Recommendation systems used in digital learning
Introduction to recommendation systems
According to R. Burke1 and L. Cui2, a recommendation system is a system that produces individualized recommendations or that has the effect of guiding the user in a personalized way towards interesting or useful items when many options are possible. Recommendation systems are gradually becoming indispensable elements in various fields to meet the ever-increasing demands of users and to help them sort through the enormous amount of information and products available. The system typically consists of four components: the user model, the recommendation object model, the recommendation algorithm and the recommendation assessment. We will see how recommendation systems can be used in the learning field.
See the different recommendation algorithms
In the literature, recommendation algorithms have generally been classified into several types, depending on how recommendations are made3:
- Contend based: the user is recommended items similar to those he has preferred in the past. Content-based recommendation systems analyze a set of items and/or descriptions previously preferred by a user and build a model or profile of the user’s interests based on the characteristics of those items.
- Collaborative filtering: the user is recommended items that people with similar tastes have valued in the past. Collaborative recommendation systems (or collaborative filtering) predict a user’s interest in new items based on recommendations from others with similar interests.
- Demographic Recommendation: users are listed according to the attributes of their personal profile, and recommendations are made based on their demographic classes e.g. according to geographic location, age, gender, etc.
- Rule Based: the recommendation is made from a mathematical function that has been modeled according to predetermined parameters. An example would be a function to recommend the best car insurance and would take in input information such as the model of your car, its mileage or the history of your incidents on your driver’s license.
- Knowledge Based: the algorithm suggests items based on logical inferences predefined4 by an expert in the field. It is a representation of knowledge (e.g., ontologies) about how an item meets a particular user need.
Each recommendation algorithm has advantages and disadvantages and is suitable for different application scenarios. It is often more appropriate to build a hybrid recommendation system that combines several algorithms.
Recommendation systems in digital learning
Recommendations systems have emerged in digital learning through Learning Management Systems (LMS) to improve learning and e-learning processes5. However, with the growth in the number and diversity of learning resources, the problem of information overload is becoming increasingly critical. Therefore, offering learners personalized learning recommendation tools is a necessity6.
Individualization in LMSs refers to the functionality that allows the system to respond in a personalized way to a learner’s needs and characteristics to help them progress. For example, levels of proficiency, prior knowledge, cognitive abilities, skills, interests, preferences, and learning styles. This type of personalization can help improve the overall quality of learning by providing recommendations for learning content that are useful for achieving the pedagogical goal but can also improve learner performance7 and satisfaction8. It is important to note that there are major differences between general purpose and learning recommendation systems, including the purpose of the system.
In areas such as e-commerce, a user is looking to purchase a product with a specific quality and price range, whereas the purpose of learning recommendation systems is to help the user, or a group of users find appropriate learning resources and activities for better achievement of the learning objective and skill development. While the principles of recommendation systems may well correspond to those of learning sciences, it is important to note that recommendation systems often need to be adapted to facilitate learning9.
Unlike marketing-oriented systems, learning recommendation systems need to engage and challenge learners. Rather than providing an environment with a lot of limitations, these systems are designed to leave a lot of room for exploration and to confront learners with unexpected content, thus enabling discovery learning. All these differences between e-commerce and learning contexts call for an adaptation of recommendation systems so that they become powerful learning tools10.
The challenges faced when recommending for learning are different from those of traditional e-commerce or other systems. The criteria needed to create a relevant recommendation must be considered:
To improve the learning process, recommendation systems must be able to guide learners and recommend content or strategies to help achieve learning objectives11.
It must be taken into account that learner proficiency evolves over time. They never reach a final level of knowledge or state of competence, but rather move on to the next level. A second specificity is the consideration of learning theories in the system. Learners must be confronted with unexpected content, as this encourages them to learn through discovery and exploration. Recommending items different from those a learner already knows stimulates critical thinking and avoids confirmation bias12 (cognitive bias, which consists in favoring information confirming one’s preconceived ideas/assumptions and giving less weight to hypotheses and information that work against one’s conceptions.)
The pedagogical relevance of the recommendations is also critical, the ultimate goal of course being to help the learner improve. Specific assessment metrics are therefore put in place for learning recommendation systems, which will be discussed in the next sections.
The choice of the algorithm or algorithms of recommendation is at the heart of the process of creating a system. This choice must therefore be adapted to the needs of the system to be set up in order to guarantee appropriate recommendations.
The results of A. C. Rivera’s study revealed that one of the most frequently used types of recommendation systems in learning is collaborative filtering (CF), which accounts for about 30 of the scientific articles reviewed. Despite its popularity, CF has some drawbacks. Among them, the lack of evaluations (parsimony), which refers to the situation in which data is scarce and insufficient to identify similarities of interest between users. There is also cold start: this problem occurs at launch when there is no assessment for new resources or when a new user has not yet assessed any items.
The most popular solution to overcome these limitations is the hybrid approach. Approximately 46 of the scientific articles reviewed are based on a combination of different types of recommendations. Hybrid systems are implemented to exploit the advantages of more than one technique while compensating for the disadvantages of each. The hybrid solution then allows the implementation of a robust system, adapted to several use scenarios.
Assessing recommendation systems in a general context is a complex process. It is even more so for a learning context, which must evaluate, in addition to the general parameters, the impact of the recommendations during learning. General assessment standards for a recommendation system include predictive accuracy, diversity, fraction of coverage, serendipity, confidence coefficient, robustness, scalability, etc.
In the case of recommendation systems for learning, it is requested that they are assessed not only according to technical standards, but also by a combination of technical and pedagogical criteria13. The evaluation is based on three aspects:
Measure the performance of the recommendation system
The objective of this evaluation is to measure the performance of the recommendation system or recommendation algorithm from a technical point of view, such as prediction accuracy, fraction of coverage and relevance of recommendations.
Measuring user-centered effects
This is to measure the general perception of the recommendation system by the user. This includes preference or satisfaction, user confidence in the system’s recommendations, user perception of novelty, diversity and serendipity14 of recommendations.
Measuring effects on learning
The purpose of this assessment is to measure the impact of the recommendations on the learning objectives. One of these objectives is to measure the learner’s learning performance, which includes measuring the learner’s level of proficiency on a particular topic and comparing the learning outcomes with the learner’s test scores. Another objective is to measure the learner’s success, which indicates the time it takes à̀ a learner to achieve a learning objective and the content consulted.
Despite their omnipresence in our digital lives, recommendation systems are not yet widespread in the creation of personalized learning paths, the reasons being the lack of maturity in the implementation of models that take into account the pedagogical challenges, but also the lack of objectivity in relation to the impact of the solutions implemented in recent years. However, this field remains promising with the expansion of e-learning and the popularization of adaptive learning.
1 R. Burke, “Hybrid Web Recommender Systems”, en, in The Adaptive Web, P. Brusilovsky, A. Kobsa et W. Nejdl, éd., t. 4321, Berlin, Heidelberg : Springer Berlin Heidelberg, 2007, p. 377-408, isbn : 9783540720782. doi : 10.1007/978-3-540-72079-9_12. adresse : http://link.springer.com/10.1007/978-3-540- 72079-9_12 (visité le 19/08/2020).
2 L.-Z. Cui, F.-L. Guo et Y.-j. Liang, “Research Overview of Educational Recommender Systems”, en, in Proceedings of the 2nd International Conference on Computer Science and Application Engineering – CSAE ’18, Hohhot, China : ACM Press, 2018, p. 1-7, isbn : 9781450365123. doi : 10.1145/3207677.3278071. adresse : http://dl.acm.org/citation.cfm?doid=3207677.3278071 (visité le 18/08/2020).
3 G. Adomavicius et A. Tuzhilin, “Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions”, IEEE Transactions on Knowledge and Data Engineering, t. 17, no 6, p. 734-749, juin 2005, issn : 1041-4347. doi : 10.1109/TKDE.2005.99. adresse : http://ieeexplore.ieee. org/document/1423975/ (visité le 19/08/2020).
4 Inferences are stages of reasoning, moving from propositions to logical results.
5 Julián, M.-P.; Jose, A.; Edwin, M.; Camilo, S., “Autonomous recommender system architecture for virtuallearning environments”, Applied Computing and Informatics, 2020.
6 M. Erdt, A. Fernandez et C. Rensing, “Evaluating Recommender Systems for Technology Enhanced Learning: A Quantitative Survey”, IEEE Transactions on Learning Technologies, t. 8, no 4, p. 326-344, oct. 2015, issn : 1939-1382. doi : 10.1109/TLT.2015.2438867. adresse : http://ieeexplore.ieee.org/ document/7115158/ (visité le 18/08/2020).
7 A. C. Rivera, M. Tapia-Leon et S. Lujan-Mora, “Recommendation Systems in Education: A Systematic Mapping Study”, en, in Proceedings of the International Conference on Information Technology & Systems (ICITS 2018), Á. Rocha et T. Guarda, éd., t. 721, Cham : Springer International Publishing, 2018, p. 937- 947, isbn : 9783319734491 9783319734507. doi : 10.1007/978-3-319-73450-7_89. adresse : http: //link.springer.com/10.1007/978-3-319-73450-7_89.
8 H. Imran, M. Belghis-Zadeh, T.-W. Chang, Kinshuk et S. Graf, “PLORS: a personalized learning object recommender system”, en, Vietnam Journal of Computer Science, t. 3, no 1, p. 3-13, fév. 2016, issn : 2196-8888, 2196-8896. doi : 10.1007/s40595-015-0049-6. adresse : http://link.springer.com/10.1007/s40595- 015-0049-6.
9S. Garcia-Martinez et A. Hamou-Lhadj, “Educational Recommender Systems: A Pedagogical-Focused Perspective”, in Multimedia Services in Intelligent Environments, G. A. Tsihrintzis, M. Virvou et L. C. Jain, éd., t. 25, Heidelberg : Springer International Publishing, 2013, p. 113-124, isbn : 9783319003740 9783319003757. doi : 10.1007/978-3-319-00375-7_8. adresse : http://link.springer.com/10.1007/978- 3-319-00375-7_8.
10 J. Buder et C. Schwind, “Learning with personalized recommender systems: A psychological view”, en, Computers in Human Behavior, t. 28, no 1, p. 207-216, jan. 2012, issn : 07475632. doi : 10.1016/j.chb. 2011.09.002. adresse : https://linkinghub.elsevier.com/retrieve/pii/S0747563211001956
11 N. Manouselis, H. Drachsler, K. Verbert et E. Duval, Recommender Systems for Learning, sér. Sprin- gerBriefs in Electrical and Computer Engineering. New York, NY : Springer New York, 2013, isbn : 97814614436059781461443612. doi : 10.1007/978-1-4614-4361-2. adresse : http://link.springer.com/10.1007/978- 1-4614-4361-2.
12 M. Erdt, A. Fernandez et C. Rensing, “Evaluating Recommender Systems for Technology Enhanced Learning: A Quantitative Survey”, IEEE Transactions on Learning Technologies, t. 8, no 4, p. 326-344, oct. 2015, issn : 1939-1382. doi : 10.1109/TLT.2015.2438867. adresse : http://ieeexplore.ieee.org/ document/7115158/.
13 H. Drachsler, K. Verbert, O. C. Santos et N. Manouselis, “Panorama of Recommender Systems to Support Learning”, en, in Recommender Systems Handbook, F. Ricci, L. Rokach et B. Shapira, éd., Boston, MA : Springer US, 2015, p. 421-451, isbn : 9781489976369 9781489976376. doi : 10.1007/978-1-4899- 7637-6_12. adresse : http://link.springer.com/10.1007/978-1-4899-7637-6_12.
14 Definition of « serendipity » according to Oxford Languages: “the occurrence and development of events by chance in a happy or beneficial way.” .
Our latest articles about the topics: learning, science and Domoscio's news.