MulOER-SAN: 2-layer multi-objective framework for exercise recommendation with self-attention networks

Abstract

Diagnosing the knowledge mastery level of students on required concepts and providing tailored exercises to them is an essential task in e-learning environments. Due to the different knowledge level of students and the large scale of exercise banks, it is difficult for general recommenders to recommend suitable exercises for each student, which decreases their learning efficiency. In this paper, we develop a 2-layer multi-objective framework for exercise recommendation with self-attention networks, abbreviated as MulOER-SAN, to capture the change in students’ knowledge acquisition and thus provide customized exercise recommendations. Note that MulOER-SAN is a 2-layer framework. Via the bottom layer, a self-attention mechanism is adopted to predict the coverage in the next exercise, which can lead the learning process in the right direction. Moreover, we implement a novel knowledge tracing model with an enhanced self-attention sub-layer, thus tracing students’ dynamic knowledge state evolution. As the top layer of the 2-layer model, exercises are filtered according to the bottom layer’s predicted results, and the candidate subsets with appropriate difficulty and novelty are generated. From the perspective of diversity, we also carefully develop a chaotic sparrow search algorithm to further filter the candidate subset to avoid redundancy of recommended results. On the above basis, a simple yet effective difficulty smoothness factor is implemented to generate a high-quality ranking list in a real teaching application scenario. Comprehensive experiments conducted on three real-world datasets have demonstrated that the proposed MulOER-SAN framework significantly outperforms state-of-the-art methods in terms of various evaluation metrics.