Predicting eLearning Dropout
September 23rd, 2009
Despite its promise, a continuing challenge to online learning/ distance learning/ and eLearning is student dropout. Studies have consistently found higher student dropout rates in these courses than in in-person courses. There have been numerous studies attempting to predict dropout, but few have gotten to the holy grail of being able to identify students who are at risk while they are in the course, based on their behavior in the course.
Ionna Lykourentzou and collegues, working in Greece, reported in Computers & Education their attempt to use machine learning to do just that. The hope is that this capability would then lead to the ability to intervene with the at-risk students. They also analyze their results in terms of accuracy, sensitivity, and precision so that the analysis takes into account both false positives and false negatives.
Machine learning generally involves the creation of algorithms that allow computers to learn from data. In this context, computers are fed data from existing classes with student variables and eventual completion or dropout of the students known. The computer then uses its algorithms to create a model that will best predict these known outcomes. Following this, new data, without outcomes is entered into the computer. The computer is then asked to predict, based on its previous learning, whether students will dropout or complete. These predictions can be compared to the actual results to understand the effectiveness of the computer model. How good was it at predicting?
Lykourentzou et al looked at three different methods of machine learning. Then, they looked at ways to combine the results of those methods, since each by itself may not be perfect. This asked the question, is it better, for example, to classify a student as a dropout if at least one of the methods classified them as such, or should we require two of the three methods to classify them as a dropout to code them as such?
The authors included gender, geographical location, work experience, education level, English language literacy, test grades, project grades, project submission date (e.g., was it late?), and number of sessions in the online environment. The accuracy of the predictors was judged at each of the seven sections of the course (more information, such as grades, was available at the end of each section). Each of the predictors ended up being important in the models, but time invariant data (such as demographics; data that does not vary through the course) was less accurate than time varying data (such as grades).
With all of the variables and the machine learning techniques, it was determined that the best way to predict dropout was to run all three machine learning techniques and classify a student as a dropout if any of the three indicated that they were classified in that category. This technique achieve accuracy ratings between 97 and 100%, sensitivity rates between 95 and 100%, and a precision rate of 100%.
There are some drawbacks to these methods. Predictions are dependent on the quality and representativeness of the training data. So, if a previous semester’s data is used to train the algorithms and there were some very idiosyncratic events in either that semester or this semester, prediction will not be as good. However, with these complex techniques running in the background, fairly simple interfaces could be created for teachers to highlight at-risk students for intervention.
Perhaps this is another way we can begin to address the dismal dropout numbers at post-secondary institutions, particularly in online courses.
Lykourentzou, I., Giannoukos, I., Nikolopoulos, V., Mpardis, G., & Loumos, V. (2009). Dropout prediction in e-learning courses through the combination of machine learning techniques Computers & Education, 53 (3), 950-965 DOI: 10.1016/j.compedu.2009.05.010
Tags: dropout
Posted in Students | Comments (0)
Leave a Reply
Recent Posts
- Games and Simulation in Science
- Simulations and Labs: Either/or?
- Open Source, E-Textbooks, Traditional Texts… oh my!
- Predicting eLearning Dropout
- Development of Help-Seeking
Categories
- Assessment (3)
- Curriculum and Materials (5)
- Policy/ Government (2)
- Publication (1)
- Research Methods (4)
- Schools (1)
- Students (13)
- Teachers (8)
- Technology (8)
- Uncategorized (3)
