Publication
Learning Analytics
The discipline of “learning analytics” consists of analysing data from learning systems, particularly e-learning, in order to understand the link between learners’ behaviours and their learning performance. That is why most MOOC platforms record all learner activities to create huge databases.
Then, automatic analysis techniques such as “deep learning” try to identify correlations between learners’ behaviour and the quality of their learning [2]. The hope is that the correlation of the large quantity of variables collected will enable the emergence of a predictive model of the quality of knowledge acquisition. However, the difficulty lies in interpreting the correlations thus obtained [2].
When we know that the average amount of time that learners spend concentrating on a training video is about 6 minutes [3], how reliable can we expect behavioural measures to be beyond that time? Furthermore, all teachers know that the most important element in determining a learner’s knowledge acquisition is prior knowledge [4].
How, then, can this essential variable be factored into correlation models? While it is reasonable to think this variable can be gauged in school or even university training, where the prior training of learners is fairly well known and measured, the situation is quite different in the case of vocational and continuing training. In the latter case, the heterogeneity of learners’ backgrounds is the rule rather than the exception. Finally, what about the influence of the pedagocial quality of content? Can we apply correlations obtained on the basis of an effective course to another, poorly constructed course? Obviously not. The overall result is that, for the time being, the results of analytical studies based on MOOCs are rather meagre [1].
But perhaps the problem lies in the methodological approach? Indeed, another technique would be to set out a clear pedagocial goal before measuring. What exactly do we want to improve? Then we need to determine a causal model between measurable values and the educational goal. This will make the influence of measured variables on this goal clear; learners can then take advice from it accordingly. As every student of statistics knows, there is a fundamental difference between correlation and causation. A correlation is an observation of the simultaneous variation of two or more variables, in other words their interdependence. But this says nothing about causality. The latter requires the prior establishment of a causal model explaining how the setting of the value of one variable influences the value of another [5].
Let’s take a caricatural example. Let’s say that on a MOOC platform we observe that learners who stop watching a video at 4pm and resume at 4.430pm do better on the final test than those who stop at 7pm and resume at 8pm. Should learners be advised to always stop at 4:00pm for 30 minutes? Certainly not. First of all, learning sciences advise taking a break after a learning period of about an hour. The 4pm break may be at the end of a one-hour period, while the 7pm break is after three hours. Then, the 4pm break could be a coffee break during which the learner consumes a stimulant (coffee, tea) that boosts his energy for the rest of the learning process. On the other hand, the learner who stops at 7pm. may have a meal after which a period of digestion is not conducive to learning. Lastly, the learner’s general fatigue at 7pm is probably greater than at 4pm, and this does not favour learning. With this simple example, we can see that correlation is difficult to interpret out of context. Moreover, some variables are not included in the correlated data. For example, what was consumed during the break? As we can see, the establishment of a coherent analytical model of e-learning is complex. However, it is a prerequisite for drawing recommendations from the analysis. The advent of microlearning, beyond its own pedagocial advantages, introduces a framework that facilitates the design of a model for measuring and interpreting learners’ behaviour. This makes it possible, in particular, to implement adaptive learning. In the e-Cadencia system of ITycom, the adaptive learning
involves adapting the learning rhythm as well as the micro-contents (microlearning contents) to the learners’ dispositions.
Do not miss our next newsletter, in which we will tell you more about learning analytics, microlearning and adaptive learning in e-Cadencia.
Dr Ph. Dugerdil
Product Owner
References
- [1]
Reich J. – Failure to Disrupt: Why Technology Alone Can’t
Transform Education. Harvard University Press,
25 septembre 2020. - [2]
Baker R.S. – Challenges for the Future of
Educational Data Mining: The Baker Learning Analytics Prizes. Journal of
Educational Data Mining, Volume 11, No 1, 2019 - [3]
Lagerstrom, L., Johanes, P., Ponsukcharoen, U. – The Myth of the
Six-Minute Rule: Student Engagement with Online Videos. ASEE Annual Conference
& Exposition, Seattle, 2015 - [4]
Ambrose S., Bridges M.W., DiPietro M., Lovett
C. Norman M.K. – How Learning Works. John Wiley, 2010 - [5]
Pearl J. – The Book of Why: The New Science of
Cause and Effect. Basic Books, 2018
Then, automatic analysis techniques such as “deep learning” try to identify correlations between learners’ behaviour and the quality of their learning [2]. The hope is that the correlation of the large quantity of variables collected will enable the emergence of a predictive model of the quality of knowledge acquisition. However, the difficulty lies in interpreting the correlations thus obtained [2].
When we know that the average amount of time that learners spend concentrating on a training video is about 6 minutes [3], how reliable can we expect behavioural measures to be beyond that time? Furthermore, all teachers know that the most important element in determining a learner’s knowledge acquisition is prior knowledge [4].
How, then, can this essential variable be factored into correlation models? While it is reasonable to think this variable can be gauged in school or even university training, where the prior training of learners is fairly well known and measured, the situation is quite different in the case of vocational and continuing training. In the latter case, the heterogeneity of learners’ backgrounds is the rule rather than the exception. Finally, what about the influence of the pedagocial quality of content? Can we apply correlations obtained on the basis of an effective course to another, poorly constructed course? Obviously not. The overall result is that, for the time being, the results of analytical studies based on MOOCs are rather meagre [1].
But perhaps the problem lies in the methodological approach? Indeed, another technique would be to set out a clear pedagocial goal before measuring. What exactly do we want to improve? Then we need to determine a causal model between measurable values and the educational goal. This will make the influence of measured variables on this goal clear; learners can then take advice from it accordingly. As every student of statistics knows, there is a fundamental difference between correlation and causation. A correlation is an observation of the simultaneous variation of two or more variables, in other words their interdependence. But this says nothing about causality. The latter requires the prior establishment of a causal model explaining how the setting of the value of one variable influences the value of another [5].
Let’s take a caricatural example. Let’s say that on a MOOC platform we observe that learners who stop watching a video at 4pm and resume at 4.430pm do better on the final test than those who stop at 7pm and resume at 8pm. Should learners be advised to always stop at 4:00pm for 30 minutes? Certainly not. First of all, learning sciences advise taking a break after a learning period of about an hour. The 4pm break may be at the end of a one-hour period, while the 7pm break is after three hours. Then, the 4pm break could be a coffee break during which the learner consumes a stimulant (coffee, tea) that boosts his energy for the rest of the learning process. On the other hand, the learner who stops at 7pm. may have a meal after which a period of digestion is not conducive to learning. Lastly, the learner’s general fatigue at 7pm is probably greater than at 4pm, and this does not favour learning. With this simple example, we can see that correlation is difficult to interpret out of context. Moreover, some variables are not included in the correlated data. For example, what was consumed during the break? As we can see, the establishment of a coherent analytical model of e-learning is complex. However, it is a prerequisite for drawing recommendations from the analysis. The advent of microlearning, beyond its own pedagocial advantages, introduces a framework that facilitates the design of a model for measuring and interpreting learners’ behaviour. This makes it possible, in particular, to implement adaptive learning. In the e-Cadencia system of ITycom, the adaptive learning
involves adapting the learning rhythm as well as the micro-contents (microlearning contents) to the learners’ dispositions.
Do not miss our next newsletter, in which we will tell you more about learning analytics, microlearning and adaptive learning in e-Cadencia.
Dr Ph. Dugerdil
Product Owner
References
[5]
Pearl J. – The Book of Why: The New Science of
Cause and Effect. Basic Books, 2018
