LORecommender: Knowledge-based recommendation strategies for personalized access to learning object repositories

The goal of our work on recommendation technologies in e-learning is to provide smart support for accessing to the Learning Objects (LOs) that exist in repositories. We have defined three recommendation strategies that make use of existing knowledge of the domain, as well as additional information from both the student and the activity, with an ontology-based semantic representation.

One of the strategies provides an student with a recommendation, list of educational resources that are adapted to the student’s learning needs. The second one promotes diversity in the recommendation list. The third strategy explores a proactive model for user interaction based on a navigation-by-proposing model. The implementation of these three strategies has lead to the proposal of a developed framework for the rapid prototyping of knowledge-based recommenders to the learning field. These three strategies have been implemented and evaluated in a computational way and in a real learning field, where teachers and students have shown their satisfaction with the recommendation strategies designed. These three strategies presented will come to address the weaknesses identified in recommender systems in education.

The knowledge sources

The knowledge sources employed by our recommendation strategies: the LOs, the domain ontology and the contextual information.

The Learning Objects

We describe the LOs using one of the most recognized metadata standards: IEEE LOM. LOM allows to tag the digital resources according to a set of predefined categories and assign values to each one. We propose to use a profile for describing LOs containing the next upper-level LOM categories: Life cycle, Technical, Educational, Relation and General.

The Life cycle category identifies the author and the licensing status of the LO. The Educational category helps to identify the type of LO (e.g., lecture note, solved example, quiz question, assignment, etc.). The Technical category groups the technical requirements and characteristics of the LO. Additionally, the Relation category identifies if an LO is a version of another one. Finally, the General category contains keywords that describe the domain learning topics covered by the LO according to the domain ontology vocabulary. So, these keywords are used to finally index the LO in the domain ontology, which is described next. While other category metadata can be exploited for retrieval purposes or only for presenting information to the user, the General category content will be especially important in our recommendation process because it will be used to compute the similarity between the LO and the learning goals that the user defines in a query.

The domain ontology

The ontology organizes the concepts that represent the domain topics using a taxonomy. It provides a general indexing scheme that includes similarity knowledge between the concepts representing these topics. As we will show, this similarity knowledge will be exploited by the recommendation strategy. The ontology also links each LO with the concepts that it covers. This information will be used by the recommendation strategy in order to determine the suitability of each LO.

ontologia
The ontology is inspired in other ontology about programming language concepts and it has been modeled using the OWL language. It has been validated through different tools: OOPS! in order to validate the quality and avoid pitfalls and Pellet to check the consistency. In addition, a group of experts validated the quality of the semantic metadata.

In this link you can find the current version of the ontology

The contextual information

We propose the use of two contextual elements in the learning environment:

  • Activity context: this knowledge is related to the inclusion of learning paths. A learning path reflects a successful sequence or order in which concepts are taught or learned in the corresponding field. Learning paths can help to filter out LOs that exemplify non-reachable concepts given the concrete cognitive state of the student. This contextual knowledge is static and specific to the learning domain. It must be defined by an expert (the instructor) before the students use the recommender. Learning paths are represented by a precedence property in the domain ontology.
  • Student context: this knowledge concerns the goals achieved by the student in the learning process. The goals achieved are represented by the concepts that the student should know and the mastery level achieved in each of them. This level is considered to be a degree of satisfaction, a metric that allows the recommendation strategy to know about the student’s knowledge level in a particular concept. This information evolves over time as long as the student interacts with the recommender and progresses in her learning.

We propose to use the activity context as a hard criterion to discard LOs that are not appropriate for the current student and the student context as a soft criterion that assigns utility to each LO.

Knowledge-based recommendation strategies for recommending learning objects

We propose three alternative knowledge-based strategies, each of them satisfies an essential requirement of recommenders of educational resources, (a) personalization, (b) overspecialization (or lack of diversity) and (c) facilities in the user interaction. Let’s briefly introduce each strategy:

  • The first knowledge-based recommendation strategy provides a high level of personalization. This strategy follows a reactive approach: the student provides an explicit query and the recommender system reacts with a recommendation response. The student poses a query using the concepts existing in the domain ontology. This query represents her in-session or short-term learning goals: the concepts she wants to learn in the session. In this knowledge-based recommender, priority is given to those LOs that are most similar to the student’s query and, at the same time, have a high pedagogical utility according to the student context information.
  • The second strategy tries to avoid the overspecialisation problem that affects to the pure-similarity recommendation strategies. This strategy includes diversity in the proposals making use of a ranking approach, called diversity-conscious ranking, inspired in diversity-conscious strategy. Introducing diversity in the recommended LOs is also crucial for making the most of the recommendation session.
  • The third strategy uses navigation-by-proposing, a simple conversational process that avoids posing direct queries and carries a small feedback overhead from the students’ perspective. Initially, the proactive recommendation result is a set of LOs presented as an assignment proposal to start with the learning activity. Priority is given to LOs that better adapt to the student context. The student can select one of the proposed LOs or she can ask for a refined proposal by entering in a conversational process. We use navigation-by-proposing due to its convenience in complex domains where the user may not be able to answer a given question because her domain knowledge is insufficient.

Los comentarios están cerrados.