Over the past couple of decades, there has been an increasing adoption of Internet technology in the e-learning domain, associated with the availability of an increasing number of educational resources. Effective systems are thus needed to help learners to find useful and adequate resources, among which recommender systems play an important role. In particular, learning path recommender systems, that recommend sequences of educational resources, are highly valuable to improve learners’ learning experiences. Under this context, this PhD Thesis focuses on the field of learning path recommender systems and the associated offline evaluation of these systems.

This PhD Thesis views the learning path recommendation task as a sequential decision problem and considers the partially observable Markov decision process (POMDP) as an adequate approach.

In the field of education, the learners’ memory strength is a very important factor and several models of learners’ memory strength have been proposed in the literature and used to promote review in recommendations. However, little work has been conducted for POMDP-based recommendations, and the models proposed are complex and data-intensive. This PhD Thesis proposes POMDP-based recommendation models that manage learners’ memory strength, while limiting the increase in complexity and data required.

Under the premise that recommending learners useful and effective learning paths is becoming more and more popular, the evaluation of the effectiveness these recommended learning paths is still a challenging task, that is not often addressed in the literature. Online evaluation is highly popular but it relies on the path recommendations to actual learners, which may have dramatic implications if the recommendations are not accurate. Offline evaluation relies on static datasets of learners’ learning activities and simulates learning paths recommendations. Although easier to run, it is difficult to accurately evaluate the effectiveness of a learning path recommendation. This tends to justify the lack of literature on this topic. To tackle this issue, this PhD Thesis also proposes offline evaluation measures, that are designed to be simple to be used in most of the application cases.

The recommendation models and evaluation measures the we propose are evaluated on two real learning datasets. The experiments confirm that the recommendation models proposed outperform the models from the literature, with a limited increase in complexity, including for a medium-size dataset. In addition, the measures proposed actually allow to characterise and differentiate the algorithms.