In this paper, we present network visualizations and an analysis of publications data from the LAK (Learning Analytics and Knowledge) in 2011 and 2012, and the special edition on Learning and Knowledge Analytics in Journal of Educational Technology and Society (JETS) in 2012.
The Society for Learning Analytics Research (SOLAR)1 provided
a dataset to solicit contributions to the LAK data challenge2
sponsored by the FP7 European Project LinkedUp3. The dataset
contains research publications in learning analytics and
educational data mining for the years 2010, 2011, and 2012
Having access to a dataset always offers new opportunities,
particularly in the educational domain, that lacks public datasets
for running experimental studies
2 http://www.solaresearch.org/events/lak/lak-data-challenge/
2.
It is often difficult for conference attendees to decide which workshops or sessions are suitable and relevant for them. Therefore, a list of recommended authors and papers based on shared interests could be supportive to plan the conference participation more efficiently and effectively. There already exist several papers published regarding awareness support for researchers (Reinhardt et al., 2012; Fisichella et al., 2010; Ochoa et al., 2009; Henry et al., 2009) and scientific recommender systems (Huang et al., 2002; Wang & Blei, 2010) but none of them has analyzed the Learning Analytics datasets for this purpose yet.
Our overall vision is to support the LAK attendees with a list of LAK authors and papers that are relevant for their own research interests. Such a recommendation could be created based on one or more of their own research papers but also on a short essay or even a tag cloud summarizing the research interest and objectives. Such a priority list can support the awareness of the attendees and empower the network of like-minded authors in the attendees’ particular research focus. In this paper, then, we aim to explore and identify like-minded authors within the LAK dataset. Supposing that we have a network of all the LAK authors and papers, the main research questions are: RQ1. How are the authors connected and which authors share more connections and are more central in terms of sharing commonalities with the others? RQ2. How are the papers connected to each other in terms of similarity? To answer these questions, we went through two main steps in our analysis: 1. Finding patterns of similarity between authors and papers, 2. Visualizing networks of the LAK authors and papers. We will now describe each step in detail.
To find relationships between authors, we first computed the similarity of the papers with the TF-IDF4 algorithm. TF-IDF can create a weighted list of the most commonly used terms in research articles. To generate the TF-IDF matrix for the LAK dataset, we first converted the LAK data from RDF to text files, which is an accepted format for the Mahout5 system. Then, we ran the default TF-IDF algorithm provided by Mahout on the text files. We removed the stop words by setting the configuration variables within Mahout to 90%. Thus, if a word appears in 90% of the document, it is considered as a stop word (e.g. and, or, the, etc.) and is removed from the similarity matrix. As a final outcome we had: • •
A so-called dictionary of all the terms in the LAK
dataset
A binary sequence file that includes the TF-IDF
weighted vectors
For computing similarity between the LAK authors, we used the
T-index algorithm
We visualized the generated graphs of authors and papers with the Welkin7 tool. Welkin takes an OWL file as input and provides visualization of the data as output. We present visualizations of the LAK authors and the LAK papers generated by Welkin in the following sub sections.
Figure 2 presents a network of the LAK authors in which red nodes represent the authors and the edges show the similarity between the publications of two authors. The result shows how the LAK authors are connected in terms of their publications' commonalities. Moreover, the network shows the users who share more commonalities than do other authors. We call them ‘central authors’. In the next section, we show how they are connected with the other authors in the network.
For some node in the network, the degree centrality shows the
total number of incoming and outgoing edges. It is a metric
commonly used for Social Network Analysis (SNA)
u1 u2 u3 u4 u5 u6 u7
Then first ten central authors
Figure 4 shows a network of the LAK papers. The red nodes are papers and the edges between them represent the similarity of the papers. By finding similar papers, we can recommend the most similar papers to specific authors. This increases the awareness of the authors about papers which are relevant to them and published in their communities.
Figure 4 shows that, some of the papers share more similarity with the others and own a higher degree number. As with the central authors, these papers will appear more often in the top recommendation list than the other papers of the dataset. One may interpret their degree as their popularity. Therefore, the papers with higher degree values are more popular and, presumably, they are more of interests to users. For the publication data, interests of users derives from the words and terms they have used more frequently in their papers.
47 24 46 23 42
The results presented here, allow us to answer our research questions in the following way: RQ1. How are the authors connected? Which authors share more connections and are more central in terms of sharing commonalities with the others? We presented a visualization of the authors’ network to provide an overview of how they are connected to each other. To justify the authors’ connections and relationships, we evaluated the degree centrality for the first ten, most central authors. Table 1 presents the first ten central authors and their degree to show the authors with the highest relevancy of their publications with others in the network. Table 1 shows the degree of the authors for sizes of neighborhoods equal to 10.
Figure 5 shows the degree centrality for the first ten papers that are most similar to the other papers. We selected the first ten top papers with the highest degrees. The horizontal axis (x) shows the top ten papers e.g. p1 is the paper with the highest similarity and thus, the highest degree value among the others shown by the vertical axis (y). Figure 5 shows degree centrality for two different sizes of nearest neighborhoods (n), 5 and 10. By increasing the n, the degree of the papers increases accordingly. As a result, we will have a larger number of top papers if n is higher (here, when n=10). In Figure 5, the degree for the first top paper (p1) is equal to 53 (n=10) and 29 (n=5). This shows how much p1 shares similarity with other papers. As a consequence, p1 can be considered as the most popular paper and it has the highest chance to appear in the top paper recommendations. We presented degree centrality of the LAK papers to give insight in their relationships in the papers’ visualized network. We selected the top ten papers that have the highest similarity with the other papers. To show which papers are placed in the top ten papers’ list, we present the title and authors for each paper. The top ten papers are not necessarily by the authors who are identified as the central authors. Although most of the central authors also appear in top ten papers’ list (see Table 2), the order is not the same. As we investigated the LAK data, we found out that some of the central authors have more than one paper. For instance, Hendrik Drachsler has contributed to four papers. In this study, similarity is calculated based on all papers of an author. So, it is quite probable that not each and every one of the authors’ papers individually has the highest similarity to the other papers. Although some of the central authors are common to the two Overall, we found that the LAK dataset can help conference attendees to become more aware of their research network, which, in its turn, is useful for sharing knowledge and experiences. However, the current dataset contains no user feedback or evaluations to evaluate either an author or a paper recommender system in terms of common metrics such as prediction accuracy and coverage of the generated recommendations. For future analysis it would be helpful if the LAK dataset also contains references to the papers. The references could be used to identify the top cited authors and papers within the LAK dataset and beyond. As a further step, we are planning to try additional social network analysis measures besides degree, such as betweenness or closeness. 6.