Supporting Learners with Recommender Services
Main impacts
- Learners are now provided with services that support them not only in retrieving their own learning resources by means of tag recommendations but also in finding new learning resources by means of resource recommendations.
- Our research on how to exploit cognitive models for the task of recommending resources and tags has increased the awareness for cognitive-inspired recommender systems in the research community.
Guiding and supporting learners via the Social Semantic Server (SSS) by providing novel recommender approaches that model individual and collaborative behavior.
The situation before Layers
Highlights
In recent years, the TEL domain has recognized the importance of recommender services for supporting learners in finding learning material in an overloaded information space. However, the algorithms behind existing TEL recommendation systems were not designed for the specific requirements of informal workplace learning.
Description
Current recommendation algorithms used in the TEL domain are not designed for the specific requirements of informal workplace learning. One of these requirements is the need for a strong personalization of the methods since informal learning happens in a self-organized way either individually or in small communities of practice. Hence, it is important for recommendation methods to model individual and collaborative behavior by taking into account the factors that shape the interaction of learners with learning resources.
One illustrative example of such interaction is the process of annotating learning resources with tags (i.e., freely-chosen keywords), which is known as social tagging. These tags are in turn used to structure, search for and retrieve content. Although there is already a large body of tag recommender algorithms out there, which aim to support users in finding descriptive tags, these approaches typically work in a data-driven fashion and neglect the cognitive factors that are important when users apply tags to learning resources.
Apart from that, there is a lack of open-source recommender services that can be freely consumed and adapted by learning tools.
What Layers did
Highlights
In the Layers project, it was our aim to design algorithms that mimic how people interact with learning resources and how these interactions are influenced by human memory processes. This was realized in the form of tag and resource recommendation algorithms, which are provided as services in the Social Semantic Server (SSS).
Description
In the Layers project, we developed various algorithms for recommending tags and learning resources based on models of human cognition. These algorithms were designed to mimic how people interact with learning resources and how these interactions are influenced by human memory processes.
In terms of tag recommendations, we incorporated the activation equation of the cognitive architecture ACT-R [1], which describes the information access in human memory. According to this equation, the relevance of an information chunk is given by its general usefulness defined by past usage frequency and recency (i.e., freshness in the sense of time) as well as by its usefulness in the current context defined by semantic similarity. We implemented this equation in the form of a novel tag recommender algorithm and showed that this approach outperforms various state-of-the-art methods (see [2]) in terms of prediction accuracy and computational complexity.
With respect to resource recommendations, we utilized a model of human category learning called SUSTAIN. This method aims at modeling attention-interpretation dynamics of a user by creating a user-specific network of topic clusters based on the resource topics the user has interacted with in the past. We used this network to evaluate the usefulness of candidate resources identified by Collaborative Filtering and recommended the most useful ones to the user. Evaluation results showed that this approach improves various Collaborative Filtering and content-based resource recommender algorithms [3].
Additionally, we implemented these algorithms as computationally efficient services that are provided by the Social Semantic Server (SSS) and thus, can be used by various Layers tools.
The situation after Layers
Highlights
Learners are now provided with tag and resource recommender services that assist them in retrieving learning material. Additionally, the algorithms are provided as open-source services that can be used by various tools in the area of informal workplace learning and beyond.
Description
Learners are now provided with services that support them not only in retrieving their own learning resources by means of tag recommendations but also in finding new learning resources by means of resource recommendations. In terms of tag recommendations this also means that novel (“innovative”) tags can be suggested to other users by our cognitive-inspired approach. This enables the emergence of innovative individual concepts, which would not be possible by a solely data-driven and popularity-based algorithm.
Furthermore, our resource recommendations also facilitate information sharing because own learning content can be helpful to others and can be reused by others if it is recommended to them. At the same time, our service could be used for the recommendation of other learners. For example, a possible collaborator could be suggested in LivingDocuments.
Finally, our algorithms are provided as open-source services that can be used by various tools in the area of informal workplace learning and beyond. This was demonstrated by the Layers tools that already incorporate our recommendation functionalities (e.g., Bits and Pieces and KnowBrain).
Impact that Layers created
Changed learning practices
Learners are now provided with services that support them not only in retrieving their own learning resources by means of tag recommendations but also in finding new learning resources by means of resource recommendations. Especially, our resource recommendations also facilitate sharing of information because own learning content can be helpful to others and can be reused by others if it is recommended to them.
Simultaneously, relevant recommendations increase the awareness of other learners’ resources. At the same time, our service could be used for the recommendation of other learners. This could promote collaboration and co-creation of resources.
Extended trust building and personal networks
Typically, learners are more likely to trust other learners that are similar to them (i.e., they intuitively trust their peer group). Thus, Collaborative Filtering based recommender approaches incorporate this by suggesting resources of like-minded learners in the network. This is also true for our cognitive-inspired SUSTAIN approach, which further utilizes the current attention by means of topic clusters of the learner. In [3] and [4] we have shown that our extended approach improves the recommendation accuracy since SUSTAIN outperforms Collaborative Filtering-based approaches.
Improved the creation and use of learning resources
Resource recommendations can help users in identifying similar resources that already exist in the learning network when creating a new resource. This is for example the case in KnowBrain and Discussion Tool, where users are notified if there are similar discussions / questions in the system already when entering a new one. In this case, the answers to the similar questions can be recommended to the learner. Additionally, this can avoid duplicating questions.
With our tag recommendation services, we supported the users in annotating their learning resources, which in turn assisted them in finding them. In [2] and [4] we have shown that our cognitive-inspired tag recommendation approach outperforms several state-of-the-art algorithms (e.g., Collaborative Filtering or Matrix Factorization-based approaches) in terms of recommendation accuracy and computational complexity. At the same time, our algorithms provided reasonable results in terms of diversity and novelty.
Enhanced digital competence
Healthcare professionals typically do not have the time to think about descriptive tags that could be applied to their learning resources, which results in sparse tagging data in tools such as Bits and Pieces. Tag recommendations overcome this problem by suggesting a set of tags that could be used by the learner to tag a resource of interest.
Changed Development Practices
Our research on how to exploit cognitive models for the task of recommending resources and tags has increased the awareness for cognitive-inspired recommender systems in the research community. Via several contributions at top-tier computer science conferences (e.g., ACM WWW, Hypertext, RecSys, EC-TEL, CIKM), we presented these types of recommender services as an alternative and/or extension to data-driven approaches to the community. Furthermore, this is again facilitated by the open-source strategy of the project (see TagRec and SocRec framework).
Links to other sections
Further Reading
Research papers: [4] [3] [2] [5] [6] [7] [8]
TagRec framework: https://github.com/learning-layers/TagRec/
SocRec framework: https://github.com/learning-layers/SocRec/
ScaR recommender framework (based on SocRec): http://scar.know-center.tugraz.at/
Slides presented by Simone Kopeinik at EC-TEL’2016 in Lyon, France about the performance of recommender services in TEL environment:
References
- J. R. Anderson, D. Bothell, M. D. Byrne, S. Douglass, C. Lebiere, and Y. Qin, “An integrated theory of the mind.,” Psychological review, vol. 111, no. 4, p. 1036, 2004.
- D. Kowald and E. Lex, “Evaluating tag recommender algorithms in real-world folksonomies: A comparative study,” in Proceedings of the 9th ACM Conference on Recommender Systems, 2015, pp. 265–268.
- P. Seitlinger, D. Kowald, S. Kopeinik, I. Hasani-Mavriqi, E. Lex, and T. Ley, “Attention Please! A Hybrid Resource Recommender Mimicking Attention-Interpretation Dynamics,” in Proceedings of the 24th International Conference on World Wide Web, 2015, pp. 339–345.
- S. Kopeinik, D. Kowald, and E. Lex, “Which Algorithms Suit Which Learning Environments? A Comparative Study of Recommender Systems in TEL,” in European Conference on Technology Enhanced Learning, 2016, pp. 124–138.
- D. Kowald, S. C. Pujari, and E. Lex, “Temporal effects on hashtag reuse in twitter: A cognitive-inspired hashtag recommendation approach,” in Proceedings of the 26th International Conference on World Wide Web, 2017, pp. 1401–1410.
- D. Kowald and E. Lex, “Studying Confirmation Bias in Hashtag Usage on Twitter,” arXiv preprint arXiv:1809.03203 - presented at European Computational Social Science Symposium, 2018.
- E. Lacic, D. Kowald, and E. Lex, “Tailoring Recommendations for a Multi-Domain Environment,” 2017.
- E. Lacic, D. Kowald, M. Reiter-Haas, V. Slawicek, and E. Lex, “Beyond Accuracy Optimization: On the Value of Item Embeddings for Student Job Recommendations,” arXiv preprint arXiv:1711.07762 - published in WSDM2018 workshop proceedings., 2017.
Contributing Authors
Dominik Kowald, Paul Seitlinger, Tobias Ley, Elisabeth Lex