A Pre-Trained Graph-Based Model for Adaptive Sequencing of Educational Documents

TL;DR

This study introduces a novel data-efficient framework for learning path personalization that operates without expert annotation, and shows that this pre-training stage substantially improves data-efficiency in a range of adaptive learning scenarios featuring new educational materials.

Abstract

Massive Open Online Courses (MOOCs) have greatly contributed to making education more accessible. However, many MOOCs maintain a rigid, one-size-fits-all structure that fails to address the diverse needs and backgrounds of individual learners. Learning path personalization aims to address this limitation, by tailoring sequences of educational content to optimize individual student learning outcomes. Existing approaches, however, often require either massive student interaction data or extensive expert annotation, limiting their broad application. In this study, we introduce a novel data-efficient framework for learning path personalization that operates without expert annotation. Our method employs a flexible recommender system pre-trained with reinforcement learning on a dataset of raw course materials. Through experiments on semi-synthetic data, we show that this pre-training stage substantially improves data-efficiency in a range of adaptive learning scenarios featuring new educational materials. This opens up new perspectives for the design of foundation models for adaptive learning.

Figures (8)

• Figure 1: Illustration of the partition of the domain: sequential corpora are designed to be followed in one single way whereas graph corpora can be navigated along a variety of paths. This dichotomy can be used to pre-train a recommender system on readily available source tasks $(\mathcal{M}_s)_{s \in \mathcal{S}}$ and fine-tune it on a target task ${\mathcal{M}}_t$.
• Figure 2: Illustration of the sampling process of a student-corpus pair. $A$ is the adjacency matrix corresponding to the set of edges $\mathcal{E}=\mathcal{E}_{\text{prereq}}\cup \mathcal{E}_{\text{pref}}$.
• Figure 3: Overview of the recommendation pipeline on one student
• Figure 4: Performance comparison (mean and confidence interval) depending on the pre-training strategy, for multiple prior knowledge distributions in the population. At each epoch, new data are collected from 5 new sessions with simulated students. In total, the training is carried out on a group of 50 students. These results are aggregated over 30 random seeds.
• Figure 5: An intuitive view of student's knowledge state model with keyword vectors