Students who use out-of-school eLearning tools, such as the mobile application Nomad Education, are seeking a complement to the education they receive in the classroom. Indeed, the conditions of classroom teaching in the National Education system (allocated time slots for each subject in the timetable, group sizes) do not always allow for adaptation to the students' needs. Some students therefore download the Nomad Education mobile application to rework certain subjects from their school curriculum with a different pedagogical approach. The use of the application is voluntary, out-of-school, and unsupervised. The objective of my work is to integrate into the existing mobile application an educational recommender system to support the learner in a personalized way in their use of the application. The challenge is to allow the student to abstract from their academic level by working on resources based on their needs and achievements and to make connections between different chapters they may encounter during their learning. For this, we want a system that is both adaptive and intelligent.

The contributions proposed aim to provide answers to the following problem:

How to propose an educational recommender system adapted for an unstructured and out-of-school use by exclusively exploiting the content of a mobile application and the learners' activity within this application?

Identifying the learning needs of each user is a complex task. Prior to this doctorate, Nomad Education applied the Item Response Theory (IRT) to its content with the aim of integrating it into the recommendation system. The integration of IRT to address this issue raises a first research question:

(QR-1): Can we use Item Response Theory (IRT) to identify the learning needs of users based on their quiz responses?

In an out-of-school context, we have only a partial view of the user’s learning. Since the use of the application is voluntary, the recommendation system must meet the specific needs of the learner and integrate into their use of the mobile application. The second research question is as follows:

(QR-2): How to propose pedagogically relevant activities adapted to the learner's specific study preferences, without a pre-established pedagogical program, in a micro-learning context?

To address my problem, I proposed a model for implementing a recommendation system for a mobile application used in an extracurricular context. This model is the subject of my first contribution.

To study QR-1, I use the Item Response Theory to identify levels of learners. I hypothesize that each level group corresponds to recommendation needs, for which I define personalization strategies.

To address QR-2, I propose a recommendation score based on three components (pedagogical, historical, and novelty) to rank the resources selected by the previously identified recommendation strategy.

Associated with this recommendation model combining IRT and recommendation score, I propose a process to implement it within the Nomad Education application. This first contribution was evaluated with teachers to validate the pedagogical interest of the recommendations and then by analyzing the use of recommendations by learners in an ecological context.

In my second contribution, I rely on the analyses made of the model proposed in the first contribution to suggest alternatives. I revisit QR-1 by studying the impact of different IRT exploitation parameters in the initial model. I propose variations to modify the IRT criteria for creating level groups, questioning the personalization strategy associated with each IRT level and allowing the learner to choose their personalization strategy outside of the one assigned based on IRT. I also revisit the presentation interface of the recommendations implemented in the first contribution to study if displaying the academic level of recommended resources can impact recommendation follow-up.

In my third contribution, I start from the analysis of the use of the recommendation system by learners to complement the answer provided to QR-2 in the first contribution. The analysis of traces has highlighted that some resources are more popular; I propose a knowledge discovery process, extracting traces of relevant information usage that can enrich the knowledge bases on which the proposed recommendation model relies. This knowledge discovery is based on the notions addressed in the recommendations and the sequence of activities within the application.