Artificial intelligence and data analytics are impacting our lives at virtually every level. The educational context is no exception. Thus is born what we call Learning analysisa research discipline that attempts to apply data analysis to improve the teaching and learning process.
Among many other types of studies, there are several research initiatives that work with increasingly sophisticated predictive algorithms that seek to anticipate factors such as a student’s risk of dropping out or the grade they might obtain.
Program to plan
Early work focused on predicting grades relied on applying a set of pre-established rules to a relatively simple set of facts. More recently, however, proposed work analyzes the entire record of a student’s interaction with their educational platform and uses complex neural networks to arrive at this grade prediction. In fact, truly amazing results are achieved.
For example, in the work presented by Alonso-Misol et al. the performance of different algorithms is compared, achieving 96% accuracy when predicting an exam’s score. This means that, in 96 students out of 100, the human puts a mark very close to the one that the algorithm said it was going to put.
It is a relatively recent discipline. It is expected that the results will get better and better, so at some point the following question might be asked: since we have a system that predicts the grade a student will get on the exam with a margin of reasonable error, can we use the prediction as the final grade, and we forget the exam?
a utopian wish
It sounds tempting, exams are the most hated activity by students. They also do not enjoy much sympathy among teachers, as grading is a very costly task. Surely both groups would applaud the end of the exams. However, the reality is quite different and it is very likely that the prediction, now successful, will only produce meaningless numbers if the test disappears.
This is due to the operation of predictive systems that rely on supervised learning techniques. Essentially, the principle of operation is as follows: current price data is analyzed and compared to past price data. If in past courses there is a pattern of activity that can be linked to earning a certain grade, then students who have that pattern in the current course will be predicted to have that grade. In other words, it is very likely that a student will obtain a grade similar to that obtained by students who have had a similar interaction with the platform.
Thus, predictive systems will be successful to the extent that the analyzed course performs equivalently to previous editions of the same course.
To fully understand this, let’s imagine a month-long course in which students must submit an activity on Friday of each week. There will be students who only generate activity on Fridays, to make the delivery. There will be other students who will generate activity throughout the week, with more intensity on Friday. What seems clear is that it will be a weekly activity pattern.
Now imagine that a new teacher comes in and decides that the activities for the four weeks are all due at the end of the month. Certainly, the pattern of student activity will change, and there will even be students who will not enter the course until last week. This change in teaching methodology will make the comparison between the current course and previous courses no longer meaningful. As a result, predictive systems will lose all their potential.
Something similar would happen when deleting the review. Although there are students who have a strong intrinsic motivation to continue their learning, the extrinsic motivation imposed by an exam is the main motivating factor to continue the activity of the course. In other words: without an exam, students would work less in the course and with a very different work pattern. Predictive systems would therefore lose their value.
Other assessment possibilities
If we want (we want) to eliminate the examination, then we must think of the whole catalog of alternative activities described in this other article. In all cases, whether it is called “exam”, “rubric”, “portfolio”, or in any other way, the activity of the students is strongly modulated by the deadlines of the activities which count for the grade.
The objective of predictive student qualification systems is mainly to detect students at risk of dropping out in order to offer them adequate support. They are also useful for anticipating the resources that will be needed. Eliminating the final exam, however, is not one of the goals of predictive grading systems.