Attentional Neural Factorization Machines for Knowledge Tracing

Abstract

To promote the quality and intelligence of online education systems, knowledge tracing becomes a fundamental and crucial task. It models knowledge state and predicts performance based on student’s learning records. Recently, factorization machine based approaches have been proposed to fit student’s learning behavior by generating interactions between underlying features. However, there are two major unresolved issues: (1) quantities of interactions are introduced along with different features involved in. Nevertheless, the redundant interactions do not effectively contribute to the model training. There is a need to extract the most important and relevant interactions. (2) The widely employed factorization machines simply process with second order interactions, leading to insufficiently expressive representations, which is not beneficial to prediction. To this end, we propose a novel Attentional Neural Factorization Machines (ANFMs) to address the above problems. First, we leverage attention mechanisms to suppress the interference of interactions redundancy by distinguishing the importance of different interactions. To be specific, explicit and implicit attention strategies are utilized respectively. Secondly, to facilitate expressive capability, we apply neural networks to the transformed interactions after attention propagation, which is able to capture high order representations. Extensive experiments on two real-world datasets clearly show that ANFMs outperforms baselines with significant margins, which demonstrates the effectiveness of our work.