These algorithms showcase what is described in the paper submitted to MedEdPublish in February 2019, called: “Social learning in large online audiences of health professionals: Improving dialogue with automated tools”.
At the individual level, each participant has a certain “degree centrality”, referring to the number of connections the person has. Most social networks have opinion leaders (hubs, in network terms) with a disproportionate high number of connections. At the same time, the “betweenness centrality” measures the number of times a person acts as a bridge between any two nodes: this kind of person is very important for disseminating knowledge across a network.
At the subnetwork level, the existence of “cliques” is important: three or more people are connected by all possible connections (such as teams are: they all know each other). This is of practical interest for the improvement of the educational proposal, as this information allows the division of groups by affinity, rather than randomly.
At the network level, attributes such as “density” can be measured: the proportion of existing connections relative to the total number of possible connections.
Intelligent group creation
A third analysis, aimed at maximizing the likelihood of dialogue in online discussion groups, was to ensure that each of the online groups had participants with the highest number of connections, and that all groups were similar. This was done by running a Matlab program, created for this purpose. This did not decrease the need to partition the audience in groups, and therefore did not alter the number of tutors needed for each of the courses, but maximized the existing connections in the groups that were created, with the potential for improved dialogue. Specifically, this last analysis aims to group the participants into subgroups, each containing at least one participant with a high clustering coefficient and her/his closest contacts (namely, all the participants that share a connection with the highly clustered participant) while maximizing the overall density of connections for the resultant subgroups and the participants affinity. In this way, each subgroup is organized around (at least) one clique, maximizing the intra-group connections, and the resultant division of participants into the subgroups is chosen such that the average of connections is the maximum that can be achieved by other rearrangements of the participants and cliques
Natural language processing (discourse analysis)
It was introduced to classify posts in online discussions. A commercial platform was used for this purpose (22). A new classifier system using this platform was trained to detect support requests in the clinical forums, and automatically derive them to support rather than post them in the clinical discussion. The participant was prompted of this possibility (to send a request to support or to otherwise post), and asked to confirm or not. Depending on the answer, the post was sent to the support system or published in the clinical forum.
Comments feed priorization
Our campus provides a “Facebook-like” dialogue with other course participants who were previously related, as well as with the Course Faculty. In Facebook and other social network applications, there are no tutors, and each member sees the dialogue around the people they already know, as well as with the hubs or opinion leaders. To implement these functionalities generated a series of challenges, such as: how to display the discussion threads and weigh each variable (date of publication, Faculty or not, contact or not, number of likes, country of origin, and so on); how to show the discussion thread to the Faculty, and many other. But in the end, it allowed to have very large audiences moderated by only two tutors, as with smaller courses, and similar results in terms of participation, completion rates, learning, and satisfaction with the overall course and with the dialogue in the course.
Recommendations (content analysis)
The recommendations of resources and activities such as multimedia content and discussion forums, was aimed to help course participants access a study material of potential interest. Also, the recommender system provided recommendations of comments published by other participants or Faculty, so they could be aware of the relevant discussions. These academic recommendations were aimed at enhancing the learning process in the educational platform. To generate the academic recommendations, the recommender system calculated for each user the relevance of the items based on the evaluations of other colleagues who belonged to their network. All these recommendations were calculated in real time and offered to the participants during their interaction with the platform (21).
The recommendations of future courses, which were of interest to colleagues in their network, could also be of interest to the participants. To generate the “commercial” recommendations, the recommender system used demographic data to characterize the preferences of participants by specialty, gender, age or region, among others. All these recommendations were generated before the start of the courses and appeared once the users logged in to the platform as well as in the automatic emails that the learning platform sent to them (21).
This video (in Spanish) illustrates the way the virtual platform introduces many of the algorithms described before.