The LAK Data Challenge asks the question “What do analytics on learning analytics tell us?” One approach to this challenge is not to answer the question, but to provide a simple, user-focused application that allows any user to easily draw their own conclusion. This was Blue Canary's driver for building the LAK Explorer (http://lakexplorer.bluecanarydata.com). Our team combined multiple tools to create a powerful search application. We extracted topics from the papers, used an autocomplete feature in the search bar, added topics as search result metadata, and provided links to similar papers all as part of the user search experience. The value is in the usability. In the same way that Google presents powerful results via a simple interface, LAK Explorer allows for seamless searching, reading, and comparing of over 400 documents. The application is also instrumented to capture user input (search terms, papers viewed) to provide closedloop analytics in the future.
The LAK Data Challenge asks the question “What do analytics on
learning analytics tell us?” The Blue Canary team tackled this
question in 2014 by using topic modeling to describe trends in the
LAK Corpus [
Building off of previous LAK entries, the Blue Canary team took a different approach. Instead of analyzing the corpus to look for trends and threads, what if we made the corpus more easily searchable so that the analytics community can browse the corpus for meaning? This was the core of our approach for 2015. The result is the LAK Explorer (http://lakexplorer.bluecanarydata.com). It’s an intuitive search application that allows users to search, browse, and find content in the corpus of papers/articles provided by the LAK Data Challenge. Our goal was to automate the processing so that the LAK Explorer could be applied to any corpus, not just specifically tuned to the LAK data. 1.1
As will be explained in this paper, the use of Turbo Topics is a key
thread to the Blue Canary team’s approach. Blei’s research [
Blue Canary retrieved the LAK Dataset from the Challenge website (http://lak.linkededucation.org/lak/LAK-DATASETDUMP.rdf.zip). Upon examination, it appeared as if this dataset did not contain the entirety of the updated 2015 content. Of the 579 content tags (<led:body> and <bibo:content> ), 108 were empty. Blue Canary inquired about the gaps but at the time of this project submission, the content was not added to the dataset.
The LAK Explorer is a fusion of existing tools and techniques for interacting with semantic data. From the simple home screen to the detailed neighborhood of papers, each component was used to add utility to the search process. Figure 1 shows how the different tools were used at different points in the search process. The context for LAK Explorer is similar to that of a search engine. The user comes to the site knowing the universe in which they are searching (corpus of papers) and some idea as to what they want to find out (search terms). However, the user doesn’t know exactly what they are looking for. In that way, the presentation of results and related information is vital to improving the utility of the application.
As with any search engine, the usability of the tool starts with the home page. For LAK Explorer, the Blue Canary team was inspired by the simplicity of Google’s ubiquitous home page. drive the efficacy of the search and leverage the linked aspect of the LAK data. The search results could be weighted on content found in the content abstract, body, the author(s), and citations.
When a user hits the search button, LAK Explorer returns the results similar to the image in Figure 4. The home page is dominated by a large text entry box. The only other significant feature on the page is a listing of topics that were extracted from the corpus of papers.
When a user starts entering text into the search box, the first thing they will notice is the use of autocomplete. Autocomplete is advantageous since the LAK Explorer deals with a fixed corpus of knowledge. Instead of tying autocomplete to a larger base of content (e.g. dictionary or DBPedia), the team tied it to the topics that were extracted from the corpus using Turbo Topics – the same topics that appear under the search bar. Blue Canary used the Typehead feature from AngularStrap (http://mgcrea.github.io/angular-strap/#/typeaheads#typeaheads) to power this feature.
Elastic Search (http://www.elasticsearch.org/) was used to drive the search engine results in LAK Explorer. Blue Canary only used basic features of this tool to drive search results. The text entered in the search box are the inputs to a keyword search algorithm. There are additional features in Elastic Search that could further The papers/articles are presented in a fashion similar to an engine like Google Scholar. The title, author(s), and abstract are presented for each result. The most prominent addition is the inclusion of the color-coded topics. LAK Explorer uses the extracted topics as another level of filtering. Upon viewing the results, the user can see what topics are relevant for each paper, click the color-coded topic, and get new search results that are sorted by the frequency of that topic.
The utility of LAK Explorer is not to just browse search results. Blue Canary wanted to make the tool helpful for actually reading the resulting papers and articles. Therefore, clicking on a paper from the search results gives the display mode shown in Figure 5. The paper is displayed in a clear/crisp format for easy online reading. The display is split into three sections. First, the user sees the topics that are most strongly associated with that paper. Then, the abstract is presented followed by the body of the paper. Another key usability point is that the topics are color-coded at the top and the coloring remains intact throughout the body of the paper. This helps draw the reader’s attention to topics that may be of particular interest.
Perhaps the strongest feature of LAK Explorer is the ability to find
similar papers. Keyword and topic searching limits similarity to
only papers that share the same frequency of that one term. The
similar papers feature uses the entirety of the paper to compare it to
other content.
Blue Canary used a technique called Doc2Vec [
Additionally, LAK Explorer displays this same vector similarity in
a neighborhood scatter plot (Figure 6.).
The Paper Neighborhood graph takes the Doc2Vec vector data
and uses t-SNE (t-Distributed Stochastic Neighbor Embedding) to
give all papers two dimensional Cartesian coordinates [
The Blue Canary team wanted to create an application that LAK researchers and practitioners would find valuable. To that end, the team focused on a few key aspects of LAK Explorer to maximize its contribution to the field.
The Blue Canary product development team gives significant weight to usability. The team might develop an incredibly useful metric, but if the user can’t easily interpret that metric, it is useless. A clean layout, color coding, and simple charts all contribute to the usability of LAK Explorer. These features were consciously added in order to improve ease of use.
As the team discovered in our LAK14 submission, extracting topics from the corpus turned out to be a simple yet effective way of absorbing the content of papers. We continued this trend for LAK15 by using the Turbo Topics to aid in the initial search and in the meta-tagging of the search results.
The LAK Explorer was named due to the fact that users will likely not be looking for a singular result. They will look for the results of their search PLUS explore other papers that are similar to their top search result.
The paper similarity tools allow LAK Explorer users to follow this natural path of exploration: 1. 2. 3. 4.
Searching for Topic X gives me an ordered list of papers
After integrating the previously described components into the LAK Explorer, the team realized that there were additional features we could add to the tool to further increase its value.
The most impactful feature to add is to track user input to the application. LAK Explorer is already instrumented to capture the terms that users search and also the papers that are viewed. The additional feature would be to expose these tracking metrics to the application’s front end so that other users can see how the community is interacting with the tool. For example, a simple listing of “most viewed papers’ would add more fidelity to other LAK Explorer visitors.
The linked aspect of the corpus could be further exploited to improve the search process of LAK Explorer. The linked data discerns between abstracts, body, authors, citations, people, and institutions. These parts can be exposed as metadata in the search results (e.g. view more papers by this author) and the data can also be used to drive search efficacy (e.g. weigh hits to the abstract higher than hits to the paper body).
This is the third year of the LAK Data Challenge and all of the participants continue to stand on the shoulders of previous contributors. Blue Canary acknowledges that previous entrants such as the ones listed in this section have developed toolsets in the same vein as LAK Explorer
RecLAK [
DEKDIV [
A team of famous LAK researchers submitted a paper to the first
LAK Data Challenge in 2013 [
As with most initiatives at Blue Canary, this was a product of teamwork. LAK Explorer was made possible by a team of players who each brought additive skills to the table. In addition to the named authors, we’d like to thank Satya Mudiam, Faiz Mohammad, Avinash Narasingam, and Kiran Reddy for their contributions.