David Monllao, Du Huynh, Mark Reynolds, Martin Dougiamas and Damyon Wiese (2018) “A Supervised Learning framework for Learning Management Systems.” Proceedings of the First International Conference on Data Science, E-learning and Information Systems.
Last October, the city of Madrid hosted DATA ’18, the latest edition of the International Conference on Data Science, E-learning and Information Systems. It is an international gathering of researchers on the cutting-edge of various fields. Education and educational technology subjects held a respectable presence.
The paper presented by Moodle at DATA ’18 was co-authored by core team members, along with computer scientists from the University of Western Australia: Huynh and Reynolds. They both show a track record of research on machine learning, including classification algorithms, automated reasoning, visual tracking and spatial trajectory prediction, for over 20 years.
Despite the absence of Elizabeth Dalton‘s name on the paper, the influence of Moodle’s Learning Analytics Lead is undeniable. The model’s attributes make it a flexible and powerful predictive engine, whose current applications were defined by Dalton and the Moodle Learning Analytics Working Group. An introduction and supporting materials are available at her Moodle Development School. While the research so far deals with the prediction side of things, the endgame is to inform the learning processes themselves. In practical terms, this means the models will seek to provide more predictive midpoints, the predictive quality of which will depend on the amount of data collected previously.
The paper introduces a framework that hopes to make Learning Analytics easier to understand, and hopefully implement, for teachers and analysts. It offers a general method of performance evaluation and tuning that defines parametres automatically from the available data. Given the broad variety of learning contexts, the proposition is both ambitious and challenging. Fortunately, Moodle’s processes for generating data from students performance and behavior (the Moodle Logs) offer clear and standardized outcomes. Its ease of access is one of the reasons why powerful analytics and reporting solutions, such as Intelliboard or Blackboard’s X-Ray, are possible in Moodle.
The paper introduces a framework rather than showing actionable data from real users. We can list its key features as its main outcomes\findings:
- The model works independently from the rest of the LMS. It can run without affecting the site’s performance, possible on a different processing unit (CPU).
- The model is of the “supervised learning” kind. Unlike an unsupervised learning model, the data is clearly labelled, the desired outcomes are known, it’s predictive power is easier to evaluate and its scope of action is limited to the specific domain in which it is implemented.
- The data that feed the model come from the Moodle Logs. They are made available through an API written in the PHP language that sends data to the model. The model has a PHP version, built as a failsafe but known to be resource intensive and less powerful; and a version written on the much better suited Python language, and using TensorFlow, the most popular general-purpose machine learning framework today.
- The model offers two predictive methods: A “logistic regression,” which weighs the attributes of each row of data and estimates parametres that minimizes the “cost” of the model; and a “Feed Forward” Artificial Neural Network, where the parameters are recursively estimated and weighted using matrix multiplication, until the model achieves a desired predictability power. Neural Networks tend to offer more accurate predictions, but they are more computationally demanding and have a higher risk of overfitting.
- The model evaluation uses the Matthews correlation coefficient, which takes into account four possible outcomes: True or False positives, and True or False negatives. A model is better if the rate of “Truths” is higher and the rate of Falsehoods is lower.
Early testing of the Moodle took advantage of an anonymized sample of nearly 47 thousand students. While the reported results show positive prediction power of the model, for both the logistic regression and the neural network, the paper does not explicitly assert that the Moodle Learning Analytics model is an actionable tool worth implementing across teaching contexts and environments.
This research falls into the “Classification” category. Academic surveys divide the field into it, along with the likes of “Clustering,” “Association” and “Predication.” Naturally, there are overlaps and discrepancies.
The researchers, from Moodle and Western University, thank Moodle and Dalton. They do not disclose the source of funding of the research nor whether there are conflicts of interest influencing the research.
The nascent field of Educational Data Mining will meet again in Montreal next July for the EDM 2019 conference. Moodle’s participation is yet to be confirmed. Among the list of proposed topics, which signal research interests in the field, the conference site includes:
- Student modeling
- Domain knowledge data and representations
- Real-world problem-solving in open-ended domains
- Informing data research with educational theory
- Affective and emotional states detection and intervention
- Social network analysis to study student interactions
- EDM techniques, approaches, methods and frameworks
- Legal and social aspects
This Moodle Technology related post is made possible by: Moonami a company that provides a full range of Moodle services that combine the flexibility, scalability, and power of Amazon’s world-leading cloud platform (AWS) with fanatical Moodle support. Click here to learn more.