-Shannon divergence method, the topic distribution of the new material was compared with the topic distributions of all authors, and authors were then ranked based on the similarity value ( de ned as 1 { JSD dissimilarity value). To avoid interest con icts, we compared the list of potential reviewers with the researcher's previous co-authors. Those co-authors were highlighted in the visualization to re ect the con icts.
This module demonstrates how research topics in the LAK or the EDM conferences trend over time (see Figure 7). In order to implement this module, data preprocessing has been done to extract the top 10 topics for each year ranked by total number of papers which contain the topic keywords. The EDM conferences, for instance, has 46 distinct topics extracted from 2008 to 2013. Frequencies for all topics were calculated across all years in a time-series format for further visualization and analysis. Using the interactive stream graph, it is clear to see the decline of certain topics, emergence of new topics, as well as trend expansions through time. The user interface enables interaction through mouse hover, reporting the chosen topic along with the number of papers associated with the topic in each year.
This analytical function is designed to nd out the researchers who have similar research interests but may have never co-authored a paper before (i.e., the researchers that can potentially become collaborators). The similarity of researchers was calculated using cosine similarity measure based on expertise of the research imported from Arnetminer. We then calculated the shortest network distance based on the co-authorships in the module of Academic Network. After that, a metric (see equation (1)) was designed to nd authors who have the potential to become collaborators. p = sim(a1 ; a2 ) (1 1=d) (1) Where p is the score for collaboration potential, sim(a1 ; a2 ) is the cosine similarity between authors a1 and a2, and d is the shortest network distance. Figure 8 shows a screenshot of the Potential Collaborator module, where the blue icon represents the current researcher and the surrounding grey icons display his/her potential collaborators. Researchers who have already co-authored papers before will have a link, while no link indicates there exist no co-authored paper in the LAK dataset.
While we only highlight some selected modules here, there are a list of other DEKDIV modules. Interested readers are encourages to explore these modules. Examples include: Conference Hot Topics: The most popular topics in this conference.
Coauthor Treemap: A treemap visualization of coauthors and their a liations.
Citation Map: The geospatial distribution of the citations of authors and their publications.
Key Concepts: Important key phrases of a given paper and their relevance.
In this paper, we presented a Linked-Data-driven scientometrics Web portal, called DEKDIV, for the LAK challenge. We enriched the initial LAK data with paper topics, geospatial locations, and author expertise. We published the data in our local triple store, and designed various interactive visualization and analysis modules to facilitate the understanding of research in the LAK community. We also developed a set of more complex service modules on top of the enriched data. They can be used to recommend reviewers for newly submitted papers, help researchers nd potential collaborators, detect trend changes in publication topics, and so forth. The developed Web portal is based on our previous work for the Semantic Web journal, and the entire system is highly modular, reusable, as well as free and open source software.